Detection of High Impedance Faults The invention described below was developed in the context of a collaboration between Siemens AG and the faculty of Applied Sciences Bio- , Electro- and Mechanical Systems at the Free University of Brussels ULB (Universite Libre de Bruxelles ) under the Leadership of Professor Maun.

The invention relates to methods and devices for detecting High Impedance Faults (HIF) occurring in an electric distribution circuit that distributes a three-phase alternating current .

Background of the invention

The publication "Field Experience with High- Impedance Fault Detection Relays " (Alvin C . Depew, Jason M. Parsick, Robert W. Dempsey, Carl L . Benner, B . Don Russe11 , Mark G . Adamiak, 2006 IEEE) describes the efforts made by the Potomac Electric Power Company to reliably detect high-impedance faults .

The international patent application WO 95/10815 discloses a method of detecting high-impedance faults in further detail . A plurality of electrical signal analysis techniques is ap ^¬ plied that provide a number of fault indicators . Depending on the outcome of said fault detection indicators a signal indi ^¬ cating a high-impedance fault is generated or not . Under certain conditions the current of High Impedance Faults is lower than the residual current during normal operation of the network; hence overcurrent devices do not detect this fault . The difficulty of the detection depends on the con ^¬ figuration of the network, the worst being the multi-grounded distribution systems , which are the most common systems in America . Solidly grounded distribution systems in Europe a re grounded at a single point, the substation . This practice together with the use of three-phase transformers in the MV/LV substa ^¬ tions means that the neutral conductor under normal condi- tions carries barely a few amperes . In contrast, the typical configuration in America are multi-grounded systems using single-phase distribution transformers . This practice means that the current unbalance due to load switching is trans ^¬ ferred to the primary distribution system, producing impor- tant neutral current . The stray current consequence of the multiple-grounding also contributes in the level of neutral current .

The residual current in multiple-grounded systems (America) , is higher than in other configurations (in Europe) . The set ^¬ tings of the overcurrent protections are 10 or 50 times less sensitive than in protections in Europe, thus the HIF detec ^¬ tion is more difficult, and it cannot be performed by the sa ^¬ me detection functions (overcurrent technology) .

Ob ective of the present invention

In view of the above, an ob ective of the present invention is to provide a method and a device that reliably indicate a possible High Impedance Fault and avoid additional efforts in data analysing if a High Impedance Fault seems unlikely .

Brief summary of the invention

An embodiment of the invention relates to a method of detect ^¬ ing a high-impedance fault occurring in an electric distribu- tion circuit that distributes a three-phase alternating cur ^¬ rent , the method comprising the steps of applying a plurality of electrical signal analysis techniques that provide a num ^¬ ber of fault indicators , and generating a signal that indi ^¬ cates a high-impedance fault depending on the outcome of said fault detection indicators . The method further comprises the steps of determining the randomness of the residual current of said three-phase alternating current prior to determining said plurality of fault detection indicators , and generating a trigger signal depending on the randomness of the residual current, wherein determining said plurality of fault detec ^¬ tion indicators requires that said trigger signal has been generated .

An advantage of the present invention is that a time- consuming application of the plurality of electrical signal analysis techniques may be avoided if the occurrence of a high-impedance fault seems unlikely . To this end, the method analyzes the randomness of the residual current prior to ap ^¬ plying the electrical signal analysis techniques and prior to determining the plurality of fault detection indicators . De ^¬ pending on the randomness of the residual current , a trigger signal is generated or not . The further evaluation including the determination of said plurality of fault detection indi ^¬ cators may then be limited to cases when the trigger signal indicates a sufficient likelihood of the occurrence of a high-impedance fault . A further advantage of the present invention is that it ad ^¬ dresses the drawbacks of multiple-grounded distribution net ^¬ works like those presently used in America .

According to a preferred embodiment, a randomness value

(hereinafter also referred to as "AAD" ) is calculated that describes the randomness of the residual current . Then, the trigger signal may be generated depending on the randomness value . Further, a first threshold value (hereinafter also referred to as "AAD___threshold" ) may be calculated based on a given number of cycles that preceded the actual cycle wherein gen ^¬ erating said trigger signal requires that said randomness value exceeds said first threshold value .

Moreover, a second threshold value (hereinafter also referred to as "rand_AAD" ) that describes the average randomness of the residual current before the actual trigger cycle (during normal operation without high-impedance fault ) may be calcu ^¬ lated, wherein generating said trigger signal requires that said randomness value exceeds said second threshold value . Preferably, generating the trigger signal requires that said randomness value exceeds said first and second threshold value .

Furthermore , generating the trigger signal may also require that a reference value (hereinafter also referred to as "nor ^¬ ma1___AAD" ) that indicates the average randomness of the resid ^¬ ual current during normal conditions falls below a maximum randomness threshold value (hereinafter also referred to as "THLDnormai AAD" ) before the actual trigger cycle .

In the latter case, the trigger signal is preferably gener ^¬ ated if said randomness value exceeds said first and second threshold value and the average randomness of the residual current falls below the maximum randomness threshold value .

Preferably, the method also includes the steps of evaluating the increase of each phase current of said three-phase alter ^¬ nating current in response to the generation of said trigger signal , and determining that no high-impedance fault occurred if all three-phases of said three-phase alternating current exhibit a similar increase of current before or after the generation of said trigger signal . In most cases , high- impedance faults are very unlikely if all three phases of the three-phase alternating current show a similar behaviour .

Further, an average difference value (hereinafter also re ^¬ ferred to as " Iextr" ) may be calculated by subtracting a pre ^¬ vious average residual current value that defines the average residual current before the generation of said trigger sig- nal , from an actual residual current value that defines the average current after the generation of said trigger signal . The plurality of fault detection indicators is preferably de ^¬ termined if said trigger signal has been generated and the average difference value is between a predefined lower threshold value and a predefined upper threshold value .

A counter may be incremented if said trigger signal is gener ^¬ ated and the average difference value exceeds the predefined upper threshold value . The plurality of fault detection indicators is preferably de ^¬ termined if said trigger signal is generated and the counter reading equals or exceeds a predefined maximum count .

An further embodiment of the invention relates to a high- impedance fault detector capable of detecting a high- impedance fault occurring in an electric distribution circuit that distributes a three-phase alternating current, the de ^¬ tector comprising a computer being programmed to carry out the steps of : applying a plurality of electrical signal analysis techniques that provide a number of fault indica ^¬ tors , and generating a signal that indicates a high-impedance fault depending on the outcome of said fault detection indi ^¬ cators , wherein the randomness of the residual current (310) of said three-phase alternating current is determined prior to determining said plurality of fault detection indicators , wherein a trigger signal is generated depending on the randomness of the residual current , and wherein determining said plurality of fault detection indicators requires that said trigger signal has been generated .

Brief description of the drawings

In order that the manner in which the above-recited and other advantages of the invention are obtained will be readily un ^¬ derstood, a more particular description of the invention briefly described above will be rendered by reference to spe ^¬ cific embodiments thereof which are illustrated in the ap ^¬ pended drawings . Understanding that these drawings depict only typical embodiments of the invention and are therefore not to be considered to be limiting of its scope, the inven ^¬ tion will be described and explained with additional speci ^¬ ficity and detail by the use of the accompanying drawings in which

Figure 1 shows an exemplary embodiment of a high-impedance fault detector, and

Figure 2 shows a flow diagram of an exemplary embodiment of a method for detecting a high-impedance fault .

Detailed description of the preferred embodiment

The preferred embodiment of the present invention will be best understood by reference to the drawings , wherein identi- cal or comparable parts are designated by the same reference signs throughout .

It will be readily understood that the present invention, as generally described and illustrated in the figures herein, could vary in a wide range . Thus , the following more detailed description of the exemplary embodiments of the present in ^¬ vention, as represented in the figures , is not intended to limit the scope of the invention, as claimed, but is merely representative of presently preferred embodiments of the in- vention .

Figure 1 shows an embodiment of a high-impedance fault detec ^¬ tor 10. The detector 10 comprises a computer 20 having a microprocessor unit 30 and a memory 40. The memory 40 stores a computer program CP that may be carried out by the microproc ^¬ essor unit 30 in order to detect a high-impedance fault oc ^¬ curring in an electric distribution circuit .

The detector 10 analyzes the residual current 310 and the 3- phase currents I 1 , 12 , 13 of a three-phase alternating cur ^¬ rent and generates a signal ST indicating whether a high- impedance fault is likely ( "HIF" ) , possible ( "Possible HIF") or unlikely ("No HIF") . An exemplary embodiment of an algorithm that may be applied by the detector 10 of Figure 1 is depicted in further detail in Figure 2. The algorithm uses the three phase currents II- 13 and outputs the label of "HIF" , "No HIF" , or "Possible HIF" .

I f a high-impedance fault appears , an increase of randomness is expected, thus the algorithm monitors the randomness (see step 100 in Figure 2 ) and triggers when there is an important increase (see step 110 in Figure 2 ) . To this end, the current of the high-impedance fault is superposed to the residual current of the pre-fault situation, thus the algorithm re ^¬ moves the current before the trigger from the current after the trigger so the extracted current is the current of the event that produced the trigger (possibly a HI F, see step 110 in Figure 2 ) . The extracted current is analyzed and classi ^¬ fied as " HIF" or as "Other event" . Apart from this process there are other criteria that complement the algorithm. The final decision is made using information accumulated during a pre-defined period of time At decision . A complete descrip ^¬ tion of the algorithm is presented hereinafter in further detail . The inputs to the algorithm are the 3-phase currents 11-13 and, if available, the sensitive measure of the residual cur ^¬ rent 310. If the residual current 310 is not directly avail ^¬ able it is calculated by the sum of the 3-phase currents II- 13.

A randomness value AAD is computed for the residual current 310, as well as a first threshold value AAD___threshold and a second threshold value rand__AAD . The second threshold value rand_AAD is calculated based on a reference value normal___AAD that defines the average randomness of the residual current

310 during normal operation (see step 100 in Figure 2 ) . These magnitudes measure, respectively, the instantaneous random ^¬ ness of the residual current 310, the randomness of the re- sidual current 310 under normal conditions of the network, the level above which the residual current 310 is considered random and the level above which the algorithm triggers . The definitions and expressions of each magnitude are shown in the following Table 1 which includes definitions and expres ^¬ sions of the magnitudes used in the description of the algo ^¬ rithm:

Acronym Meaning Description/Equation

310 Residual Current The sum of the 3phase cur ^¬ rent

AAD Accumulated AAD measure the randomness

Absolute of the signal by quantify ^¬

Differences cycle ing the changes in the am ^¬ per cycle plitude and the content of non-harmonic components

Nacc

AAD = X 3 / 0 , 3/0 _{i + ¾7C}

k -

Nacc Number of samples Nacc determine the number accumulated of differences cycle per cycle that are accumulated for calculating the AAD . spc Samp1es per cycle normal AAD Value of AAD when It represents the random ^¬ the network ness of the residual cur ^¬ works under normal rent 310 during normal op ^¬ conditions eration of the network . It is calculated as the a er ^¬ age of AAD during 20 cycles during which the algorithm does not trigger . rand AAD Minimum AAD indicat ^¬ Value of AAD above which ing randomness the signal is considered random. It depends on normal AAD: if normal AAD<=

Ci*spc*Nacc, rand AAD= C ₂ *spc*Nacc if normal AAD>

Ci*spc*Nacc, rand AAD = normal AAD*2.5 Ci = 1, 25E- 4; C ₂ = 3, 125E-4

THLD normal AAD Threshold that indi ^¬ THLD normal AAD indicates cates the the ma imum level of norsuperior limit for mal AAD that allows the normal AAD algorithm to work correctly . Above this value, the randomness of the re ^¬ sidual current 310 under normal conditions is too random and the algorithm cannot work .

THLDnormal AAD = C3 , ^* C3 = 1E ^~

3 *spc*Nacc

AADmean mean value of AAD AADmean is calculated each during 5 cycles 5 cycles as the average value during this period

AAD Threshold Threshold for AAD It is a dynamic threshold, that determines updated each 5 cycles , when the algorithm which value is calculated triggers based on the avarage

value of AAD during the previous 5 cycles

Threshold = C4 * AADmean i-i c ₄ = 1,4 The main condition for the good performance of the algorithm is that the residual current 310 during normal operation of the network is regular or not random, so that normal_AAD is low . Therefore, the value of normal__AAD has to be checked . I f it is lower than a maximum randomness threshold value THLD normal___AAD then the residual current 310 is considered regu ^¬ lar enough and the algorithm for triggering runs . Otherwise, the algorithm breaks , indicating that the load of the network is too random.

The value of normal_AAD is updated several times per day in order to be adapted to the changes in the network . So the al ^¬ gorithm will be aware of the moments when the conditions of the network are so bad that high-impedance fault detection cannot be done .

The algorithm is designed to trigger when there is a change in the residual current 310 linked to an increase of random ^¬ ness . High-impedance faults cause changes in the residual current 310 and increase the randomness of the current, but also inrush currents or load switching activities do . The al ^¬ gorithm has to trigger in any of those cases , and later it will distinguish between high-impedance faults and other events .

There are two requirements for triggering : that the instanta ^¬ neous value of AAD is higher than the threshold AAD___threshold and that the value of the instantaneous AAD is high enough so it indicates randomness . The AAD_threshold adapts its value each 5 cycles of current . If the instantaneous value of AAD passes this threshold, it means that the random of the resid ^¬ ual current 310 at that moment has notably increased, because a change in the residual current 310 has occurred . On the other hand, the instantaneous value of AAD has to be repre ^¬ sentative, has to be higher than a minimum level of AAD that reveals randomness . This minimum level is rand___AAD, which is updated depending on the value of normal_AAD ( further expla ^¬ nation in Table 1 ) . When a trigger is produced the algorithm extracts the compo ^¬ nent of the current related to the change that made the algo ^¬ rithm trigger (see step 120 in Figure 2 ) . This current compo- nent , lextr, (hereinafter also referred to as average differ ^¬ ence value lextr) is analyzed in order to decide if it is re ^¬ lated to a high-impedance fault or to another event . The al ^¬ gorithm also considers some other cases : the triggers related to 3-phase events, the very low amplitude extracted currents , and the too high amplitude extracted current .

If the trigger is due to a 3-phase event (see step 130 in Figure 2), the event is not a high-impedance fault because high-impedance faults are single-phase faults . Therefore, af- ter each trigger, the algorithm obtains the extracted currents in each of the 3 phases (ΔΙ in 3 phases ) . They are cal ^¬ culated by subtracting the phase current before the trigger from the phase current after the trigger . ΔΙ in 3 phases represents the 3-phase current of the event that causes the trigger . I f the event is a single-phase-event , the extracted currents in two phases have to be negligible, and the e x ^¬ tracted current in one phase has to be similar to the ex ^¬ tracted current of the residual current 310, le tr. On the contrary, if the extracted currents in the 3-phases have similar amplitudes , the event is a 3-phase event, so it is not a high-impedance fault and the algorithm breaks and out ^¬ puts the label "No HIF" .

If the average difference value le tr is higher than THLD sup__Iextr the output is "No HIF". By definition, the amp1i- tude of high-impedance faults is low, e . g . between 1A and 70A-100A. In practice it needs to be considered that high- impedance fault detection is complementary to overcurrent protection, thus the maximum amplitude considered by high- impedance fault detection is the setting of the overcurrent protection . THLD sup___Iextr is given by the limit of the over- current protection of each network, and we estimate this value between 10 OA and 20 OA . If the amplitude of Iextr is lower than THLD inf_Iextr, the algorithm memori zes the trigger by increasing a counter by " 1 " (see step 140 in Figure 2), but the algorithm does not compute the classification of Iextr. Due to the inaccuracy of the current measurement and of the extraction method there is noise in Iextr. If the amplitude of Iextr is not much higher than the amplitude of the estimated noise, Iextr is consid ^¬ ered too noisy to be analyzed . However, the fact that the al ^¬ gorithm triggered is taken into account is meaningful . In case the event analyzed is a high-impedance fault the algo ^¬ rithm will trigger several consecutive times during a long period, which can be several seconds or even days . Conse ^¬ quently, the information of the numbers of triggers during a period of time At decision is an input for deciding if the event is high-impedance fault or is not .

If the amplitude of lext is between THLD inf_Iextr and THLD sup__Iextr the algorithm memorizes the trigger by increasing the counter by " 1 " , and Iextr is classified as high-impedance fault or as "Other event" (see step 150 in Figure 1 ) .

In case the event is a high-impedance fault the extracted current Iextr is the current of the fault, so it would have the typical characteristics of high-impedance faults (main harmonic the 3rd harmonic, phase of the 3rd harmonic constant around 180°, effect of the arc at the current zero- crossing...) . Therefore, a given list of indicators (for in ^¬ stance 14 indicators as listed in the following Table 2 ) that reveal the typical characteristics of high-impedance faults may be calculated from the Iextr. Using this input, the clas ^¬ sifier offers the output " HIF" or "Other event" . The output of the classifier is accumulated during the period of time At decision, and is used for taking the final decision . The fol- lowing Table 2 lists indicators and their characteristics in an exemplary fashion : Indicators Characteristics rms value Typical amplitude and limit

of algorithm

AADaverage Randomness

Fast change

Slow change

Main Harmonic 3 ^rd harmonic as main harmonic

Mean Amplitude of 3 ^rd component

H

Correlation 3 ^rd H/ fund Amplitude of 3 ^rd harmonic is

not proportional to fundamen ^¬ tal amplitude

Mean Phase 3 ^rd H Constant phase of the third

Standard deviation harmonic, close to 180

of Phase 3 ^rd H

Ratio 0 Cross Effect of the arc at the cur ^¬ rent-zero-crossing

Min ADPN Asymmetry between the nega¬

Mean ADPN tive and positive part of the

Ratio Alternance cycle

sign

Ratio H3 Asymmetry at the current- zero-crossing

It is evident that more or less indicators or other types of indicators than those listed in Table 2 may be used . The list in Table 2 represents a preferred embodiment, only .

The decision logic (see step 160 in Figure 2 ) indicates the final decision ( "HIF" , "No HIF" or "Possible HIF" ) based on the information of the numbers of triggers (see Table 2 ) and the output of the classifier during At decision (see steps 140 and 160 in Figure 2 ) . The output will be "HIF" if there were several triggers and a determined number of them were classified as high-impedance faults . The output will be "No HI F " if there were not enough triggers or if the number of them classified as " HIF" was lower than the minimum number needed for being suspicious of a high-impedance fault . And the output will be "Possible HIF" if there were several trig ^¬ gers and most of them were related to an Iextr lower than THLD inf_Iextr, or if the number of triggers classified as high-impedance fault was higher than the minimum number needed for being suspicious of a high-impedance fault but lower than the number that determines it as a high-impedance fault .

The extraction of the " suspicious event" as detailed above represents an important advantage compared to existing meth- ods . By removing the component of the residual current that is due to the background load, the current of the event is obtained that has ust appeared . So even if the current of the event is very low, it is extracted and analysed looking for characteristics of high-impedance faults .

The classification may be developed using data-mining techniques , and it can be improved as the database of residual currents in case of a high-impedance fault and residual cur ^¬ rents in case of other suspicious events is extended . The classifier may be a one-class classifier using a Support Vec ^¬ tor Machine . A Support Vector Machine may be trained and tested using a database of previous high-impedance faults and other events . Adding and removing data from the original da ^¬ tabase may be carried out to improve the classifier . An auto- matic system design for this function may be used . Some pa ^¬ rameters such as normal____AAD and rand_AAD are specific for each network and each moment, so the method may adapt to the customer . The design of the algorithm allows the possibility of future improvements that will be possible after testing the high- impedance fault detection method and increasing the training database . These improvements a re related to the definition of THLDnormal__AAD, to the extraction algorithm and to the data- mining technique .

Instead of defining THLDnormal___AAD as a constant (C3 = 1E-3 *spc*Nacc) it could depend on the amplitude of lextr. Con ^¬ cerning the extraction method, the calculation of lextr can be improved if the two currents that are subtracted (current before the trigger and after the trigger) are synchronized considering the possible error in frequency .

Related to the data-mining technique , the algorithm may use a one-class support vector machine with negative examples , but with a complete database it can be considered a two-class classification, such as random forest, decision rules