Technical Field

[001] This disclosure relates to methods and systems for monitoring of photovoltaic modules (also referred to as “solar panels” or “solar energy generators”).

Background

[002] Solar energy facilities (SEFs) are designed to produce renewable energy by harnessing the inherent energy given off by the sun’s rays. Arrays of photovoltaic modules are employed to convert this energy into electrical energy, which may then be supplied to the national grid, used to power a private region or enterprise, or stored by suitable means. There exists a wide range of potential problems which may affect levels of energy production for any given SEF; these include wear & tear and normal panel degradation over time, faults in photovoltaic modules, wiring faults, inverter faults, lightning, meteorological conditions, daily and annual changes in the strength of the sun’s rays, dust, cloud cover, fog, smoke, birds and their droppings, snow accumulation, panel-to-panel imbalance, salt-spray accumulation, pollen, hail, fire, vermin attack, solar eclipses or any number of other perturbations. Some of these faults will occur instantaneously, or at least seemingly so. Others may affect one or more solar panels in the form of subtle deteriorations in performance over an extended period of time. Others still may occur intermittently, for example in conjunction with certain patterns of weather. Today, most of these faults remain undetected for extended periods, and result in power production losses.

[003] With reference to Figure 1 , many present solar energy facilities, or solar energy generators 100, comprise collections of solar modules or panels 102, 104 connected in string series to a combiner box 106, 108, which is connected to a DC/AC inverter 110, 112, before being exported to a grid or to storage, typically via arrangements of transformers and/or switchgears 114. This string series arrangement presents problems; for instance, if one panel in the string experiences a drop in power production, then the performance of each of the other panels in the same string will fall to a newly-established minimal energy production level, or “lowest common denominator”. Deploying panels in string series also necessarily involves working with higher voltages, increasing the risk to human operatives and the risk of arc fires occurring. String or central inverter faults can lead to major system shutdowns and significant losses in power production until repaired or replaced by a technician. In large-scale and domestic facilities, inverter failure is by far the most commonly-recorded fault type. In addition, string or central inverters must be replaced regularly (every 5-10 years) and have a cost which is high, relative to other operations and maintenance expenses during systems lifespans. Furthermore, inverter tripping (switching-off) can be caused by any number of module and cabling issues within a solar energy facility, which in turn causes shutdown of energy production for varying periods. In spite of this, it is still conventional across the solar industry today to rely on such string series arrangements, especially for particularly large-scale Solar Energy Facilities.

[004] It is currently common practice in the solar energy industry to develop projects with the aim of achieving short-term power production. Consequently, facilities are frequently operated for a decade or two, then any panels and other equipment are recycled or disposed of, before starting again afresh. This is clearly not ideal either from a technical perspective, or from an environmental perspective. Little regard is paid to instances of panels which are performing sub-optimally (for instance, faulty panels), as the facility will typically have a large quantity of solar panels which are deemed to function “sufficiently well”. There is a common prejudice among those working in the field of solar energy that the faults which presently go undetected, or have only a minor impact on the detected performance levels, are somehow “trivial” and not worth addressing. Solar energy facility operators who desire to find and remedy cases of under-performance may visually inspect the panels (either by eye, or by using drone-mounted cameras flown over the facility, for example) to find faults, but often the faults which are discovered by this visual inspection are too far advanced to address easily and should desirably have been addressed much earlier. That is, by the time that a panel is so badly damaged that the damage is visible to an observer, it is often too late to fix it. There is little utility in being able to detect instances of under-performing panels long after the opportunity to correct those under-performing panels has passed (and/or after the manufacturer guarantee or warranty has expired). There is therefore a need for a method by which an operator may detect such issues early, and thereby expand the useful life of the solar energy facility.

[005] When monitoring is carried out by existing SEFs, the monitoring is generally performed at a low level of sensitivity, and at a very low granularity - for example, a monitored statistic might be the power output from one string of panels. An operator might notice that there is something wrong with production when the power produced by the facility (or a given string of panels) drops off in magnitude, but it is not always possible to determine precisely what has gone wrong, or whereabouts in the facility the fault(s) has occurred. When photographs (including infra-red photographs or electro- fluorescent images) are taken of panels (e.g. by a drone or other suitable field-based or factory-based equipment), these are annotated manually and have low sensitivity; they are not optimally useful for purposes of geolocation, and they lead to very poor fault detection. Moreover, the use of photographs alone cannot predict upcoming faults or performance issues - it can only indicate that something has gone wrong post facto. Because such faults go undetected over long periods, they are often detected as a result of what are termed “catastrophic failures”, i.e. instances where panel function has ceased completely or almost completely. Earlier detection strategies may allow for failing performance to be detected and/or predicted up to several years in advance of such problems.

[006] The other problem with present monitoring systems, aside from the generally low granularity, is that SEF operators monitor only the absolute objective output values for the relevant panel performance metrics (e.g. power). This means that it is not possible to easily track long-term performance of individual panels in the array relative to one another, because one cannot make meaningful comparisons between snapshots of data collected at different points in time. For instance, comparing performance throughout a facility at 5pm on a sunny day in March against performance in the same facility at 1pm on a rainy day in July is likely to yield wildly differing figures for absolute performance values. This makes it difficult to detect when panels need to be repaired, replaced or reoriented.

[007] Recently, “solar micro-inverters” have emerged as a new technological means of transforming the direct current (DC) electricity produced by solar panels into alternating current (AC) at the output point of the panel itself, so that these AC outputs can be connected in parallel, rather than in series, before being exported to storage or to the grid. These microinverters are generally configured to each couple to a single solar panel; however, some models are known which can be coupled to more than one solar panel. Alternatively or additionally, micro-inverters may be built into the solar panels themselves, rather than being fitted as independent pieces of equipment. Micro-inverters can help to optimise performance at the level of the individual panel (or panels) to which they are coupled, by employing maximum power point tracking algorithms on a per-panel basis. Micro-inverters also have a secondary effect - namely, they are able to measure (or compute) and subsequently transmit one or more performance metrics, such as power, for their associated panel (or for each of their associated panels) as an output signal via a suitable medium. At present, these microinverters have typically seen use in domestic or very small-scale solar installations. Existing medium- or large-scale solar energy facilities continue to employ string series arrangements, and thus lack high-granularity monitoring capabilities (at best, the overall performance of a unitary “block” of panels can be monitored).

[008] In many cases, testing the performance of panels in a solar energy facility to locate and repair faults currently comprises sending technicians into the field to travel around a facility and test the panel outputs manually, which is an undesirably slow method; the panels may be spread around a very wide geographical area, particularly in facilities intended to power a national utility grid, so the ability to quickly determine the location of one or more faulty (or otherwise underperforming) panels matters hugely. An individual string series can vary from 4 to 72 panels and occasionally many more, and, in turn, multiple strings can be joined in a string connector or combiner box; it is therefore not uncommon for several hundred or even several thousand solar panels to be interconnected in series in some facilities. In turn, each individual solar panel may contain many (in some examples, between about 60 and about 96) solar cells and inter-cell junctions, each with the potential to fail and thus reduce that panel’s energy production capacity. Even as solar panel technology improves overall, these large- scale string series arrangements represent a considerable potential for fault occurrence over the lifetime of a typical facility (i.e. two or three decades). Finding the actual locations of faults (or impending faults) which may affect a particular solar panel in these agglomerations remains a major technical challenge. The challenge in accurately locating solar panels having faults (or impending faults) is further exacerbated when multiple faults occur concurrently within the same series, or within the same facility. In large-scale facilities, many hundreds or thousands of solar panels may be continuously or intermittently under-performing simultaneously at any time. These instances need to be detected in parallel, which is not possible with traditional string-mounted arrays. As the size of the facility increases, the risk of this happening increases dramatically, and current methods of fault detection and location for large-scale facilities are simply not able to adequately deal with such cases.

[009] Many areas of land, such as disused landfill sites, have the potential to be suitable locations upon which to build a solar energy facility but are not at present overly suitable, due to the unavoidable variability in substrate settlement and sedimentation rates and the inherent complexity of disused landfill geotechnics. In recent years, there has been increasing resistance to the use of agricultural land for the installation of solar energy facilities. However, sites which are based on landfill cannot be used for agriculture anyway, and so they represent an ideal opportunity for the development of solar energy facilities - provided that a solution can be developed to combat the fact that the differential settlement from these sites frequently causes panels to move out of optimal orientation (panel tilt), resulting in a loss of production from the facility as a whole, without it being possible to determine what has gone wrong. Indeed, the fact that unstable substrate settlement will clearly lead to the solar panels in a facility becoming incorrectly aligned is a significant technical barrier to the commercial exploitation of unstable land for the development of solar energy facilities. One possible explanation for the lack of uptake of micro-inverter technology for large-scale solar farming is that at first glance, one might incorrectly conclude that there is no need to monitor individual panel performance levels, particularly on stable ground; since facilities which are not based on disused landfill do not experience problems caused by differential settlement, the panels of such a “normal” site may be thought to already perform “well enough”. In such cases, the addition of solar micro-inverters for the purposes of performance monitoring at the level of individual panels (including the corresponding additions or changes to monitoring software which would have to be made) might seem needlessly expensive.

[010] Figure 2 depicts an example of an unstable substrate (such as a landfill site) undergoing differential settlement. The solar energy facility 204 on the left of the figure represents the ideal orientation of solar panels for solar radiation 200 incident thereon, and the solar energy facility 206 on the right of the figure represents what may happen over time when the solar panels are arranged on an unstable substrate. Landfill is unstable for several reasons - heterogeny of the waste material (horizontally, vertically and during the period of waste deposition), with different components of waste breaking down at different rates; the possibility of excess water percolation into the site; internal leachate pressures; engineering of landfill prior to waste deposition; pluviometry and resultant erosion; subterranean accumulations and flows of liquids and gases; different tensile properties and compressibilities of waste materials; variations in particle size of waste components - all of these factors play a part in affecting how the landfill settles. Substrates may remain highly unstable and unpredictable even after three to five decades of being left to settle; moreover, depending on the particular geotechnics at work, a substrate might settle down, well up, undergo twisting or shearing forces, or remain level.

[011] Modelling the settlement patterns of such unstable substrates in advance is possible, but highly problematic due to the geotechnical complexities involved. However, such predictions with respect to predicted changes in tilt angle for each solar module within a given facility (and thus energy generating potential) are well beyond the technical capacity of these modelled predictions. Consequently, it is not ordinarily possible to reliably predict future energy production rates for such facilities, or to ensure their long-term durability. At best, solar energy facility operators can attempt to react to geotechnical changes as they are occurring, or after they have occurred. Capped landfill sites present a particularly complex challenge for solar energy facility placement, due to traditional panel fixations not being able to penetrate more than 20-50cm below the surface. Since the panels cannot be properly anchored, their orientation is likely to be significantly impacted by differential settlement if and when it occurs. The above problems are compounded by the fact that, when arranged in the manner currently used by solar energy facilities (i.e. string series), once a single panel becomes disorientated from its optimal orientation, the entire series of panels falls back to the lowest common denominator, as discussed above.

[012] As such there is a need for a method or system which is able to address one or more of the above-identified problems. Summary

[013] This disclosure relates to a computer-implemented method for monitoring an array of photovoltaic modules, the method comprising: receiving a performance metric for each photovoltaic module in the array of photovoltaic modules; computing a numeric array based on the performance metrics, the numeric array comprising, at each array position, a relative performance score for one of the plurality of photovoltaic modules located at the corresponding position within the array of photovoltaic modules; wherein the relative performance score for the one of the plurality of photovoltaic modules is computed with respect to the performance metrics of a subset of the plurality of photovoltaic modules; and providing an output based on the numeric array. Following initial data stratification and data manipulation, sequential data analyses may allow levels of underperformance to be ranked. Any such ranking from the best to worst performing solar panels over a window of time may be conducted by purely nonparametric population analysis, namely, independently of population means, variance, and/or deviation from an expected or an observed mean value. This process may employ a comparator capable of encompassing milliseconds to decades of data and functioning independently of constantly-varying levels of solar irradiation.

[014] In one embodiment, providing an output based on the numeric array may comprise generating an image from the numeric array by using the relative performance score at each array position and adjacent array positions to assign a value for a pixel or group of pixels at the corresponding position in the image and displaying said image. Displaying the image may comprise displaying the image as either a synthetic grey-scale or a synthetic colour pixelated image. The method may comprise receiving a plurality of sequentially acquired performance metrics for each photovoltaic module in the array of photovoltaic modules and computing a plurality of numeric arrays based on the plurality of sequentially acquired performance metrics, wherein providing an output based on the numeric array comprises generating a plurality of images from the plurality of numeric arrays and displaying said plurality of images as an animation.

[015] In one embodiment, providing an output based on the numeric array may further comprise analysing the numeric array to detect one or more underperforming photovoltaic modules. Analysing the numeric array to detect one or more underperforming photovoltaic modules may comprise applying an image analysis algorithm to the numeric array. Analysing the numeric array to detect one or more underperforming photovoltaic modules may comprise applying a machine learning algorithm to the numeric array or an image or animation generated therefrom. Providing an output based on the numeric array may further comprise identifying a probable cause for the underperformance of the one or more underperforming photovoltaic modules or a numerical indicator of the future likelihood of a further reduction in performance. The probable cause for the one or more underperforming photovoltaic modules may be identified based on one or more of: a visual form of shapes detected within an image generated using the numeric array; a relative pixel intensity comprised of pixels corresponding to underperforming modules within an image generated using the numeric array; and the consistency or intermittent nature of shapes and spots present in an image generated using the numeric array. In embodiments where a plurality of numeric arrays are computed, the probable cause for the one or more underperforming photovoltaic modules may be identified based on the consistency or intermittent nature of shapes and spots present in a series of images generated over time using the numeric arrays. Providing an output based on the numeric array may further comprise prompting a user to repair, maintain, replace or adjust a component of the solar energy facility in order to improve a performance of the one or more underperforming photovoltaic modules at known locations within a solar energy facility. The method may comprise receiving an input from the user after providing said output; and adjusting how underperforming photovoltaic modules are detected and/or predicted based on the received input.

[016] In one embodiment, providing an output based on the numeric array may further comprise analysing the numeric array to predict future underperformance of one or more photovoltaic modules.

[017] In one embodiment, computing the relative performance scores may comprise, for each photovoltaic module, calculating an average value of the performance metric over a given sliding window of time. In embodiments where a plurality of numeric arrays are computed, this sliding window average may be calculated, for example, once a second, or several times per second.

[018] In one embodiment, the measured or computed performance metrics may be received for each photovoltaic module in the array of photovoltaic modules in parallel or within a predetermined time.

[019] In one embodiment, the subset of the plurality of photovoltaic modules may comprise photovoltaic modules having a performance metric less than or equal to a predetermined performance metric. The predetermined performance metric may be determined based on a plurality of well-performing photovoltaic modules during a predetermined period of time and may restrict the effect of variation in the level of solar irradiation observed during said period of time. The plurality of well-performing photovoltaic modules may have a predetermined variation in the level of solar irradiation. For example, the predetermined period of time might be a value of seconds in the range 1 to 60 seconds, a value of minutes in the range 1 to 60 minutes, or a value of hours in a range 1 to 24 hours. The subset may consist of the worstperforming X% of the plurality of photovoltaic modules in terms of absolute performance metric over a given period, for some value of X. This percentage may be in the range of 0 to 20%, 0 to 10%, O to 5%, or O to 2.5%, including, at least, 0.5%, 1.0%, 1.5%, 2.0%, 2.5%, 3.0%, 3.5%, 4.0%, 4.5%, 5.0%, 5.5%, 6.0%, 6.5%, 7.0%, 7.5%, 8.0%, 8.5%, 9.0%, 9.5%, and 10.0%. Using the worst-performing panels for the subset in this way produces the benefit of improved computational efficiency and reduced data processing requirements, because only the performance metric data from the most significant minority of under-performing modules is used in producing each array of relative performance scores. Computing the relative scores may further comprise mapping each performance metric of the subset of the plurality of photovoltaic modules to an integer value on a negatively-skewed linear scale, wherein the values of the linear scale range uniformly from a highest integer, corresponding to the lowest performance metric value recorded for said subset, to a lowest integer, corresponding to the highest performance metric value recorded for said subset.

[020] In one embodiment, the method may comprise receiving a plurality of sequentially acquired performance metrics for each photovoltaic module in the array of photovoltaic modules; and computing a plurality of numeric arrays based on the plurality of sequentially acquired performance metrics; wherein computing the numeric array further comprises generating stacked arrays by averaging and/or summing a plurality of numeric arrays representing performance metric values measured or computed over a predetermined period of time.

[021] In one embodiment, each numeric array and/or each image may be associated with one or more items of metadata relating to the measured or computed performance metrics from which the numeric array or image is derived.

[022] This disclosure also relates to a monitoring system for a solar energy generator, the system comprising: an array of photovoltaic modules; a plurality of measuring devices coupled to the photovoltaic modules and configured to measure or compute a performance metric for each photovoltaic module of the array of photovoltaic modules; and a processor configured to perform the method of any preceding embodiment.

[023] In one embodiment, the measuring devices may comprise one or more of a microinverter, a DC-current measuring device, and a temperature sensor. [024] In one embodiment, the measuring devices may be configured to continuously measure or compute the performance metrics at regular intervals.

[025] In one embodiment, the system may comprise a surge protection and/or automatic shutdown system.

[026] In one embodiment, the array of photovoltaic modules may comprise a plurality of solar energy generators geographically distributed, wherein each solar energy generator comprises a plurality of photovoltaic modules.

[027] In one embodiment, at least a portion of the plurality of photovoltaic modules may be situated on an unstable geological substrate, the unstable geological substrate comprising land which is prone to flooding, land which has a substantially high sand, loam and/or silt content, or a landfill site.

[028] In one embodiment, the measuring devices, inclusive of micro-inverters, DC/DC converters, temperature sensors (i.e. entities capable of measuring electrical parameters or temperature), and contact-less measuring devices, may be contained within a solar panel junction box and may be coupled therein with one or more of: an aluminium electrolytic capacitor; a battery-less connectivity module; a Bluetooth low energy transmission module; and a 3G, 4G and/or 5G telecommunications module.

[029] This disclosure also relates to a non-transitory computer-readable medium comprising instructions which, when executed on a computer, will cause the computer to carry out the method of any preceding embodiment.

[030] This disclosure also relates to the production of a standardised linear comparator of relative energy production efficiency for individual solar panels that can be manipulated mathematically and function independently of sunshine levels that vary constantly due to weather, time of day and season and that can operate over short (for example, milliseconds) or extended (for example, decades) periods of time via the computer-implemented method of any of the appended claims. When deployed in conjunction with micro- inverters, the claimed systems and methods obviate the need for string and central inverters in solar installations and afford operational systems resilience to solar power production due to inverter redundancy at the level of individual solar panels. The following unique features have been found either independently or in various possible combinations to deliver hitherto unheard-of levels of acuity for parallel fault detection, fault prediction, and fault location, even among up to several million solar modules: use of a standardised performance measure that can be employed independently of levels of solar irradiation, thus facilitating unified analysis over days, seasons, years, extremes of weather and climate and even solar eclipses; an initial triage based on overall population performance to focus on under- performing solar modules among a vast quantity of well-performing modules in order to reduce the complexity of subsequent analysis; a high level of statistical confidence underwritten by the Law of Large Numbers, with under-performance being defined using up to several hundred million measurements acquired by a control population of well-performing modules and employed to confidently detect even the most subtle levels of intermittent or continuing underperformance in power generation, this statistical confidence being further enhanced as the number of panels within a Solar Energy Facility increases and as the frequency and sensitivity of performance measurements increase; advanced image analysis conducted on summations of numeric arrays composed of sliding window averages, where each sliding window average is acquired over a short interval, but the averages are analysed collectively over longer periods to determine a fault type; and retrospective use of artificial intelligence, machine learning techniques (including convolutional neural networks) and/or principal component analysis on multidimensional datasets, so as to learn to mathematically characterise instances of fault occurrence very early, and thereby how best to predict repetition of similar faults in the future.

Brief Description of the Drawings

[031] Embodiments of the invention will be described, by way of example, with reference to the following drawings, in which:

[032] Figure 1 illustrates a traditional configuration employed in some solar energy facilities, in which solar panels are mounted in a string series arrangement;

[033] Figure 2 illustrates an example of differential settlement in an unstable substrate negatively impacting the ability of panels in an array to harvest solar energy; [034] Figure 3 illustrates an arrangement of photovoltaic modules and solar micro-inverters, DC/DC converters, temperature or other sensors in accordance with an embodiment of the invention;

[035] Figure 4 illustrates an example of a set of performance metrics being used to compute a set of relative performance scores in accordance with an embodiment of the invention;

[036] Figure 5 illustrates an example of a portion of a numeric array of performance scores in accordance with an embodiment of the invention;

[037] Figure 6 illustrates a sequence of steps of a computer-implemented method for monitoring an array of photovoltaic modules in accordance with an embodiment of the invention;

[038] Figure 7 illustrates a computer-generated image representing several faulty or underperforming solar panels in accordance with an embodiment of the invention, where the ‘Sea of White Squares’ corresponds to those modules that have performed well during the period of analysis or averaged analyses and the accompanying ‘grey-scale’ (bottom left) corresponds to degrees of underperformance or frequency of fault occurrence;

[039] Figure 8 illustrates a computer-generated image representing a rectangular array of intermittently under-performing solar panels caused by abrasions and moisture infiltration into cables connecting the affected panels in accordance with an embodiment of the invention, where the ‘Sea of White Squares’ corresponds to those modules that have performed well during the period of analysis or averaged analyses and the accompanying ‘grey-scale’ (bottom left) corresponds to degrees of underperformance or frequency of fault occurrence;

[040] Figure 9 illustrates a computer-generated image which may be indicative of a region in a solar energy facility having experienced differential settlement in accordance with an embodiment of the invention, where the ‘Sea of White Squares’ corresponds to those modules that have performed well during the period of analysis or averaged analyses, and the accompanying ‘grey-scale’ (bottom left) corresponds to degrees of underperformance or frequency of fault occurrence;

[041] Figure 10 illustrates a computer-generated image which may be indicative of differential settlement affecting a set of panels mounted on a single panel support system in accordance with an embodiment of the invention, where the ‘Sea of White Squares’ correspond to those modules that have performed well during the period of analysis or averaged analyses, and the accompanying ‘grey-scale’ (bottom left) corresponds to degrees of underperformance or frequency of fault occurrence; [042] Figure 11 illustrates an example of a graphical user interface of a computer system generating an output comprising a prompt for a user to rectify a solar panel fault in accordance with an embodiment of the present invention;

[043] Figure 12 depicts a simplified network of interconnected solar energy facility sites to illustrates how analysis, data and trends conducted, gathered or identified at a particular solar energy facility (small-scale domestic or large-scale commercial) may be used to improve monitoring capabilities at other solar energy facilities in accordance with an embodiment of the invention; and

[044] Figure 13 illustrates a scree plot showing the results of principal component analysis of solar panel performance data in accordance with an embodiment of the invention.

Detailed Description

[045] This disclosure presents systems and methods for solar panel array performance monitoring within solar energy facilities which enable and facilitate the detection and/or prediction of faults and performance issues which may affect photovoltaic modules or other important equipment.

[046] Figure 3 illustrates a solar energy facility 300 comprising a plurality of solar panels 302 arranged in a 2-dimensional (2D) array in accordance with an embodiment of the invention. In some embodiments, the array may be square or rectangular, or it might be a more unconventional shape. A solar energy facility may comprise only a single array of panels 302, or there might be multiple arrays in the same facility. In some embodiments, such as that which is depicted in Figure 3, solar panels 302 are precisely arranged in an ordered grid of rows and columns, but it will be appreciated that they need not be in such an exact arrangement. In preferred embodiments, the number of solar panels 302 in an array is at least one hundred, and could be in the order of thousands, or hundreds of thousands. In some embodiments, the solar energy facility may comprise single-axis and/or dual-axis solar tracking systems for the solar panels 302. In other embodiments, however, the panels 302 are fixed-angle installations. The panels 302 may be ground-mounted, or mounted on water or on roofing of any kind, including those deployed for shade creation.

[047] Figure 3 also illustrates a plurality of solar micro-inverters 304. In the figure, one microinverter is coupled to each solar panel. The (alternating-current) electrical outputs from the micro-inverters are connected 306 in parallel and exported to the grid via a transformer, to storage, or to some other suitable use 310. At the same time, outputs from the micro-inverters 304 representative of a performance metric measured or computed by the micro-inverter for its associated solar panel are transmitted to a software system 312 of the solar energy facility 300. In some preferred embodiments, the performance metric is a measurement of current and/or power, a metric which is frequently facilitated by solar micro-inverters. However, those skilled in the art will recognise that the performance metric or metrics could be any one or several of voltage, resistance, power, maximum power point, current frequency/amplitude, waveform or leakage or any other relevant metric. It should be noted that not all of the metrics in the above list can be directly “measured” per se, for instance the maximum power point metric, which is why the term “measured or computed” will be employed herein.

[048] Although Figure 3 depicts an embodiment in which there are wired connections 308 between the micro-inverters 304 and the software system 312, those skilled in the art will recognise that the data values measured or computed by the micro-inverters 304 may be transmitted to the software system 312 by any suitable means, including wired or wireless connections. In some embodiments, the micro-inverters 304 may be omitted and the one or more performance metrics transmitted to the software system for each of the solar panels may be measured or computed by other measuring means or devices, including DC-current measuring devices, DC/DC converters (DC power optimisers) and temperature sensors. In some embodiments, each micro-inverter is coupled to (and configured to measure and/or compute one or more performance metrics for) one solar panel, but in other embodiments micro-inverters may be coupled to 2, 4 or 6 solar panels each, by way of non-limiting examples. For example, the measuring devices may comprise any of the abovementioned measuring devices (i.e. entities capable of measuring electrical parameters or temperature), and contactless measuring devices, contained within a solar panel junction box (e.g. a single junction box, or a plurality of junction boxes), and the measuring device may be coupled in the solar panel junction box with one or more of: an aluminium electrolytic capacitor; a battery-less connectivity module; a Bluetooth low energy transmission module; and a 3G, 4G and/or 5G telecommunications module. The measuring devices may also include a separation of measuring and transmission tasks.

[049] As discussed above, the solar micro-inverters 304 may be built-in to the panels or solar panel junction box, or may be separate entities. The solar micro-inverters may be configured to convert DC to AC on an individual panel basis.

[050] The micro-inverters 304 are connected in parallel, such that the majority of electrical cabling throughout the solar energy facility 300 carries a significantly lower voltage (in some embodiments, less than 250V) than would be carried in the case of a DC string series connection, particularly for a series comprising a large number of solar panels. Power transmission in the present invention is therefore achieved with a much greater degree of safety than that achieved by string series arrangements. In some embodiments, the solar micro-inverters may use maximum power point tracking (MPPT) algorithms to optimise power production from solar energy on a per-individual-panel basis. The MPPT algorithms allow the micro-inverters to adapt to the individual characteristics of each solar panel, thereby avoiding mismatches between panels. This use of MPPT is beneficial because it ensures that the maximum power available from each photovoltaic module is exported, regardless of the performance of other modules in the array, which might experience performance drops caused by any of the issues described above. In some embodiments, the use of these micro-inverters affords the solar energy facility a concomitant surge protection and automatic shutdown capacity. In some embodiments, computations may be carried out by one or more processors in the solar micro-inverters; an MPPT algorithm is one example of such a computation. In some embodiments, absolute values of one or more performance metrics may be computed by one or more processors in the solar micro-inverters. In some embodiments, relative performance scores may be computed by one or more processors in the solar micro-inverters. However, in some embodiments of the present invention, some or all of the above computations may be carried out by a software system in the solar energy facility, which will be discussed in greater detail below.

[051] In some embodiments the array of solar panels, or at least a portion thereof, may be situated on an unstable substrate. This may be a disused landfill site, or a low-lying area of land which is susceptible to flooding, or an area having soil which is rich in sand, loam and/or silt. The landfill may in some embodiments be a capped landfill site, i.e. one having a pollutioncontainment geomembrane placed on the surface.

[052] In some embodiments, the performance metric used in the computer-implemented method is computed by using a “sliding window” to calculate, for each solar panel, the average output signal over the previous X units of time for some fixed X. For instance, in some embodiments, the performance metric for any given panel may comprise the average value of that panel’s measured output over the last five minutes. In some preferred embodiments, performance metrics are measured and/or transmitted at least once per second, optionally four or more times per second, and optionally many more than four times per second and may have a high level of sensitivity. The higher the number of measurements per sliding widow of time, the more likely it becomes that high levels of statistical confidence can be achieved when detecting even subtle levels of panel under-performance, as a consequence of the Law of Large Numbers. However, it will be appreciated that the benefit of the invention may still be realised even at much lower measurement frequencies. Sliding windows having relatively short lengths (such as sliding windows spanning only a few minutes) may be particularly suitable, because all of the measured values found in any such window will reflect an approximately similar level of solar irradiation. The use of values which are averaged over a sliding window can in some cases help to mitigate the effect of outlier measurements (which may be caused by instrument noise or other irregularities). In some embodiments, several performance metric measurements for a solar panel might be based on sliding windows which overlap or partially overlap; embodiments are also contemplated in which performance metrics are computed based on the averages of measurements taken over non-overlapping windows of time. It will be appreciated by those skilled in the art that the benefit of the invention may still be realised even in the event that sliding window averages are not employed at all.

[053] In some embodiments, continuous performance monitoring of the solar panels may be carried over a long time period, consequently generating and storing large volumes of data for future analysis. In some embodiments, performance measurements (which may be cumulative performance measurements) are taken at least once a day, preferably at least once every 3 seconds, and most preferably many times per second (e.g. 10, 20, 30, 40, 50, 60 or more times a second). In preferred embodiments, the measurements are sensitive to several decimal places for milliamp and/or Watt measurements. In some embodiments, the measurements are sent to the software system in real time as they are recorded. Embodiments are contemplated wherein measurements are stored on the micro-inverter devices before transmission to the software system, and/or wherein the software system polls the micro-inverters. The solar micro-inverters (or other relevant hardware) may be configured to continuously measure and/or compute their associated performance metric or metrics at regular intervals. In some embodiments, the measuring and/or computing of the performance metric or metrics by the measuring device may occur independently of their transmission to the software system, since there may exist energy-related advantages in separating signal recording from signal transmission for long-term systems durability. Alternatively or additionally the solar micro-inverters (or other relevant hardware) may be configured to transmit the measured and/or computed performance metric or metrics at regular intervals and / or as a function of availability of power generated by the SEF, as the latter undergoes significant diurnal variation.

[054] The software system may be configured to continuously receive measured or computed performance metric data from the micro-inverters or other hardware. In some embodiments, the software system may collect this data by polling the micro-inverters. The software system, in some embodiments, is configured to store, recover, visualise and/or display (via a graphical user interface) data representing values of the measured or computed performance metric for every panel in an array or for every panel in the solar energy facility, optionally coupled with local measures of e.g. weather. The absolute data values may be conveyed from the solar micro- inverters to the software system via a local area network employing Ethernet cables, internet-of-things and/or smartphone communication technologies such as 3G, 4G or 5G, by way of non-limiting examples. In some embodiments, the software system may use one or more hard disk drives for data collection and storage. In some embodiments, the software system may be linked to a standalone data collection and/or storage facility. A relational database or any other kind of database may be employed in connection with some embodiments, as will be understood by those having ordinary skill in the art.

[055] Making reference now to Figure 4, the absolute empirical values of the measured or computed performance metrics are processed as relative measurements with respect to the performance levels of the other solar panels. The absolute panel performance data is transformed into a plurality of relative solar panel performance scores. In some embodiments, the relative performance scores are transformed onto a uniform scale, for example, the range of integer values from 0 to 255. Prior to the transformation from absolute values to relative scores, a subset of the solar panels in the array may be identified, and used to determine the performance scores. In some embodiments, the identified subset is a subset of the solar panels having the lowest measured or computed absolute performance values of all those in the array. Such a subset may be identified by, for example, taking the worst-performing X% of panels in the array, for some value X, going by the measured or computed performance metric. In Figure 4 for instance, which represents an illustrative and non-limiting example, the worstperforming 2.5% of panels in the array have been identified as the subset, these panels being likely to be associated with statistical significance and distinct from a much larger population of well-performing panels. Of course, it will be understood by those having ordinary skill in the art that this threshold could be greater or lower, depending upon a range of factors, without departing from the scope of the present invention. For instance, the subset may comprise up to the worst-performing 15% of panels, such as the worst-performing 1 %, 2%, 3%, 4% or 5% of panels. The threshold may be manually altered if the solar energy facility contains less than a particular quantity of panels (e.g. less than 1000). In some embodiments, the subset may be identified by selecting all the panels whose output is more than a predetermined number of standard deviations lower than the mean of the population. In some embodiments where a subset of the array of panels is identified as the basis for generating the relative performance scores, and the subset does not comprise the entire array, those panels which are not members of the subset may be assigned “default” relative performance scores by the transformation, to ensure that it is well-defined over its whole domain. By way of non-limiting example, if the worst-performing 2.5% of panels are assigned relative scores in the range of integer values from 0 (the highest performance values in the subset) to 255 (the lowest performance values in the subset), all of the panels outside of this subset (i.e. the bestperforming 97.5% of solar panels) may be assigned a relative score of 0 as well. [056] Figure 4 depicts, in the uppermost row 400, a collection of absolute values of measured or computed performance metrics. The row is shown as being truncated, but it can be assumed that it contains 320 entries, one for each of the 320 solar panels in an array (this relatively small number being chosen purely for ease of illustration). The row below 402 depicts the result of extracting the lowest 2.5% of these absolute values (i.e. the values pertaining to the worst-performing 2.5% of the solar panels in the array). For the avoidance of doubt, each of the values in the second row is taken from somewhere within the first row, as this first step is merely a process of selection, filtering, stratification or extraction. The latter process dramatically reduces dataset complexity of otherwise similarly-behaving information (i.e. acceptable levels of module performance). A linear transformation is then applied to values in the second row, so as to map the lowest of the eight absolute values to the highest possible relative score (in this case, 255), and to map the highest of the eight absolute values to the lowest possible relative score (in this case, 0). Solving algebraically for these two known mappings, we can see in this case that since 255 = 1.4m + c and 0 = 16.2m + c, we get the

. 1275 , 20655 . . .. values m = - and c - - , yielding:

74 74 ^J °

1275 20655 which can be used to determine the remainder of the relative scores.

[057] As an optional last step, the values in the third row 404 may in some embodiments be rounded in order to obtain the “final” relative performance scores. Whilst this is not necessary to achieve the benefit of the present invention, rounding the values (for instance, to integers) may be particularly useful when employed alongside other optional aspects of the present invention, such as the generation of images from arrays of the relative performance scores. In this particular illustrated example, the values are being rounded to the closest or nearest integer; however, embodiments are contemplated in which values may be rounded up to the next integer (the mathematical “ceiling” function) or down to the next integer (the mathematical “floor” function). The fourth row 406 illustrates final (round) values. Finally, it should be noted that in the example depicted in Figure 4, the other 312 absolute performance metric values from the first row 400 not appearing in the second row 402 will be assigned a default value of 0 by the transformation.

[058] Embodiments of the present invention may comprise methods and systems which rely upon analysis of relative performance score data, either by computational techniques or by a user being provided with images or visualisations as output, to detect and predict actual and/or potential faults affecting solar panels and other equipment in and around the solar energy facility. Focusing on a particularly poorly-performing subset of the panels, and thereby limiting the analysis which must be carried out to examine only the scores from the very worst- performing panels, makes such analysis easier, more efficient, conceptually simpler, and vastly more effective. For instance, computer-implemented analysis based on artificial intelligence and machine learning techniques (such as those discussed below) will be computationally faster and more straightforward, because there is a smaller volume of data to work with, and also a lower quantity of noise in that data. In the case of embodiments which render one or more computer-generated images as part or all of their output, focussing on the lowest-performing subset of the panels for the creation of relative scores subsequently leads to it being less computationally intensive to render the necessary images for presentation to the user, particularly when summative animations containing many thousands of sliding windows are to be analysed; it also makes it easier for the user to perform the task of identifying underperforming panels (and their locations within the array). When referring to “underperformance”, what is generally meant herein is a comparative drop in performance relative to the correctly-behaving population of panels (for instance, the best-performing 97.5% of panels in the array), which effectively operate as a kind of yardstick or control group against which performance may be tracked.

[059] This use of a subset provides a method dependent upon the Law of Large Numbers whereby underperformance can be underwritten by high levels of statistical confidence and increasingly so as the number panels contained in an array increases, as the frequency of measurement increases and/or the level of detection sensitivity and its associated dynamic range is enhanced.

[060] In the exemplary embodiments described herein, specific methods of computing relative performance scores have been described, which comprise assigning each absolute performance metric value in a subset to an integer value on a linear scale, the values of the linear scale ranging uniformly from a lowest integer, corresponding to the highest performance value in the subset, to a highest integer, corresponding to the lowest performance value in the subset. In this way, the relative scores which are generated truly are “relative”, in the sense that they are representative of solar panel performance levels in comparison with the rest of the identified subset, on a unified scale independent of significant variations in the level of solar irradiation, rather than being attached to any objective measure by which solar panel performance might be assessed (such as the current or power output). Of course, as will be apparent to those skilled in the art, there will exist a range of other methods of computing relative performance scores for solar panels which do not depart from the scope of the present invention; relative scores may be computed based on an ordered ranking of the solar panels’ performance metrics, to give just one example.

[061] By using relative scores to monitor solar panel performance, rather than absolute performance data, it is possible to make meaningful quantitative comparisons independently across daily and yearly cycles of solar radiation, and across extremes of weather and climate, as the levels of solar irradiation increase and diminish in intensity. Comparisons can be made between data which is measured a few minutes after sunrise, and data which is measured in the midday heat; additionally or alternatively, one can meaningfully compare solar panel performance data gathered on a sunny day to that which is gathered on an overcast, foggy, or rainy day. It will also be possible to compare the present year’s performance data with data from a year to decades away in the future, when climate change may well have had a substantial effect on the measurements. Comparisons of data can be made across different seasons, latitudes, levels of elevations above sea level, and so forth. In this way, sets of measurements can be compared numerically on the same unified performance scale even though their absolute values may vary greatly, and might otherwise not be meaningful when placed side-by-side. Were one only to consider absolute objective measurements, it may be more difficult to analytically identify evolving patterns over time, because some data (e.g. values representing solar panel performance early in the morning, late in the evening, during the winter, or during adverse weather) would give rise to a considerably different-looking set of values to those recorded at noon on a sunny day. By standardising all of the measurements on a uniform, linear, relative scale, it becomes possible to introduce all or almost all of the collected performance data into an analysis, rather than merely of a subset thereof. The additional data which is provided in this way allows a user or computer analysis to draw more nuanced conclusions, to make more accurate predictions, and to detect faults earlier with a higher degree of statistical confidence.

[062] Referring now to Figure 5, collections of relative performance scores for solar panels in the array may be stored or represented in the form of one or more numeric arrays, in order to facilitate analysis. A numeric array may represent a “snapshot” of relative performance scores for each solar panel in the array; that is, it may comprise a collection of relative scores generated from absolute performance data values which were measured, computed and/or transmitted at the same or approximately the same point in time. In some embodiments, the relative scores in a numeric array may be derived from a set of absolute values which stem from sliding-window averages, as described above, where the start and end points of the sliding windows are the same or approximately the same. In some embodiments, numeric arrays may be generated by summing or averaging other numeric arrays, as described in more detail below. It should be noted that the term “array” is used here in the general sense of a two-dimensional data structure, and is not intended to limit the invention to embodiments whose implementation involves the use of an “array” within the particular meaning used in the context of any specific programming language. Those skilled in the art will appreciate that other data structures may be employed in order to implement the present invention, without departing from its scope.

[063] Positions within the one or more computed numeric arrays correspond to positions in the array of solar panels in the solar energy facility, and vice versa. A mapping may be established between the actual physical locations on the ground of solar panels (noting that, since solar panels perform at their best when given an unobscured view of the sky, two coordinates are sufficient to represent panel positions) and each numeric array. In some preferred embodiments, panels that are adjacent in real life will have their relative performance scores appear juxtaposed on each generated numeric array, and, likewise, scores at adjacent positions in each generated numeric array correspond to adjacent solar panels in the field. Of course, it is not always necessary that the array of solar panels be arranged in a rectangular formation, or aligned in a perfectly accurate grid arrangement, provided that it can be established which solar panels in the physical array correspond to particular relative scores in the numeric array. In some embodiments, the mapping between solar panels and numeric array positions may be a logical or simplified mapping, rather than one based purely on geographical truth. In some embodiments, a plurality of different solar panel arrays (which might be located at different sites or sub-sites) may be represented in one single numeric array.

[064] In some embodiments, each numeric array (or the outputs generated therefrom) may be associated with a set of contextual metadata relating to the performance metrics or snapshot in time from which the array or image is derived. By way of non-limiting example, the metadata may include auxiliary details such as ambient temperature; wind strength; saturation deficit; rainfall; ages of solar panels; sub-contractor(s) responsible for installation of panel supports, solar panels, cabling or connector boxes; dates of installation; panel types; and/or panel geographical location data. Any or all of this metadata may be used to gain additional insight into the data reflected in the numeric arrays, and thereby to enable more accurate and reliable fault detection and prediction capabilities, as will be discussed below.

[065] In some embodiments, a “stacked” numeric array may be computed, by computing the element-wise sum of a plurality of other numeric arrays. In some embodiments, these other numeric arrays may each be associated with sliding windows of time, as previously discussed. Similarly, a numeric array may be computed by taking an element-wise average of a plurality of other numeric arrays. For example, a numeric array may be obtained by computing the average of all of the numeric arrays of relative performance scores for a particular solar array over the course of the previous five years, which would yield an array reflecting the “overall” long-term relative performances of the solar panels contained therein. When working with numeric arrays based on sliding window averages, such an evaluation would be representative of the collection of all the five-minute sliding windows recorded over the five- year period. In some embodiments, panels may be represented based on the percentage of the total sum (from the stacked image) that they represent. For instance, each solar module may be represented visually or numerically as a percentage of the total grey-scale intensities measured for each and every module over the exemplar five-year period and visualised in a stacked composite image. Stacked or averaged arrays may be used as input for computer- implemented analysis techniques for fault detection and/or for fault prediction (such as those described below). Alternatively or additionally, stacked or averaged arrays may be output visually for display to a user (also described below) as a single composite image or an animation of each sliding window contained therein and over time. Whereas individual arrays representing instantaneous “snapshots” or representing data pertaining to a narrow timeframe may be very useful for identifying instantaneous or intermittent faults, stacked or averaged numeric arrays are highly-suited for the identification of panels having a level of performance which is never catastrophic, but may be consistently and continuously sub-optimal (for instance, a panel which is partially misaligned as a consequence of differential settlement of the ground beneath it) or intermittently under-performing. Stacking the numeric arrays can help to filter out or highlight the effect of brief “hiccups” over time to reveal the panels or groups of panels which are afflicted with the most serious long-term power production issues or those that underperform intermittently. As will be appreciated by those having ordinary skill in the art, the subset of worst-performing solar panels in any given array is unlikely to remain constant over time, and faults may arise due to all manner of causes; these may be intermittent, occur over an instant and persist thereafter, or become evident over time due to gradual aging and degradation of module performance.

[066] Figure 6 depicts a computer-implemented method 600 for monitoring an array of solar panels, in accordance with an embodiment of the present invention. As shown in Figure 6, once the computer has received the performance metrics for each of the solar panels 602, and computed one or more numeric arrays of relative scores based on the performance metrics 604, an output will be provided 606, based on the content of the one or more numeric arrays that have been computed. In some embodiments, providing this output may comprise generating one or more images or visualisations based on one or more of the computed numeric arrays, and showing these to a user via suitable hardware. For example, embodiments may include the generation of composite images by assigning greyscale intensity values to pixels or groups of pixels based on corresponding relative performance scores in a numeric array. In some embodiments, pixels corresponding to the highest relative score may be assigned black, and pixels corresponding to the lowest relative score may be assigned white, representing normal performance for a solar panel. Intermediate greyscale intensity values may be assigned accordingly based on the remaining relative scores in the array. In some embodiments, relative scores in an array could be used to generate an image by using the scores to assign colour properties such as the hue, saturation, brightness or alpha values of pixels or groups of pixels. Other ways of mapping relative performance scores to pixel colour values will be apparent to those skilled in the art. In some embodiments, each relative performance score of the numeric array may be used to generate an intensity value for a single pixel. In other embodiments, a relative performance score for a solar module may correspond to a group of several pixels. Embodiments are contemplated in which, by way of illustrative and non-limiting examples, each solar panel may be represented by a square of 4, 9, 16, or 25 pixels, or may be represented by a rectangular array of pixels.

[067] In some embodiments, a user may be presented with an image, representing an instantaneous snapshot, or representing the result of stacking or averaging multiple arrays and/or sliding window averages for each solar module in an array gathered over a given period of time, so that they may visually perform analysis of its content. The given period of time may span several minutes, or several decades, or any magnitude therebetween. In some embodiments the user may be presented with a plurality of such images, either concurrently or in sequence. Displaying the images in sequence may comprise presenting the images to the user one after another in succession, like individual frames of an animated film. Whereas individual images may be useful for the identification, by a viewer, of faults which have already occurred in the solar energy facility, such an animated sequence of multiple images might assist a viewer with the task of intuitively predicting when and where particular performance issues will occur in the near future, by “following the pattern” shown in the succession of images. For example, the sequence may assist the user in the prediction of faults such as those associated with improperly oriented panels caused by differential settlement of land. In some embodiments, artificial intelligence techniques may be applied to such successions of images in order to further enhance the user’s ability to extract nuanced detection or predictions of underperformance for a given dataset, namely, well-beyond the level of data complexity (series of evolving image sets) when intuitive prediction I analysis defies human intelligence.

[068] Reference is now made to Figures 7-10, in which exemplary illustrations of computergenerated greyscale images are depicted. The present invention enables the detection of a wide variety of fault types, including in many cases the identification of the fault “type” itself. That is, the information present in the numeric arrays and/or the computer-generated images of the present invention is sufficient to allow a human user or a computer program (such as an adequately trained machine learning application) not only to infer that a fault has occurred, or will occur imminently, but also to make precise predictions about what it is likely to be that has gone wrong, or will go wrong. Figure 7, for example, depicts varying levels of fault occurrence either constantly or intermittently and affecting several solar panels; the pattern shown in Figure 8 comprises a set of multiple under-performing solar panels in a rectangular formation, which is likely to be indicative of a fault (abrasion of interconnecting cables caused by loose panel attachments, for example) somewhere in one of the facility’s cabling or connector arrangements and accompanied by moisture ingress - this is evidenced by the lower intensity of grey-scale shown for the affected panels. The pattern seen in Figure 9 features a small number of very poorly-performing panels in a tight cluster alongside a larger number of slightly under-performing panels in an amorphous “blob” formation, all surrounded by a sea of white; this kind of pattern is likely to be a strong indication to the user that the under-performing panels have been affected by differential settlement of the underlying land. Figure 10 shows a gradient of underperforming panels sloping towards a lower-right corner, which a user or the computer software will be able to associate with differential settlement affecting one particular corner of a panel support system, for instance. By establishing a suitable mapping between positions in the numeric arrays (and in the computer-generated images) and locations of panels in the real world, it is possible to easily discern via computational methods or by visual inspection where in the solar panel array the fault lies, such that field technicians can locate one or more specific underperforming solar panels and investigate, correct, repair, replace or reorient panels, meaning that they can be found and fixed in very little time.

[069] In some embodiments of the present invention, the output provided by the computer system (see Figure 6) may comprise an indication that a fault has been detected and/or a prediction that a fault may occur at some point in the future. The indication and/or prediction may comprise additional information including but not limited to the location of the fault, the relative seriousness of the fault, the probable cause of the fault, and/or the degree of statistical probability with which the indication or prediction is made. In some embodiments, such an indication or prediction may be output in addition to one or more computer-generated images such as those described above. In some embodiments, the output may comprise only one or more images. In some embodiments, the output may comprise only indications and/or predictions of suspected or imminent faults. Other possible output combinations will be apparent to those skilled in the art.

[070] In those embodiments in which the output comprises an indication that a fault has occurred or a prediction that a fault may occur imminently, such indications or predictions may be generated by analysis carried out by a computer program.

[071] Regarding automated fault detection, it is often very difficult for operators of solar facilities employing string series arrangements of solar panels to detect that a single panel within a string is underperforming relative to other panels in the array, particularly if the only available monitoring data relates to overall “per-string” production data, or if the only available measurements are absolute values. Even when an instance of an under-performing panel is recognised, it is difficult to determine the physical location of said panel, and also to know what type of fault has occurred without inspecting and/or testing the panel in the field. Techniques from the fields of image analysis, pattern recognition and machine learning may therefore be applied to the numeric arrays generated in accordance with the present invention, in order to accomplish a variety of tasks. In some embodiments, these computer-implemented methods may be used to determine which panels are performing most poorly, and where in the solar energy facility these are located. Importantly, these methods facilitate the detection of potentially a multitude of underperforming modules occurring simultaneously in parallel. Alternatively or additionally, computer-implemented methods may be employed to classify the numeric arrays of relative scores based upon which type of fault or performance problem they are deemed most likely to represent.

[072] We now consider automated fault prediction. Predictive maintenance techniques aim to optimise the way in which components of systems are repaired, adjusted, reorientated and/or replaced by collecting and processing relevant data in order to estimate when such maintenance will need to be performed. These techniques are already employed in some commercial applications, such as the monitoring of performance of CPU cores in supercomputers and datacentres in order to predict core failures; evidently, such uses are rather distinct from the field of solar energy, and at the time of writing, whilst there exist some applications of preventative maintenance in solar energy facilities, these are relatively low- resolution technologies in comparison with the present invention. Nevertheless, predictive maintenance has proved to be surprisingly successful in the context of the present invention. One key question which may be asked of an automated fault prediction system is this: given all of the relative performance data available across all numeric arrays computed up to the present moment (and, optionally, also given the contextual metadata associated with these arrays), when is the next fault likely to occur, which panel or panels will it affect, and which category or type is the fault most likely to belong to?

[073] In the preferred embodiments described herein, due to the high granularity of monitoring, the high frequency, precision and sensitivity of the measurements, and/or the sheer quantity of photovoltaic modules present in a given array, the detection and prediction capabilities of the software system carry a high level of statistical confidence, thanks to the Law of Large Numbers. There will be statistically powerful differences between panels operating normally and those in the worst-behaving subset, particularly for arrays comprising very large numbers of panels. In the latter case, control populations of normally-behaving modules can produce many hundreds of millions of measurements or more over extended periods, thereby affording very high statistical confidence with respect to detections of even very minor decreases in panel performance. Without the high-granularity monitoring used by the present invention, alterations in performance caused by panel orientation or tilt would otherwise go undetected. It will be appreciated by those having ordinary skill in the art that the automated fault detection and prediction methods described herein will be most effective when carried out in conjunction with very large solar arrays, comprising very many panels; smaller arrays or domestic installations may not be capable of producing the same volumes of data that a large-scale solar energy facility can produce and, as such, may be less likely to enjoy the high degrees of accuracy and associated statistical confidence available to the fault detection or prediction software described herein that are afforded to such large facilities. Statistical analysis, machine learning and other artificial intelligence techniques are of course less effective in general when the quantity of available data is limited, and so the present invention is most effective in connection with these large-scale solar energy facilities.

[074] In some embodiments, providing the output comprises calculating statistics for solar panel performance, either in a given instant, or over a fixed time frame; these statistics may include values for mean, median, standard deviation, standard error, population variance or any other relevant statistic, be that parametric or non-parametric. Additionally or alternatively, classical statistical analysis tools may be employed in order to determine the statistical significance of a detected or predicted fault, in comparison with the null hypothesis - that is, the probability that the detection or prediction is the result of pure chance, caused by sensor noise, randomness, imperfections in equipment and/or bugs or inaccuracies in computer software or hardware. A variety of statistical methods may be used to analyse the one or more numeric arrays, including but not limited to analysis of variance, hidden Markov models, or Monte Carlo methods.

[075] In some embodiments, providing the output comprises using established image analysis techniques to analyse the composite images and hence find shapes, spots and forms which might be indicative of actual faults, potential future faults, differential settlement of the land underlying the array, or other problems giving rise to instances of under-performing panels. In some embodiments, edge detection within in all directions may be used to identify shapes, spots and forms, namely, a difference from one pixel to any, or all, adjacent pixels is determined using Fourier transformed data, for example, based on the two-dimensional arrays of relative scores. This may be performed to avoid the identification of a false positive, and may be used, for example, to detect land subsidence and for the area over which the ground has subsided to be mapped more accurately. “Adjacent”, with respect to pixels or array positions (which may be representative of solar panels), may refer to one or more of the eight pixels or positions surrounding a given pixel or position in a grid, i.e. the pixels or positions above, below, to the left of, to the right of, above and to the left of, above and to the right of, below and to the left of, or below and to the right of the given pixel or position. Other valid embodiments in which the adjacent pixels or positions are only the four orthogonally adjacent pixels or positions (i.e. above, below, left and right) or some other set of pixels or positions (e.g. for alternative pixel or position layouts, such as triangular or hexagonal grids) will be known to those skilled in the art. In some embodiments, the analysis may comprise the application of one or more machine learning techniques to an array (or several arrays) of relative performance scores. These machine learning techniques may include decision trees, support vector machines, regression or neural networks including convolutional neural networks. In one embodiment, for instance, a convolutional neural network may be trained on a dataset comprising a considerable number of two-dimensional arrays of relative scores, each array being labelled with data indicative of a particular type, cause or category of fault; the neural network will learn to accurately classify new arrays of relative scores and assigning each one a fault type or category based on its numeric content. The computer analysis may additionally or alternatively be configured to identify the existence and/or the geographical location of a fault from a given numeric array or image.

[076] In some embodiments, a convolutional neural network may be trained to identify or predict faults by examining a plurality of different numeric arrays or images (or animated films or sliding window averages contained therein) which represent a set of evolving snapshots of the performance of an array of panels over time. In one embodiment, for instance, a convolutional neural network may be trained on a dataset comprising a considerable number of sets of two-dimensional arrays of relative scores, each set being based on a series of absolute performance values for each solar panel in the array obtained at a set of different times. Each “set of snapshots” in the dataset may further be labelled with an indication of whether or not the evolving snapshots led to a fault or other cause of underperformance (such as differential settlement of the ground). By training the network on such a dataset, it will learn to read in a set of numeric arrays (i.e. a three-dimensional input) representing the evolving state of the array, and use it to accurately predict whether a fault somewhere in the array is imminent. In some embodiment, the labelling data may be further augmented by the inclusion of metadata related to observed fault locations, times of occurrence, and/or fault types, such that, when trained, the neural network is able to accurately predict how long it will be until the fault occurs, what kind of fault will occur, and where in the physical array of panels the fault is most likely to occur. Other suitable implementations of convolutional neural networks or other machine learning models will be known to those skilled in the art.

[077] In various embodiments, the computer analysis may conduct shape detection and/or spot detection via a range of methods in order to detect and/or predict potential faults based on the one or more numeric arrays of relative performance scores. In one embodiment, for instance, the computer analysis comprises applying centroid detection, followed by one or more edge detection and/or edge propagation techniques. Preferably, the computer analysis performs these techniques in multiple directions over the arrays (e.g. up, down, left and right). Most preferably, the computer analysis performs these techniques in all eight orthogonal and diagonal directions (i.e. up, down, left, right, and diagonally therebetween). This enables features of interest associated with the individual solar panels to be more easily detected against the local background of their adjacent solar panels. Shapes and outlines may be represented using Fourier descriptors and deployed against Fourier-transformed numeric arrays. Additionally or alternatively, the computer-implemented steps responsible for fault detection or prediction may comprise segmenting images into similar or dissimilar regions, to find areas of interest with respect to the local background region. Some embodiments may make use of region homology, contextual algorithms, thresholding techniques, or other similar methods known to those skilled in the art. The thresholding techniques may include detection of Gaussian filtered maxima. Once a shape or spot has been identified by one of the techniques described herein, the computer-implemented method may, as previously discussed, use the form of the shape or spot to infer the probable cause of a fault or predicted fault, particularly in the case of inferring types of differential settlement of unstable areas of land or cabling defects that intermittently affect groups of interconnected modules. The position within the array may be used to determine where the fault is, relative to the ground or to the rest of the panels. In the case of “stacked” images, the method may use the level of grey-scale intensity with respect to the summed grey-scale intensities detected for elements of the pixelated image over time to detect or predict faults and/or to infer their probable cause.

[078] Many of the methods and techniques discussed above enable and/or make use of spot detection with respect to the local background in an “image” (i.e. numeric array), which has particularly special significance for the present invention; on different physical slopes, for example, the levels of solar irradiation striking individual solar panels within a given facility at any given point in time, or within a given sliding window of time, may not be identical across the entire surface of the solar energy facility. The methods and techniques mentioned above are intended to eliminate the bulk of the otherwise overbearing influence of the absolute “local background” values.

[079] Referring now to Figure 11 , in some embodiments, the output which is generated based on the one or more computed numeric arrays of relative performance scores may comprise a prompt for the user to perform or arrange repairs, to fix, re-orient or replace a panel or other piece of equipment when the system detects a fault having occurred, and/or to preemptively carry out preventative maintenance when the system predicts that a fault will or is likely to occur soon. The user may be prompted to act on the information generated by the computer-implemented analysis, for example by tightening connections, moving panels, repairing or replacing cables, module connectors or combiner boxes, or taking any other appropriate steps in response to prompts which may be given as output. In the exemplary embodiment shown in Figure 11 , a display of the computer system 1100 illustrates a visual representation of the solar energy facility in the top-left 1102, in addition to a computergenerated image indicating a position of a solar panel within a particular array in the solar energy facility 1104 (bottom-left). Also depicted on the display is a set of location data pertaining to an identified solar panel fault 1106 (top-right), above the output of a computer analysis intended to identify the probable type or cause of the detected solar panel fault 1108 (for example, based on the output of a convolutional neural network trained to classify numeric arrays of relative scores or animated image sets). Embodiments may include either more or less information in the prompt than that described above in association with Figure 11 , and may differ in content, without departing from the scope of the present invention.

[080] Through facilitating timely replacement or reorientation of underperforming panels, the present invention is able to give rise to improved levels of long-term energy production, particularly throughout the second and third decades of a solar energy facility project. More dependable power production may be achieved, thanks to the system’s provision of an ongoing ability for operators to fix deficiencies on an individual panel basis, to improve performance over the entire project lifetime. Details such as energy production levels, rate of replacement of panels and other equipment, and necessary maintenance to the system may be reliably forecasted in advance, in order to keep energy production going for longer than would otherwise be possible. Because the present invention is able to inform the user of the geographical location of any detected or predicted fault, technical staff gain the ability to find the fault more quickly than would be the case if, for instance, the only known information was the identification of a fault having occurred somewhere within a string of panels. Moreover, because many embodiments of the present invention are able to output details relating to the probable nature of the fault, faults can also be rectified much more quickly than previously possible. By way of illustration, consider the opportunity to send one or more technical operatives into the field equipped with the necessary tools and equipment required to fix the specific identified fault type, in contrast with having to go into the array to locate the faulty panel(s), determine the nature of the fault, and then return later on with the appropriate equipment needed to rectify the fault. Those skilled in the art refer to the relevant measures of Mean Time to Detect (MTD) and Mean Time to Repair (MTR) which each contributes to the extent of lost energy production. The present invention is designed specifically to reduce both MTD and MTR. In some embodiments, prompts to users may additionally or alternatively utilise geographical and relative panel performance data, to generate and output an optimised panel cleaning programme in order to minimise wasted time, for example.

[081] After a fault has been detected, whether in the field or by computer analysis, retrospective learning techniques may be applied, analysing datasets with the benefit of hindsight in order to observe and learn the numerical traits or patterns that characterise the evolution of a given fault type over time, so that increasingly early and reliable predictions of said fault type may be made, reducing the negative impact of its occurrence in the future. In some embodiments, an empirically observed fault type may be logged in the software system and used automatically as training data for the convolutional neural network responsible for fault detection and/or prediction. In some embodiments, a user may manually update a model according to the detected fault and the previous output of the software system.

[082] Figure 12 depicts a simplified network 1200 of interconnected solar energy facility sites1202. In some embodiments of the present invention, data, patterns or trends collected and/or identified from one solar energy facility site 1204, whether automatically by software, or resulting from manual information entry or selection by a human user, might be used to detect faults at another site. As described above the present invention has been applied to a single array of photovoltaic modules. However, when using the data measured at several different solar energy facilities (or solar energy generators), the array referenced herein may be formed of several solar energy generators geographically distributed, each comprising a plurality of photovoltaic modules each varying in size and number of panels contained therein. In this way, the present invention can be applied to a housing estate, county or state, or even country, by combining individual solar energy facilities or generators mounted on individual houses, for example, to form one large solar energy facility collectively comprising an array of photovoltaic modules, which can be monitored using the methods described herein.

[083] With reference made now to Figure 13, principal component analysis (PCA) may be applied in some embodiments to determine or discover quantitative associations between the contextual metadata associated with instances of underperforming or faulty solar panels, and the subsequently arising variations in relative panel performance measurements. This application of PCA helps to establish a statistical link between such faults and the circumstances which are associated with them. For example, in the scree plot depicted in Figure 13, the first four principle components (the four leftmost bars in the plot) may relate to circumstances such as (for example) high rainfall, high wind speed, a long-past installation date, and a particular sub-contractor who was responsible for the installation. As can be seen from the plot, these four factors when taken together account for at least 90% of the variance in the data, yet when examined in isolation they are not statistically significant - embodiments of the present invention may therefore apply such an analysis to find insights and draw conclusions regarding possible correlations between measured panel performance and recorded contextual data.

[084] The term “comprising” encompasses “including” as well as “consisting” e.g. a composition “comprising” X may consist exclusively of X or may include something additional e.g. X + Y.

[085] Unless otherwise indicated each embodiment as described herein may be combined with another embodiment as described herein.

[086] The methods described herein may be performed by software in machine readable form on a tangible storage medium e.g. in the form of a computer program comprising computer program code means adapted to perform all the steps of any of the methods described herein when the program is run on a computer and where the computer program may be embodied on a computer readable medium. Examples of tangible (or non-transitory) storage media include disks, hard-drives, thumb drives, memory cards, etc. and do not include propagated signals. The software can be suitable for execution on a parallel processor or a serial processor such that the method steps may be carried out in any suitable order, or simultaneously. This acknowledges that firmware and software can be valuable, separately tradable commodities. It is intended to encompass software, which runs on or controls “dumb” or standard hardware, to carry out the desired functions. It is also intended to encompass software which “describes” or defines the configuration of hardware, such as HDL (hardware description language) software, as is used for designing silicon chips, or for configuring universal programmable chips, to carry out desired functions.

[087] Those skilled in the art will realise that storage devices utilised to store program instructions can be distributed across a network. For example, a remote computer may store an example of the process described as software. A local or terminal computer may access the remote computer and download a part or all of the software to run the program. Alternatively, the local computer may download pieces of the software as needed, or execute some software instructions at the local terminal and some at the remote computer (or computer network). Those skilled in the art will also realise that by utilizing conventional techniques known to those skilled in the art that all, or a portion of the software instructions may be carried out by a dedicated circuit, such as a DSP (Digital Signal Processor), programmable logic array, or the like.

[088] It will be understood that the benefits and advantages described above may relate to one embodiment or may relate to several embodiments. The embodiments are not limited to those that solve any or all of the stated problems or those that have any or all of the stated benefits and advantages. [089] The steps of the methods described herein may be carried out in any suitable order, or simultaneously where appropriate. Additionally, individual steps may be deleted from any of the methods without departing from the spirit and scope of the subject matter described herein. Aspects of any of the examples described above may be combined with aspects of any of the other examples described to form further examples without losing the effect sought. Any of the steps or processes described above may be implemented in hardware or software.

[090] It will be understood that the above descriptions of preferred embodiments are given by way of example only and that various modifications may be made by those skilled in the art. Although various embodiments have been described above with a certain degree of particularity, or with reference to one or more individual embodiments, those skilled in the art could make numerous alterations to the disclosed embodiments without departing from the scope of this invention.