TITLE OF INVENTION APPARATUS AND METHOD FOR PREDICTION OF ALARM, IRREGULARITY

AND UNDESIRED MODE

TECHNICAL FIELD

[0001]

The present invention generally relates to a management system such as an energy management system predicting alarms, irregularities, and undesired modes (e.g. overheating of units, running out of fuel, malfunction due to electrical or thermal stress, breakdown due to aging components exposed to perpetual strain) and a technical system with middle or high complexity showing dynamic behavior over time and undergoing deviations from normal operation modes in a significant way.

In particular, the present invention relates to an apparatus and a method for prediction of alarms, irregularities, and undesired modes in a technical system.

BACKGROUND ART

[0002]

Complex technical systems are usually operated by supervisory control systems, and therefore they may cause a large number of alarms such as alarm signals showing the occurrence of danger or imminent danger. Multiple alarms may require suitable alarm management to be established. Commonly, dangers are risk of damage or loss of equipment, harm to people, risk of economic loss either due to operation in an unfavourable operation mode or the inability to provide contracted services to customers or the violation of existing regulations. Beneficial for proper alarm management are schemes (e.g. systems or apparatuses) that can predict alarms (e.g. irregularities or undesired modes) and provide suggestions or recommendations regarding countermeasures to prevent an alarm situation.

[0003]

Condition monitoring is a well-established field in the technological domain. Certain parameters are observed or estimated in order to determine the condition of a machine, an engine, or a plant. Detection schemes (e.g. detection schemes based on correlation) allow operators to detect a possible anomaly in industrial equipment as disclosed in NPL 1 entitled "An improved correlation-based anomaly detection approach for condition monitoring data of industrial equipment". In addition, predictive maintenance can be applied in order to prevent unnecessary outage periods of plants, machines, or engines as disclosed in NPL 2 entitled "An introduction to predictive maintenance". Predictive maintenance may incur additional costs that must be traded off against the monetary or other types of costs this special type of maintenance strategy involves; however, predictive maintenance has not necessarily proved to be advantageous. It depends strongly on the accuracy of the predictive, prognostic, or forecasting capability of the apparatus used for predictive maintenance applications.

[0004]

Technologies relevant to prediction of alarms, irregularities, and undesired modes are disclosed in the related art. For example, PLT 1 discloses an irregularity prediction system, namely an information monitoring and abnormality predicting system for a library. PLT 1 teaches a library system including a database, a prediction model, an alarm system, a terminal, and a server, wherein the alarm system provides a message alarm when a serious condition occurs; hence, PLT 1 is more related to an ICT (Information and Communication Technology) system. PLT 2 discloses a method of forecasting maintenance of a machine based on two or more parameter variation curves, wherein claim 1 recites "...measure a parameters of the machine, predict parameter variation curves each parameter variation curve representing parameter at a different confidence level..."; hence, PLT 2 may refer to the idea of prediction and uncertainty. PLT 3 discloses a system and method for determining regularity of respiration such as a radar-based physiological motion sensor, wherein an irregularity is detected by processing one or more frames of a respiratory waveform to obtain information regarding the regularity of respiration. PLT 4 discloses a system and method for generating alerts in a patient monitoring system, wherein this system monitors parameters or tracks data values so as to determine a range of data values with a first limit and a second limit and to emit an alert in the case of reaching the first limit or the second limit. PLT 5 discloses an online game irregularity detection method for the purpose of detecting fraud in online games with limited information (e.g. game logs). By using logging data and a sophisticated method, it is possible to reproduce or simulate games so as to check scores for plausibility.

[0005]

PLT 6 discloses a supervisory control system for a power system designed to detect an abnormality by way of state prediction for predicting plausible states via observation of a power system according to an evaluation function. PLT 7 discloses a model prediction control apparatus to prevent failure or erroneous operations in a controlled object such as a radiator cooling control system of a vehicle. PLT 8 discloses a feedback control apparatus using a servo motor by way of simulation using a machine model. [0006]

In any system having a certain level of complexity, it is difficult to carry out prediction since, apart from external variables, the concrete operation mode determines a possible occurrence of an erroneous event concerning an alarm, an irregularity, and an undesired mode. With a larger prediction horizon, it is difficult to exactly predict erroneous signals by simply replicating the behavior of a system involving a simulation setting, and therefore any methods based on pure time domain system operation simulation may fail to do so. The ability to predict an alarm, an irregularity, and an undesired mode allows for taking preventive or mitigating action in order to prevent erroneous events (or undesired modes). Alleviating the impact of erroneous events or informing the occurrence of erroneous events with a high probability is an important contribution to more economic, safe and efficient operation of technical systems.

CITATION LIST PATENT LITERATURE

[0007]

PLT 1 : Chinese Patent Application Publication CN 102609789A

PLT 2: United States Patent Application Publication US 2009/0037206A1 PLT 3: United States Patent Application Publication US 2010/0249633A 1 PLT 4: United States Patent Application Publication US 201 1/0298621 Al PLT 5: United States Patent Application Publication US 2005/0288103 A 1 PLT 6: Japanese Patent Application Publication No. 201 1 -24286

PLT 7: Japanese Patent Application Publication No. 2006-172273

PLT 8: Japanese Patent Application Publication No. 200-316937 NON-PATENT LITERATURE DOCUMENT

[0008]

NPL1 : Shisheng Zhong, Hui Luo, Lin Lin, Xuhun Fu, "An improved correlation-based anomaly detection approach for condition monitoring data of industrial equipment", 2016 IEEE International Conference on Prognostics and Health Management (ICPHM)

NPL2: R. Keith Mobley, "An Introduction to Predictive Maintenance 2nd Edition", 2002, Elsevier publisher

SUMMARY OF INVENTION TECHNICAL PROBLEM

[0009]

The prediction of alarms, irregularities, and undesired modes can only be valuable in different application areas if the "false predicted alarm rate" and the "unpredicted alarm rate" are sufficiently low, the prediction horizon is sufficiently large and if it makes the best use of available data considering modern computing and simulation technologies paradigms.

[0010]

The present invention aims to outperform conventional systems suffering from the following deficiencies.

(1) The conventional systems cannot make use of modern prediction, data-driven or simulation techniques for undesired mode prediction.

(2) The conventional systems cannot replicate operation strategies for the system at hand to improve prediction performance. (3) The conventional systems cannot use physical time evolution simulation combined with a prediction of operation characteristics for the system, and therefore they cannot make the best use of both domains such as time operation simulation and longer time operation characteristics or meta-variables prediction.

[0011 ]

These deficiencies may cause the drawback that the information at hand of the system cannot be used in the best way since an overall highly detailed physical model and exact prediction of necessary signals in the whole prediction horizon cannot be realistically available.

[0012]

It may be the best way possible to combine operation general characteristics prediction with operation time evolution simulation based prediction (or discrete time simulation based prediction) so as to exploit the special strengths of both domains. This is achieved by a special architecture and unit with special functionality in it for prediction and using two types of prediction in a unified undesired mode/irregularity predictor. The architecture allows for self-learning and improving their predictive power by use of a "prediction process rule base".

[0013]

The present invention aims to provide an apparatus and a method for prediction of alarms, irregularities, and undesired modes in a technical system such as an energy management system so as to outperform conventional systems by addressing their drawbacks.

SOLUTION TO PROBLEM

[0014] The present invention is directed to an apparatus and a method for prediction of alarms, irregularities, and undesired modes in a technical system having complexity. Namely, the present invention introduces an alarm/undesired mode predictor having the following features.

( 1) The alarm/undesired mode predictor combines two types of prediction in an advantageous way.

(2) The alarm/undesired mode predictor controls the sequence and data flow in the apparatus by using its own rule base (e.g. prediction process rule base).

(3) The alarm/undesired mode predictor outputs alarm/undesired mode occurrence information and associated necessary action to take in order to prevent or mitigate alarms.

[0015]

The present invention provides the two type of prediction (A), (B) as follows.

(A) The first type of prediction, called "operation-time evolution-based simulation", is based on the simulation of a technical system that is characterized by the computation of future commands with a reduced control model of the system and the application of this future commands in a simulation environment with a sophisticated simulation model considering also signals that have not been used for command computation.

(B) The second type of prediction, called "operation characteristics simulation", is based on characterizing parameters of the period to predict and either by using data- driven models (generated by a model generation unit) or by looking for similar (defined by a similarity index) periods (i.e. periods which characterizing parameters resemble the most with the period to be predicted), which derives the prediction of important parameters of the system in the considered period.

[0016] In a first aspect of the invention, an alarm/undesired mode predictor is devised for a technical system predicting an alarm or an undesired mode that deviates from normal operation. The alarm/undesired mode predictor includes an operation command prediction module configured to predict an operation control command for robust control of the technical system based on a first predicted signal concerning an operation policy of the technical system; an operation simulation module configured to generate a first prediction result affecting the normal operation of the technical system according to time-evolution based simulation based on the first predicted signal, the operation control command, and a second predicted signal concerning an operation schedule for the technical system; a model database configured to generate a simulation model based on a physical model and/or a statistical data-driven model for the technical system; an operation characteristics simulation module configured to generate a second prediction result according to operation characteristics simulation based on the first predicted signal and the simulation model; and an alarm/undesired mode detector configured to predict occurrence of the undesired mode based on the first prediction result and the second prediction result so as to cause a preventive action for preventing the undesired mode in the technical system.

[0017]

In a second aspect of the invention, an alarm/undesired mode prediction method is devised for a technical system predicting an alarm or an undesired mode that deviates from normal operation. The alarm/undesired mode prediction method includes the steps of: predicting an operation control command for robust control of the technical system based on a first predicted signal concerning an operation policy of the technical system; generating a first prediction result affecting the normal operation of the technical system according to time-evolution based simulation based on the first predicted signal, the operation control command, and a second predicted signal concerning an operation schedule for the technical system; generating a simulation model based on a physical model and/or a statistical data-driven model for the technical system; generating a second prediction result according to operation characteristics simulation based on the first predicted signal and the simulation model; and predicting occurrence of the undesired mode based on the first prediction result and the second prediction result so as to cause a preventive action for preventing the undesired mode in the technical system.

ADVANTAGEOUS EFFECTS OF INVENTION

[0018]

It is possible to achieve more robust and more reliable prediction for alarms, irregularities, and undesired modes in a technical system. In addition, it is possible to provide prediction for alarms, irregularities, and undesired modes in a highly complex system. Moreover, it is possible to provide the prediction for better planning of countermeasures and maintenance actions, in any technical systems such as a battery management system undergoing temperature fluctuations, a fuel tank management system undergoing shortage of fuel, and a maintenance system of machinery.

BRI EF DESCRIPTION OF DRAWINGS

[0019]

Fig. 1 is a block diagram showing the overall functionality of an

alarm/undesired mode predictor according to the present invention.

Fig. 2 is a block diagram showing the sub-functionality of a model generation unit included in a model database of the alarm/undesired mode predictor. Fig. 3 A shows an example of possible realization of feature selection in conjunction with two graphs showing original features and selected features by means of the model generation unit.

Fig. 3B shows an example of possible realization of model generation in conjunction with a graph showing the relationship between predicted values and actual vales of BTin (i.e. internal temperature) by means of the model generation unit.

Fig. 4 is a simplified block diagram showing control/simulation models in association with their corresponding modules extracted from the alarm/undesired mode predictor shown in Fig. 1.

Fig. 5 is a graph showing a SoC (State of Charge) pattern according to a control policy using parameters on a normal day (or a colder day under normal operation).

Fig. 6 is a graph showing SoC patterns along with an outside temperature profile on a hot day.

Fig. 7 is a graph showing SoC patterns along with an outside temperature profile on a very hot day.

Fig. 8 is a block diagram of the alarm/undesired mode predictor with the interpretation of two modules for the battery overheating situation.

Fig. 9 is a graph showing improvements of SoC patterns along with an outside temperature profile on a very hot day.

Fig. 10 is a graph showing improvements of SoC patterns associated with relevant signals along with an outside temperature profile on a hot day resulting a change between battery inside temperature profiles.

Fig. 1 1 is a schematic view showing a microgrid including renewable generation, storage, loads, and a control device in connection with an alarm/undesired mode predictor and an unreliable grid access.

Fig. 12 is a graph showing a tank level pattern indicating the development of a fuel tank level over time.

Fig. 13 is a graph showing comparison between tank level patterns varied due to the presence/absence of an alarm/irregularity/undesired mode predictor (AIUMP) over time.

Fig. 14 is a block diagram of an alarm/undesired mode predictor adapted to a fuel tank refill strategy using two types of predicted fuel trajectories y ₀ cbs and yotebs.

Fig. 15 is a graph showing a method for determination of similar patterns by use of appropriate distance measures in a feature space defined by feature 1 and feature 2.

Fig. 16 is an explanatory drawing indicating a mathematical method of mapping the distances of most similar patterns in order to determine a single weight and its similarity index needed to compute prediction signals.

DESCRIPTION OF EMBODIMENTS

[0020]

The present invention will be described in detail by way of examples with reference to the accompanying drawings, wherein parts identical to those shown in various drawings will be denoted using the same reference sings; hence, detailed descriptions thereof will be omitted as necessary.

[0021 ]

1. Framework of the invention

Fig. 1 is a block diagram showing the basic framework of the invention, i.e. the overall functionality of an alarm/undesired mode predictor 100 utilizing two types of predicted signals as follows.

(1 ) Type 1 : Predicted signals are used to compute robust operation control commands, e.g. predicted loads, predicted wind/PV (photovoltaic) generation, predicted grid availability, etc.

(2) Type 2: Predicted signals are not used to compute operation control commands, however, they affect some aspects of operations depending on system and system topology. They are not used to keep the operation policy (e.g. a strategy and algorithm) not too complex so that it can be executed on some restricted hardware in the battery charging depending on temperatures of the battery, casing, shelter, etc. They are related to some process that is not automated or only semi-automated (e.g. a refueling schedule of the BTS Diesel container).

Predicted internal signals (state sequence) are computed based on operation commands concerning a physical model and/or a hybrid model (incorporating physical and data-driven properties) since they depend on the operation strategy, policy, and mode of a system.

[0022]

The alarm/undesired mode predictor 100 includes an operation robust command prediction module 101 , a time-evolution based operation simulation module 102 (e.g. a CFD (Computational fluid dynamics) simulation module), a model database module 103, an operation characteristics simulation module 104, and an

alarm/undesired mode detector 105 implementing an operational logic with a prediction process rule database 106 and an output rule database 107. The model database module 103 includes a model generation unit 200, a physical model repository 1031 , a hybrid model (i.e. combined physical and data-driven models) repository 1032, and a statistical data-driven model repository 1033. Fig. 2 is a block diagram showing the sub-functionality of the model generation unit 200 including a data preprocessing unit 201 , a feature selection unit 202, and a core model generation unit 203.

[0023]

As shown in Fig. 1 , the alarm/undesired mode predictor 100 is supplied with two types of predicted signals (namely, type- 1 and type-2), wherein type- 1 refers to signals (or variables) used to compute operation strategy (or policy) parameters. The operation robust command prediction module 101 predicts future commands for the operation policy of the system. The operation simulation module 102 receives policy parameters so as to calculate the operation by use of typical time-evolution based sophisticated simulation considering the physical model of the system and the operation strategy/policy (with its associated policy parameters/commands from operation robust command prediction module 101) and predicted signals which have not been used to compute commands of the system. The reasons why type-2 predicted signals have not been used for command computation could be that they would make the problem of command computation too complex and too expensive computationally while under normal circumstances their impact on the system behavior is of minor importance or negligible. However, sometimes they can have a profound impact on the system behavior. Therefore, they are considered in the operation simulation module 102.

[0024]

The operation simulation module 102 outputs a sequence of predicted internal states. As opposed to the time-evolution based operation simulation module 102, the operation characteristics simulation module 104 considers the more data-driven aspect and/or the observation of the system on a longer time horizon or tries to find out periods with similar (similar to the period to be predicted) characteristics of some important parameters in the past.

[0025]

An example could be an EMS (Energy Management System) with a complicated interplay of renewable, storage and conventional fuel-based generation with complex generation and consumption patterns where the fuel consumption over longer time is predicted or its characteristics are predicted, which will be discussed later in conjunction with Fig. 11.

[0026]

In Fig. 1 , the alarm/undesired mode detector 105 has its internal functionality operational logic in connection with the prediction rule database 106 and the output rule database 107. The alarm/undesired mode detector 105 can control the operation robust command prediction module 101 , the operation simulation module 102, and the operation characteristics simulation module 104. Due to the ability and the internal knowledge (i.e. prediction/output ru le databases 106, 107) of the alarm/undesired mode detector 105, it is possible to simulate special meaningful (significant) days by way of the operation robust command prediction module 101 and the operation simulation module 102 as well as the results used in alarm/undesired mode detector 104 which takes the operation characteristics aspect into account. Herein, the separation between operation simulation/prediction and operation characteristics simulation/prediction could be typically achieved by time scale separation since, for example, thermal dynamics is usually slower than electric, the depletion of a tank for providing fuel for a generator is a slow process, typically slower than the thermal dynamic in and around the EMS.

[0027]

The case where the impact of type-2 predicted signals accumulates over time (i.e. their impact is of integral type) should not be neglected at this time. Therefore, they are considered in the operation characteristics simulation/prediction. The alarm/undesired mode detector 105 includes two types of rule bases, i.e. the prediction rule database 106 and the output rule database 107. The purpose of the prediction rule database 106 is to determine how the interplay of operation characteristics simulation/prediction and time evolution model-based simulation/prediction should be carried out according to the guideline for the operational logic. A simple example of a concrete realization of this interplay is given later in the description of the second embodiment configured to introduce a similarity index, wherein if the time interval to be predicted from the perspective of the interesting variables resembles more certain time intervals in the past, the needed input for the alarm/undesired mode detector 105 is composed from the operation characteristics simulation module 104 rather than from the time-evolution based operation simulation module 102.

[0028]

The output rule database 107 determines based on the prediction results, concerning alarms, undesired modes, and irregularities, what countermeasure should be taken in order to avoid the alarm state of the system. What this means in practice becomes more evident later considering the embodiments of this invention.

[0029]

The alarm/undesired mode predictor 100 has the model database module 103 serving as a model repository. The model database module 103 stores three types of models, namely physical models, hybrid models or combined/mixed physical and data- driven models, and pure statistical data-driven models. Herein, physical models, which are stored in the physical model repository 103 1 , are typically needed for the operation robust command prediction module 101 and the time-evolution based operation simulation module 102. In addition, combined/mixed physical and data- driven models and pure statistical data-driven models, which are stored in the hybrid model repository 1032 and the statistical data-driven model repository 1033 respectively, can be used for the operation characteristics simulation module 104.

[0030]

The model database module 103 further includes the model generation unit 200, details of which are shown in Fig. 2, Figs. 3A and 3B. Fig. 2 shows the sub- functionality of the model generation unit 200, which further includes a data preprocessing unit 201 , a feature selection unit 202, and a core model generation unit 203. The data preprocessing unit 201 is used for cleaning and resampling data from historical data and for organizing data concerning N-time ahead information. The feature selection unit 202 is used to pre-train the model to identify its important features so as to select the top features influencing the model most (where K is an integer arbitrarily selected). The core model generation unit 203 is used for fitting the model with reduced features and for providing the model and its model structure as the output of the model generation unit 200.

[0031 ]

Fig. 3 A shows an example of possible realization of feature selection with the feature selection unit 202, while Fig. 3B shows an example of possible realization of model generation with the core model generation unit 203. Fig. 3 A shows an example that the top fifteen features (where K=15) are selected from among numerous features with respect to outside/internal temperature and SoC (i.e. the state of charge) of a battery over time in conjunction with two graphs showing original features 301 and selected features 302 based on their importance scores. By preprocessing data, the relevant features 302 having higher scores are selected from among all considered features 301 and used for prediction of future signal values. Fig. 3B includes an example of a graph showing the relationship between predicted values and actual values of BTin (i.e. internal temperature) based on mapping feature vectors. Herein, reference numeral 303 shows the resulting true/predicted plots for a prediction with the reduced feature set involving the resulting weights (or parameters) described in the data-driven model.

[0032]

Fig. 4 is a simplified block diagram showing control/simulation models in association with their corresponding modules extracted from the alarm/undesired mode predictor shown in Fig. 1. Fig. 4 shows the distinction between a control model 401 (contained in the operation robust command prediction module 101 ) and a

sophisticated simulation model 402 (contained in time-evolution based operation simulation module 102). Herein, the control model 401 (e.g. a state-space model of low order) contains the most important aspects and is a lower complexity model for determining optimal commands, while the sophisticated simulation model 402 (e.g. a CFD model) is a detailed and accurate model of high complexity.

[0033]

Specifically, the control model 401 is typically a lean model containing the basic and necessary aspects of the technical system for the computation of right operation/control commands. The need for a lean model has its roots in the associated optimization/computation problem which should be feasible and have the property to be satisfactorily solvable (dependent on its application, it is necessary to suffice suboptimal solutions as well) in reasonable time or within a predefined time interval. On the other hand, the sophisticated simulation model 402 can be a comprehensive and very detailed model of the technical system (considering partial differential equation and stochastic partial differential equations) since it is not used for optimization but only for simulation purposes which are usually a straightforward computation with a low number of iteration needed (exemplary the solution of a differential equation by numerical integration is mentioned).

[0034]

It is possible to name various examples as technical systems such as a car, an engine, and an aircraft. The control model 401 used for computation of ideal commands in an electronic control unit (ECU) is a simple model since it is the model considered for real time command computation. This computation must be terminated within a certain time limit which is of the hard type. Heretofore, the framework of the invention is described for thorough understanding of the invention idea in practical use, which will be followed by the explanation of various embodiments.

[0035]

2. First Embodiment

The present invention can be exemplified using embodiments with respect to a microgrid system with a combustion engine generator and a fuel tank as shown in Fig. 11. Fig. 11 is a schematic view showing a microgrid 1101 in connection with an alarm/undesired mode predictor 1103 and an unreliable grid access 1105, wherein the microgrid 1101 includes a tank 1102, a combustion engine generator (or a diesel engine generator), a battery (or an energy storage) 1104, a control device, loads, and a photovoltaic (PV) unit. The first embodiment relates to the battery 1104 used in the microgrid 1101. Sometimes, power from an electrical power grid is available, but it is not assured that power can be sustained in any occasions. In the "no grid case", the battery 1104 is discharged to provide power for effective loads (wherein "effective load" = load - renewable power generation) or charged by a diesel engine generator and/or renewable power sources. Since renewable power sources and a diesel engine generator with nonlinear characteristics are available, it is possible to compute an optimal charge/discharge pattern. Usually, the computation of this pattern may not take into account the battery inside temperature, because it is too complex to consider in the computation of the optimal battery charge/discharge command.

[0036]

Next, the first embodiment of the invention will be described with respect to the prediction for battery-temperature undesired modes with reference to Figs. 5 through 10. Fig. 5 is a graph showing a SoC (State of Charge) pattern according to a control policy using parameters on a normal day (or a colder day under normal operation). Specifically, Fig. 5 shows a SoC pattern 501 for a "no grid" period (see straight lines). It is assumed that an optimal charge/discharge policy can be characterized by the parameters (e.g. pmin l , pmin2, pmin3, pmaxl , pmax2, si , s2). ^~ Considered here is the problem of battery inside temperature. If the battery inside temperature is higher than a certain threshold, charging is no longer possible or charging would excessively damage the battery. In addition, when the battery reaches a certain temperature, discharging is no longer possible or discharging would excessively damage the battery. Therefore, the event of reaching the relevant thresholds (concerning alarms, irregularities, and undesired modes) should be considered in the first embodiment.

[0037]

The role of the operation robust command prediction module 101 is to generate battery charge/discharge commands as they are expressed by the already introduced policy parameters (pminl , pmin2, pmin3, ...) minimizing diesel fuel consumption, guaranteeing general energy efficiency, low operation and maintenance cost for the microgrid 1 101. The operation robust command prediction module 101 could be realized using the alarm/undesired mode predictor 1 103. However, it would already be integrated in the operation/controller for the system. The operation robust command prediction module 101 generates commands based on a simplified model, e.g. the control model 401 shown in Fig. 4. The time-evolution based operation simulation module 102 is based on a more complex model, e.g. the sophisticated simulation model 402 shown in Fig. 4. In the case of the battery, it could be integrated nonlinear charge/discharge efficiency characteristics or the like.

[0038]

Fig. 5 shows a battery charge/discharge pattern based on the assumption about low outside temperature; therefore, the battery inside temperature does not affect the charge and discharging process of the battery. As opposed to Fig. 5, Fig.6 shows the situation on a hot day. Fig. 6 is a graph showing SoC patterns 601 , 602 along with an outside temperature profile 603 on a hot day. Herein, the SoC pattern 601 is associated with the low battery inside temperature, while the SoC pattern 602 is associated with the reduced battery charging power due to alarms without adaptation of policy parameters. The protection system may reduce the charging rate based on a threshold value of the battery inside temperature (where the threshold value leads to an event of alarm 1) in order to save the battery from damage. However, this may lead to non-optimal operation since optimal commands (e.g. policy parameters) are based on the assumption that the full charging rate is applied as shown on Fig. 5. This results in the SoC pattern 602 far from the "optimal" SoC pattern 601 in this situation. It is possible to still charge the battery at the end of the "no grid" period, but it is not good from a fuel saving perspective.

[0039] Fig. 7 is a graph showing SoC patterns 701 , 702 along with an outside temperature profile 703 on a very hot day. Specifically, the SoC pattern 701 is associated with low battery inside temperature, while the SoC pattern 702 is associated with the reduced battery charging power due to an event of alarm 1 and inhibition of battery discharge due to an event of alarm 2 without adaptation of policy parameters. On a very hot day as shown in Fig. 7, an event of alarm 2 could happen such that the battery cannot be charged/discharged at all. The timing at which battery inside temperature again reached a temperature enabling charging is marked as an event of alarm 3. In this case, the deviation of the SoC pattern 702 (i.e. an actual pattern) from the SoC pattern 701 (i.e. an optimal pattern) could be very pronounced at the end of the blackout of the battery still with high SoC value.

[0040]

Fig. 8 is a block diagram of the alarm/undesired mode predictor 100 addressing the aforementioned problem of the battery inside temperature alarm. Fig. 8 shows the interpretation of two modules 102 and 104 for the battery overheating situation. Herein, the time-evolution based operation simulation module 102 is used to simulate the SoC of the battery while the operation characteristics simulation module 104 is used to simulate the battery inside temperature (whose evolution is to some extent also dependent on the SoC).

[0041]

Fig. 9 is a graph showing improvements of SoC patterns 901, 902 along with an outside temperature profile 903 on a very hot day. Herein, the SoC pattern 901 is associated with low battery inside temperature while the SoC pattern 902 is associated with high outside temperature while using the alarm/undesired mode predictor 100 and its resulting recalculation of optimal policy parameters. Fig. 9 demonstrates the beneficial effect of the invention with respect to the improved SoC pattern 902 (i.e. the resulting battery charge/discharge pattern) compared with the SoC pattern 702 shown in Fig.7. Due to the alarm/undesired mode prediction, it is possible to prevent any alarm since the alarm/undesired mode predictor 100 outputs the rule of recalculating optimal patterns considering the reduced battery charging power. Then, the alarm/undesired mode predictor 100 is able to compute new policy parameters (e.g. s' l , p'maxl , p'min2, p'max2) and as soon as it predicts an impending alarm. By doing so, energy efficiency for fuel saving is kept optimal considering circumstances such as the reduced maximal battery charge power due to high outside temperature.

[0042]

Fig. 10 is a graph showing improvements of SoC patterns 1001 through 1003 associated with relevant signals along with an outside temperature profile 1007 on a hot day resulting a change between battery inside temperature profiles 1008, 1009. Herein, the SoC pattern 1001 is associated with low battery inside temperature; the SoC pattern 1002 is associated with high battery inside temperature without alarm/undesired mode prediction; the SoC pattern 1003 is associated with

alarm/undesired mode prediction and its resulting recalculation of optimal policy parameters. In addition, patterns 1004 through 1006 relate to diesel engine generator (DEG) power, wherein the pattern 1006 indicates diesel engine generator power resulting from commands for high battery inside temperature without alarm/undesired mode prediction; the pattern 1005 indicates diesel engine generator power resulting from commands for low battery inside temperature; the pattern 1004 indicates diesel engine generator power due to high battery inside temperature with alarm/undesired mode prediction and its resulting recalculation of optimal policy parameters.

Moreover, the battery inside temperature profile 1008 is made without alarm/undesired mode prediction while the battery inside temperature profile 1009 is made with alarm/undesired mode prediction and its resulting recalculation of optimal policy parameters.

[0043]

Fig. 10 shows more microgrid signals to get a better insight into the reason behind the fuel saving. Fig. 10 shows comparison between the SoC patterns 1001 through 1003, wherein the SoC pattern 1001 indicates a battery charge/discharge pattern on a cold day where no alarm is triggered; the SoC pattern 1002 indicates a battery charge/discharge pattern on a hot day along with the outside temperature profile 1007 triggering an alarm; the SoC pattern 1003 indicates a battery charge/discharge pattern on a hot day by way of the alarm/undesired mode prediction. In Fig. 10, the DEG power pattern 1005 corresponds to the SoC pattern 1001 ; the DEG power pattern 1006 corresponds to the SoC pattern 1002; the DEG power pattern 1006 corresponds to the SoC pattern 1004.

[0044]

It can be seen from Fig. 10 that, in the case of a hot day and no alarm prediction available, the DEG runs longest since battery cannot discharge due to high battery inside temperature. Due to the nonlinear nature of the power generation of the DEG, this situation can be translated into a highly inefficient use of the DEG. In addition, it can be seen that, due to use of the alarm/undesired mode prediction, the battery internal temperature profile 1009 is kept below the battery alarm temperature (see a dashed line). This bound is violated and therefor an alarm would be triggered without using the alarm/undesired mode prediction as shown in the battery inside temperature profile 1008.

[0045] It is possible to summarize the first embodiment concerning the prediction of the battery temperature undesired mode as follows.

(a) The problem of the battery inside temperature arises when the battery inside temperature is higher than a certain threshold and therefore charging is no longer possible or charging would excessively damage the battery. In addition, the problem arises when the battery inside temperature reaches a certain temperature and therefore discharging is no longer possible or discharging would excessively damage the battery. Therefore, it is necessary to consider an event of the battery inside temperature reaching the relevant thresholds.

(b) The role of the operation robust command prediction is to generate battery charge/discharge commands. The operation robust command prediction could be realized in the alarm/undesired mode predictor 100. However, it would already be integrated in the operation/controller for the system. This operation robust command prediction is used to generate commands based on a simplified model, e.g. the control model 401. The time-evolution based operation simulation is based on a more complex model, e.g. the sophisticated simulation model 402. In the case of the battery, it could be integrated nonlinear charge/discharge efficiency characteristics or the like. The role of the operation characteristics simulation would be to determine the battery inside temperature based on empirical values according to data driven "model less" or "model based" approach (or a regression model).

(c) The operation strategy considered for this system is exemplary shown to reduce the charging rate based on a threshold value of the battery inside temperature (reaching this threshold value would lead to an event of alarm 1. Typically, this event happens on a hot day and the battery has already been charged for a while. However, this event may lead to non-optimal operation since optimal commands are based on the assumption that the full charging rate is applied as shown in Figs. 5, 6, and the resulting pattern far from being optimal for this situation.

(d) On a very hot day, an event of alarm 2 could happen such that the battery cannot be charged/discharged at all (see Fig, 7). The timing at which the battery inside temperature again reaches a temperature enabling charging is marked as an event of alarm 3. In this case, the deviation from the optimal pattern could be very pronounced at the end of the blackout-cycle of the battery still with high SoC.

(e) The idea is here that the alarm/undesired mode predictor 100 can predict the occurrence of alarms so as to activate the corresponding rule derived from the output rule database 107. In this case, the charging rate is reduced in a pre-emptive way. The alarm/undesired mode predictor 100 recalculates for this situation the expected temperature, and if no undesired mode is detected, it would apply the corresponding command. In this case, considering the affecting temperature profile, it is possible to achieve the optimal operation.

[0046]

3. Second Embodiment

Next, the second embodiment of the invention will be described with respect to the prediction for an undesired mode (or an alarm) in a fuel tank running out of fuel with reference to Figs. 1 1 through 16. Fig. 12 is a graph showing a tank level pattern 1201 indicating the development of a fuel tank level over time. Herein, an alarm or an undesired mode is an event of the tank 1 102 of the diesel engine running out of fuel. Fig. 12 shows the typical tank level pattern 1201 according to a regular refill schedule over periods m[i-5] through m[i]. It can be seen that alarms or undesired modes (i.e. "no fuel" events) can be usually prevented by a regular refill schedule. However, in a period (marked with a star symbol) following the period m[i], an alarm or an undesired mode happens for a short time while due to no fuel available.

[0047]

Fig. 13 is a graph showing comparison between tank level patterns 1301 , 1302 varied due to the presence/absence of an alarm/irregularity/undesired mode predictor (AIUMP) over time. Fig. 13 shows an advantage of an alarm/irregularity/undesired mode predictor (AIUMP) to introduce the tank level pattern 1302 replacing the "original" tank level pattern 1301 . The AIUMP can predict an undesired mode 1304 (indicating a tank running out of fuel) occurring in the period (marked with a star symbol) in the tank level pattern 1301 ; hence, the output rule database 107 outputs the rule for "earlier refill at least 2 days earlier" so as to implement an earlier refill action 1303, thus preventing a tank from running out of fuel.

[0048]

Fig. 14 is a block diagram of an alarm/irregularity/undesired mode predictor (AIUMP), i.e. the alarm/undesired mode predictor 100 adapted to a fuel tank refill strategy using two types of predicted fuel trajectories yocbs and yotebs, wherein the AIUMP predicts the tank fuel level trajectory (TFLT) by combining two types of simulation/prediction so as to detect the possible occurrence of an undesired mode. Specifically, the operation characteristics simulation module 104 calculates the predicted fuel trajectory y ₀ cbs while the time-evolution based operation simulation module 101 calculates the predicted fuel trajectory 104 y ₀ teb _s . The alarm/undesired mode detector 105 controls the sequence and the control/data flow inside the AIUMP so as to calculate a single predicted tank fuel level trajectory y _pre d according to the following formula.

ypred = Wocbs + (1 ^~ Yotebs μ <≡ [0 ^j 1]

where μ denotes a similarity index (i.e. a value ranging between 0 and 1 ) and describes how well past periods of tank depletion resemble the period to be predicted. With the similarity index equal to 0, the prediction of the TFLT is only based on pure time-evolution based operation simulation. With the similarity index equal to 1 , a kind of weighted average of the most similar past situations is taken for prediction.

[0049]

Fig. 15 is a graph showing a method for determination of similar patterns by use of appropriate distance measures (e.g. di through ds) in a feature space defined by feature 1 and feature 2. Herein, reference sign 1501 denotes an area of relevance in the feature space of past periods; reference sign 1502 denotes an outstanding point in the feature space of the period which should be predicted with regard to the property interested therein; reference sign 1503 denotes a close point in the feature space indicating the past period with similar characteristics. Fig. 16 is an explanatory drawing indicating a mathematical method of mapping the distances of the most similar patterns in order to determine a single weight and its similarity index needed to compute prediction signals. Herein, reference sign 1601 denotes an interval to be mapped in a distance space, and reference sign 1602 denotes a resulting interval in a weight space.

[0050]

Figs. 15 and 16 show the details of a possible computational implementation. In order to predict the TFLT with the operation characteristics simulation module 104, feature signals of the period to be predicted are determined and compared with past situations, wherein feature signals are represented as a point in a two-dimensional space for the purpose of easy understanding. The closer a past situation is (closeness is evaluated with an appropriate distance measure), the more the interesting parameters of this situation contribute to the predicted TFLT. Due to a limit for consideration, the distance to the features of the situation (defined by its interval and period) to be predicted must be at least diimit for consideration. Fig. 16 shows how this computation can be achieved mathematically by appropriately mapping of the distance interval [0; diimit] ( 1601 ) to the weight interval [0; 1 ] ( 1602). In order to calculate the similarity index, the average weight is determined and a nonlinear function s() is applied with the main purpose of achieving a first degree of similarity even if not perfect similarity is achieved. The normalized weights are used for calculating the predicted fuel trajectory y ₀ cbs by combining the different situations in the past with higher priority given to more similar situations. In addition, the predicted fuel trajectory yotebs is calculated by simulating the operation of a microgrid (used for solving hybrid differential algebraic equations or the like). If past simulation shows too much difference (defined by some distance measure in the feature space), the prediction is carried out based solely on y ₀ tebs.

[0051 ]

4. Third Embodiment

The third embodiment example goes along the lines of the second

embodiment. The alarm/irregularity/undesired mode predictor (AIUMP) is used to plan and predict needed maintenance actions. In order to do so, the wear of machine parts is predicted by combining two methods (i.e. operation characteristics computation and time-evolution based operation simulation).

[0052]

Lastly, the present invention is not necessarily limited to the foregoing embodiments, which can be modified or changed in design within the scope of the invention defined by the appended claims. INDUSTRIAL APPLICABILITY

[0053]

The present invention is applied to the prediction of alarms, irregularities, and undesired modes which may possibly occur in energy management system and transportation system by use of two types of simulation in association with model database. The present invention is applicable to any other types of

management/maintenance such as electric power management and communication manag ement.

REFERENCE SIGNS LIST

[0054]

100 alarm/undesired mode predictor

101 operation robust command prediction module

102 time-evolution based operation simulation module

103 model database

1031 physical model repository

1032 hybrid model (combined physical, data-driven model) repository

1033 statistical data-driven model repository

104 operation characteristics simulation module

105 alarm/undesired mode detector

106 prediction rule database

107 output rule database

108 internal logic control & data signal

00 model generation unit

01 data preprocessing unit 202 feature selection unit

203 core model generation unit

401 control model

402 sophisticated simulation model

1 101 microgrid

1102 tank

1 103 alarm/undesired mode predictor

1 104 battery

1 105 unreliable grid access