Detection and location of faults in an electricity distribution system

Abstract

A method is disclosed for detecting early transient (pecking) faults in a low voltage feeder from an electricity substation. The method comprises: monitoring a voltage and / or current waveform by at least one probe at at least one point on the feeder; detecting an event of interest based on one or more characteristic(s) of the voltage and / or current waveform exceeding a predetermined threshold; recording a time series of samples of the voltage and / or current waveform associated with the event; applying the time series of samples to the input of an artificial neural network, the output of the artificial neural network corresponding to estimated distance of a candidate fault from the probe; and locating the candidate fault according to the output of the artificial neural network.

Description

[0001] DETECTION AND LOCATION OF FAULTS IN AN ELECTRICITY DISTRIBUTION SYSTEM

[0002] The present invention relates to systems and methods for detecting and locating faults in an electricity distribution network, particularly in low voltage feeders from substations to end users, especially when the feeders are installed underground.

[0003] BACKGROUND TO THE INVENTION

[0004] Electricity is supplied to homes and businesses by feeder cables from substations. A feeder will typically run from a substation, and supply a low voltage to perhaps a few hundred buildings along a few streets.

[0005] The substation is supplied by a higher voltage, typically 11kV in the UK. The substation includes a transformer to reduce the voltage to a low voltage supply (nominally 230V in the UK) suitable for safe use in homes and offices. The low voltage feeder comprises three phases and neutral, with homes typically being supplied with a single phase and neutral in an alternating pattern along the street (e.g. the first house is fed phase L1 and neutral, the second house is fed phase L2 and neutral, the third house is fed phase L3 and neutral, the fourth house is fed phase L1 and neutral, and so on). Buildings requiring a three-phase supply are fed all three phases and neutral. The alternating pattern may not be observed exactly, especially where supplies to new buildings have been provided over time. However the overall aim is to balance the load across the three phases as closely as possible.

[0006] In the simplest case, a feeder is provided in a radial configuration, i.e. the feeder runs from the substation transformer, along the street, until it reaches an end. However, it is common for feeders to have spurs, i.e. the feeder splits into two branches so that there are two (or even more) “ends” distant from the substation. Also, especially in large cities, a feeder may be connected to multiple substation transformers in a mesh (interconnected) network.

[0007] The configurations described are typical of UK distribution networks. However variations are possible and the configuration in some other countries is different. The invention is applicable for finding faults in low voltage feeders of various kinds.

[0008] Feeders are typically installed underground. If there is a fault in the cable, usually the road will have to be dug up. This is expensive and disruptive, and also takes time. While the fault is being repaired, homes and businesses may be without power. Faults detected early and repaired through routine maintenance are, on average, 25% of the cost of faults fixed as emergency repairs.

[0009] One common type of fault in an underground low voltage feeder is an early-stage transient fault, known as a pecking fault. This is often caused by water ingress into the cable, but more generally it is a transient insulation failure of some kind.

[0010] The common way a pecking fault will arise is that water will seep into the cable and joints due to a sealing failure, which can be induced through natural degradation, or when the joint was originally made, or civil engineering works induced ground movement, or fault in the design, or choice of cable, or jointing method. This causes an arc either between two phases or between a phase and neutral or a phase and ground. A very high current will flow for a very short period of time, but often at first the current is not high enough, or for a long enough duration, to blow the substation fuse. The arc evaporates the water and the fault then disappears for a period of time. Customers may notice a flickering of lights but the supply will generally continue. However, each time this happens damage may be caused to fuses, cables, and transformers. The watertight integrity of the cable, which is already compromised, is also likely to become worse over time, so that next time water seeps into the cable the fault current will be somewhat higher and / or the fuse is more likely to blow.

[0011] Eventually, the fault current will be high enough to blow a substation fuse. This will leave customers without an electricity supply until the fuse is replaced. A distribution network operator will send a crew to the substation to investigate the fault and replace the fuse. However, because the arc has evaporated the water which caused the fault in the cable, the report will often be “no fault found”. The fuse will be replaced and the supply restored, but the underlying problem will not be detected. The fault will inevitably recur, with repeated disruption to customer supply.

[0012] Typically, the fuse can be changed around five times before the fault becomes a permanent fault, where the fuse fails or another type of protection device activates immediately on replacement / reclose. Once the fault is permanent it can be located and repaired. However, customers will suffer a loss of supply on four or five occasions before this happens, and by the time the fault is found it will be so serious that a more lengthy outage is required to repair it before supply can be restored. This results in disruption to customers, and fines for distribution network operators. Existing technology to detect and locate faults generally works by sending a signal down the cable and detecting where it “bounces” off the fault due to the change in impedance at the fault, impedance to fault or Time Domain Reflectometry (TDR) techniques. These techniques can typically locate a fault to + / - 10-20m on a simple network with no spurs and this location can be further confirmed, provided an operative can attend before the gaseous arc products have dispersed. However, this may still require several holes to be dug to locate and repair the fault. In particular, where a feeder includes spurs I branches, the “distance from substation”, however accurate it is, will correspond to multiple physical locations and there is no way to distinguish between them. Moreover, this technique is only effective for fairly late-stage faults, typically when the substation fuse has already blown several times. In the early stages of the fault, before it becomes a “permanent fault”, once the arc has evaporated the water in the cable there is no significant change of impedance and no fault will be found. This means that faults are not fixed until they become serious, which in turn means that customers will lose supply while the fault is repaired.

[0013] Another problem with known detection and location technology is that the required devices generally require a power outage to install and to maintain which is required on a regular basis.

[0014] It is an object of the invention to solve these problems, in particular to provide methods and systems for detecting and locating faults more accurately, and sooner, before they cause power outages.

[0015] STATEMENT OF INVENTION

[0016] According to the present invention there is provided a method of detecting and locating an early transient fault in a feeder from an electricity substation, the method comprising: monitoring a voltage and / or current waveform by at least one probe at at least one point on the feeder; detecting an event of interest based on one or more characteristic(s) of the voltage and / or current waveform exceeding a predetermined threshold; recording a time series of samples of the voltage and / or current waveform associated with the event, from a point just before to just after the event simultaneously on all phases and neutral on all detectors on that feeder, applying the time series of samples to the input of an artificial neural network, the output of the artificial neural network corresponding to estimated distance of a candidate fault from the probe; and locating the candidate fault according to the output of the artificial neural network.

[0017] To practice the method of the invention, a probe or sensor needs to be installed on the feeder. Typically, a probe or sensor will be installed at a substation (the “beginning” of the feeder) and can measure voltage between each phase and neutral. Current can also be measured on each phase and on neutral. In some embodiments, current can also be measured on each phase and on the neutral where the voltage recording is taken from the bus bar driving all feeders; this allows identification of the feeder conducting the fault.

[0018] Such a probe can easily be installed on the feeder, without turning off the supply.

[0019] The probe continually monitors the voltage and / or current waveform(s) at the point where the probe is installed. Events of interest are detected according to a rules-based system. Typically, the rules-based criteria for detecting an event include some measure of difference in the waveform as compared to a typical or steady-state waveform. For example, a current difference of more than 200A for a duration of at least 1 ms. Difference metrics can be devised which compare the monitored waveform continuously, and output a “disruption” metric, which may depend for example on the magnitude and duration of a difference between the monitored waveform and a reference waveform. In some embodiments, the disruption metric may be calculated in aggregate across a number of waveforms, for example across three voltage waveforms in a three-phase system. In this case the disruption metric may depend on the aggregate of the magnitude and duration of the difference between each monitored waveform and a corresponding reference waveform.

[0020] In some embodiments, a “dictionary” of reference waveforms may be provided for continual comparison against monitored waveforms. Alternatively, a reference waveform can be continually updated as an average waveform over a few AC cycles based on previously monitored data.

[0021] Preferably, a first probe is provided at a first point on the feeder, and a second probe is provided at a second point on the feeder. Each probe is preferably provided at what can be considered an “end” of the feeder; however, in real installations there may be constraints on where probes can practically be installed which will lead to less optimal locations. By providing multiple probes, faults can be located even on more complex feeders with branches and spurs and even spurs off of spurs. Also, it has been found that providing multiple probes leads to a smaller uncertainty or ambiguity in the location output even on a straightforward feeder with no spurs.

[0022] On a radial feeder the current and voltage can be measured at the substation, where the feeder connects to the substation transformer. At the other end of the feeder (there may be multiple “other ends” where there are spurs of the feeder) voltage is measured.

[0023] A feeder in a “mesh” network may connect to multiple transformers at multiple substations. In this case current and voltage can be measured by probes at each substation.

[0024] Where multiple probes are provided, preferably the monitored waveforms are treated together in the intermediate stages of the method. In other words, the events of interest may be detected based on criteria applied to both of the waveforms measured by the different probes. For example, one characteristic of an event of interest may be a dip in voltage of more than a threshold magnitude and lasting for longer than a threshold duration, which is detected simultaneously (or within a predetermined short time period which may be for example 1 - 5ms) by probes located at different points on the feeder. Events of interest may be identified based on a combination of characteristic(s) crossing thresholds, and the number of places on the feeder in which the characteristic is observed to cross the threshold.

[0025] Where multiple probes are provided, preferably a time series of samples is recorded by each of the probes, and the multiple sampled waveforms are then applied to the same artificial neural network, which is configured to output a vector corresponding to an estimated distance of the candidate fault from each probe.

[0026] Preferably an ensemble of artificial neural networks is provided. The ensemble may comprise for example between around 10 and 100 neural networks. Each neural network is configured to accept input of a time series of samples from one or more recorded waveforms, and output an estimated distance or distances from one or more probes. The results from the ensemble of networks are then combined to estimate a most likely candidate fault location.

[0027] Each artificial neural network is preferably trained using training data generated by a simulation or “digital twin” of the feeder. The digital twin may be created in known simulation software, for example Simulink (RTM). The digital twin is preferably created with parameters which can then be randomised within constraints to produce training date. Pecking faults, or transient faults, are simulated at various points along the simulated feeder in the digital twin and the effect on measured voltage and / or current waveforms is recorded. An instance of training data therefore includes the measured voltage and / or current waveform(s), and the location of the simulated fault (i.e. the distance of the simulated fault from the simulated measurement point). Many millions of instances of training data can be created. Parameters of both the simulated faults and simulated network characteristics (e.g. line impedance, harmonics, loads) can all be randomised to create a large diversity of training data. Preferably, the diversity of the parameters used to generate the training data should be slightly wider than the extremes of the parameters likely to be seen in real life. Training data is therefore generated for a very large number of different simulated faults, and for a very large number of variations on the simulated feeder.

[0028] Preferably at least 200,000, and more preferably up to several million, instances of training data are generated in simulation.

[0029] Preferably the training data is then used to train a large ensemble of neural networks. For example around 300 or more neural networks may be generated. The hyperparameters of the neural networks (for example, the number of hidden layers, and the number of nodes per hidden layer) are randomised across the ensemble, within some bounds. The ensemble therefore comprises several hundred neural networks, each one having slightly different hyperparameters. All of the neural networks are then trained using the training data, and during the training process the networks are monitored for convergence. Neural networks which diverge are discarded. In a typical embodiment the result may be between around 10 and 100 trained and converged neural networks.

[0030] Preferably, the converged neural networks are then verified against data obtained by measurements on the real network. It is possible to create characteristics similar to those caused by a pecking fault on a real network. Suitable equipment can be connected to be network at various different locations (with measured distances from the probes) and the corresponding waveforms from the probes recorded. This is used as verification data to check the converged neural networks. Neural networks which do not show sufficient accuracy against this verification data may be discarded from the ensemble.

[0031] Verification using measurements of simulated pecking faults on the real network may also be used to refine the “digital twin” model. For example, if a systematic offset error in the location of faults estimated by the ensemble of networks is found, then it may be that the length of a particular cable run has simply been wrongly estimated from available data when creating the digital twin model. The digital twin model can be refined and an updated version used to generate new training data, which can be used to train a new ensemble of neural networks hopefully leading to improved results.

[0032] It is the remaining ensemble of neural networks, after dropping networks which do not converge and networks which perform poorly in verification, to which the time series of samples is applied in order to estimate the location of candidate faults (which may or may not be real pecking faults). The output from the ensemble of neural networks is a set of estimated distances between each probe and the candidate fault. This can be aggregated in various ways to derive a single most-likely location for the candidate fault, and to understand more fully the characteristics of the candidate fault, including whether it is in fact a pecking fault I early-stage transient fault. In some embodiments the most useful output may be to simply plot a histogram of the various estimates, from which an operator can visualise the range of the estimates and make a decision accordingly.

[0033] The decision made may be to dig in a particular place to further investigate and / or repair the fault. However, further investigation without digging up the road, and even without visiting the site, may be possible and useful. The output from the detection system is an estimated location of a candidate fault, that is, something which might be an early-stage transient fault or might be something else. It is found that where the ensemble of networks outputs a fairly wide range of estimated locations, it is less likely that the candidate fault is in fact an early-stage transient fault. A very narrow range of estimated locations, i.e. when a large number of the neural networks in the ensemble agree on an estimated location, is more likely to indicate an early-stage transient fault. However, at this stage collateral information may also be used to determine the likely cause of the event of interest. For example, a detected event which always occurs in the position of a restaurant and always occurs about an hour before the restaurant opens is likely in fact not to be a fault in the feeder at all, but rather a load event caused by switching of a large load. This can be determined without digging up the road, and potentially without even visiting the location if mapping data is good enough. Thus by accurately locating detected events, a better understanding of the causes of the event, and a better decision about the action required, is possible.

[0034] The time series of samples may be pre-processed before applying it to the input layer of the neural network(s). It has been found that pre-processing the time series by subtracting a reference waveform so that the time series presented as input represents a difference from the reference waveform, provides improved performance. The reference waveform is the steady-state waveform recorded by the probe, preferably just before the event was detected. The idea is to remove some noise which is consistently present on the feeder. In such embodiments, the same pre-processing will be applied both to simulated training data and measured verification data.

[0035] Embodiments of the invention have been found to successfully locate pecking faults to an accuracy of around ±3-5m. This is a significant practical and economic improvement on known systems with an accuracy of ±10-20m in terms of the amount of roadway which needs to be dug up to pinpoint and repair the fault. Moreover, this accuracy is achieved on early-stage transient faults, which are not even of significant magnitude to blow fuses. In many cases a repair can be carried out at this early stage without an interruption to the supply. At worst, the supply will only be interrupted while the repair is actually being carried out, rather than for the much longer period when a serious fault has become permanent and a crew needs to be dispatched, the fault located, and then repaired. The invention therefore allows much more proactive preventative maintenance of electricity distribution networks, leading to improved availability of supply and better service to customers.

[0036] The artificial neural networks may be convolutional neural networks. Other types of neural network, for example recurrent neural networks, have been tested as well, but convolutional networks provide the best performance. Each network may have for example between around nine and sixteen layers (the number of layers is one of the hyperparameters which is randomised across the ensemble).

[0037] BRIEF DESCRIPTION OF THE DRAWINGS

[0038] For a better understanding of the present invention, and to show more clearly how it may be carried into effect, reference will now be made by way of example only to the accompanying drawings, in which:

[0039] Figure 1 is a schematic of a typical feeder feeding a number of streets from a substation, being monitored by the method according to the invention;

[0040] Figure 2 is a diagram of part of a simulation of a feeder of the type shown in Figure 1 ; Figure 3 is a map overlaid with a plot of a feeder of the type shown in Figure 1 , and further overlaid with output of the method according to the invention showing located candidate faults; and

[0041] Figure 4 shows waveforms associated with candidate faults identified and located by the method of the invention, together with histograms showing the candidate fault location.

[0042] DESCRIPTION OF PREFERRED EMBODIMENTS

[0043] Referring firstly to Figure 1 , a schematic of a low voltage feeder 10 from a substation 12 is shown. The substation 12 can be considered as effectively a power source for the feeder 10. In practice the feeder 10 is connected to the secondary of a transformer which steps down the upstream voltage (for example around 11 kV) to a low voltage (nominally 230V RMS each phase to neutral in the UK) for distribution to homes and businesses by the feeder 10.

[0044] The feeder comprises conductors for each of three phases and a conductor for neutral. Depending on the system there may also be a separate earth conductor. For clarity the three phase and neutral feeder is shown as a single line in Figure 1. The feeder, especially in an old system which has been repaired, expanded and upgraded over decades, may comprise several different links, possibly using different cable types, jointed together. In the simplest case the feeder runs in a single line, for example along a street to the end of the street, but in most cases more complex arrangements are found. As shown in Figure 1 , there may be a spur 12 off the feeder, and a spur 14 off the spur 12.

[0045] The feeder serves loads, i.e. houses, commercial buildings, and anything else that needs an electricity supply. Typically, in a mainly residential street, most of the loads will be single phase and will alternate across the three phases. On other words, the house 16a may be fed phase L1 and neutral, the house 16b may be fed phase L2 and neutral, the house 16c may be fed phase L3 and neutral, the house 16d maybe fed phase L1 and neutral, and so on. Some loads, usually larger or commercial buildings, will be connected to all three phases.

[0046] The current and voltage is monitored by probes, in this case by two probes. Probe 18 is provided at the “beginning” of the feeder, i.e. at the substation. Probe 18 measures current in each of the three phases, and neutral current. Probe 18 also measures the voltage between each of the three phases and neutral. A second probe 20 is provided at the “end” of the feeder. The second probe measures voltage between each of the three phases and neutral (there is typically no current to measure since it is at the end of a radial feeder).

[0047] Note that, as shown in this example, there may be several positions which could be considered the “end” of a radial feeder, because there are spurs and spurs off of spurs. For example, the second probe could equally have been positioned in the position marked A or the position marked B. However, two probes are found to provide sufficient accuracy in many embodiments - it is not necessary to provide a probe on every end.

[0048] It may not even be strictly necessary to provide a probe at an “end” at all, for example a probe might be installed at either of the positions marked C or D. Usually installing probes at “ends” is most effective, and usually installing probes at “ends” will be easier due to access to the cable to install them in the first place. However in the wide variety of different installations there will inevitably be exceptions where for some reason it is more convenient to install a probe, for example at the position marked D.

[0049] In some installations, “mesh” rather than “radial” feeders are used, where feeders run from one substation transformer to another substation transformer, serving loads along the way. In other words, there may be two, or even more than two, power supplies on one feeder. Apart from the fact that it is possible to measure nonzero current at more than one end of such a feeder (information which may or may not be used in embodiments of the invention) this makes little difference in practice to the way the method works.

[0050] The feeder 10 is likely to be buried in the ground. Water can seep into the cable through old, damaged or worn insulation, joints, etc. This can cause an arc between conductors, which will cause a transient high current condition and a very brief dip in voltage. This is unlikely to cause a serious problem when it first happens, and beyond for example a brief flickering of lights may normally go unnoticed. However it is an early indication of a fault which will only become worse over time, and eventually may cause a lengthy outage. The object of the method according to the invention is to detect these faults at an early stage, and locate them so that remedial action can be taken, e.g. by digging up the road (but only a small section of road since the fault is accurately located) and replacing a section of cable (again, accurate location of the fault will mean only a short section needs to be replaced). In order to achieve this, a simulation or “digital twin” of the feeder is produced. Figure 2 shows a partial model of a feeder produced in the Simulink (RTM) modelling software. The model includes the power source 22 (i.e. the substation transformer and all the upstream components of the electricity supply). The power source 22 is modelled as an ideal AC source 24 plus an inductance 26 and resistance 28. Measurements can be made on the real power source to ascertain suitable values for the inductance 26 and resistance 28 of the supply. Each phase of the three phase supply has an inductance 26 and a resistance 28 and in principle the supply impedance is not necessarily the same for all three phases, although in most cases it would be expected to be similar.

[0051] A harmonics block 30 is also provided as part of the model, or digital twin model, of the power source 22. The harmonics block introduces coherent noise onto the supply, which is found on real supplies and has various causes, both upstream and downstream of the substation. In the model it is simply treated as a characteristic of the power source 22. The harmonics again can be measured on the real supply and the parameters of the harmonics block 30 in the model can be set accordingly.

[0052] A probe 32 is provided in the model, just as it is in the real network (see Figure 1). The probe 32 measures the voltage between each phase and neutral, and measures the current in each phase, and measures the current in the neutral conductor. The probe 32 is modelled immediately next to the power source 22. A further probe (not shown in Figure 2) is placed at the end (or at a point which can be considered an end) of the feeder in the model.

[0053] Sections of cable, i.e. transmission lines 34a, 34b, 34c are modelled. The characteristics of the feeder sections can to some extent be measured and can also be estimated from plans and other information. The type of cable, its age, its length, etc., may all affect the parameters of the model. The model used may be the “pi model” which is known in the art and is built-in to known simulation software, for example Simulink (RTM).

[0054] Loads are modelled. A load block 36 may in practice correspond to a group of buildings rather than a single building in some embodiments. Loads have a resistance, inductance and capacitance and obviously in practice the load presented by a particular building or group of buildings will change quite a lot over time. Finally faults are modelled. A fault block 38 simulates an early transient (“pecking” fault), i.e. a short-duration low resistance path between two phases, or a phase and neutral, which then quickly disappears.

[0055] In some embodiments, the pecking fault simulation may be enhanced. In such embodiments the simulation is enhanced to provide simulations of more than one pecking fault occurring within the sample window. This phenomena has been observed to occur on live networks.

[0056] The simulation can be used to generate training data. Preferably, hundreds of thousands or even millions of rows of training data are generated. In successful embodiments, the inventors have found that between 200,000 and a few million rows can be produced. Each row represents a simulation run over for example five or six mains cycles (i.e. about 100ms at 50Hz). In each simulation, there is a simulated pecking fault at a particular location (expressed as a distance from the or each probe). Across the many different simulations, many different pecking faults at different locations will be represented. Also, the parameters of the model are randomised within bounds, so the many different simulations will represent many slightly different models with different parameters.

[0057] The parameters which particularly may be randomised include the source impendence, the harmonics block, the load and fault parameters and the number and proximity (in time) of pecking faults. Since neural networks are generally found to be better at interpolation than extrapolation, it is preferable to randomise the model parameters over a slightly larger range than is likely to be experienced in the real feeder.

[0058] Each row of training data includes a fault location (i.e. distance from the or each probe) together with the waveform data (voltage and current data) from the probe(s). This can be used to train a neural network, the object being to produce a neural network which can be fed an input comprising a time series of waveform data, and provides an output corresponding to a fault location - or rather, a candidate fault location, since the waveform data fed to the neural network may or may not correspond to a fault, but corresponds to a locatable event. By locating the event and comparing the location with historical and collateral data, a decision can be made as to the probability that the event is indeed an early transient fault.

[0059] An ensemble of neural networks are then created. The neural networks may be for example convolutional neural networks. The hyperparameters of the neural networks (e.g. the number of layers) are randomised across the ensemble, so that each neural network in the ensemble is slightly different. Then, the training data generated by the simulation is applied to the neural networks to train the neural networks. In the training process, neural networks which diverge are dropped. Typically, around 300 neural networks may be created in the ensemble, and perhaps between 10 and 100 will remain after dropping nonconverging networks. Thus an ensemble of trained and converged neural networks is obtained.

[0060] Once this is done, real data from a real feeder can be applied to the ensemble of neural networks. The input to the neural networks is a time series of voltage I current data from the probe(s) on the network, for example a time series corresponding to about 100ms. However, it is not a case of simply applying every 100ms time window of recording from the real feeder to the neural networks, or even any arbitrary sample of a 100ms time window of recording from the real feeder to the neural networks. This is because the neural networks have been trained on the assumption that the input data contains a fault to locate. The vast majority of 100ms-long recordings from the probes on a real feeder will not contain a fault, due to the type of faults of interest being transient (short duration) faults. Applying every 100ms-long recording to the neural networks would firstly be impossible, and produce far more output data than could usefully be analysed, and secondly the output would in any case be meaningless - the output from the neural networks is supposed to be a fault location, but if there is no fault to locate then the output will not mean anything.

[0061] Therefore the method first identifies “events”, which can be considered “candidate faults”. An event is a time window during which a particular characteristic, or set of characteristics, of the voltage and current waveforms pass a threshold. For example, a current surge of more than 200A and of duration longer than 1 ms may be a suitable rule to identify an event. Note that the rules to identify an event should be selected so that real pecking faults are unlikely to be missed (i.e. a low false-negative rate is desirable). However, it is to be expected that some, perhaps as many as 50% or even more of the identified events, will not be real pecking faults (i.e. a relatively high falsepositive rate is acceptable). The point at this stage is not to identify pecking faults with certainty, but to identify locatable events which can be fed into the neural networks to produce useful output. The output from the neural networks may then be reviewed together with the original waveforms, historic outputs, and collateral data (e.g. maps of the area or data collected by physically visiting the area) to determine the likelihood that what is being detected is in fact an early transient fault. Preferably, the model and the trained ensemble of neural networks are validated by creating simulated faults in the real feeder (i.e. by applying equipment which causes short duration high current events at known locations on the real feeder). This may be done for example a few dozen times with the applied simulated fault at different known locations. The voltage and / or current waveform data is then collected and fed into the ensemble of neural networks, and the output compared to the ground truth known location of the simulated faults. This validation process may confirm that the simulation model is good enough and the neural networks have been correctly trained, but if not, it may usefully provide an indication of what is wrong. For example, commonly cable lengths in the simulation models are estimated based on maps which may be approximate or out of date. The parameters of the simulation model can be improved based on the real data obtained, then a new ensemble of neural networks can be trained on the improved model, which will hopefully succeed when validated.

[0062] Figure 3 shows an example of how output from the method of the invention can be usefully presented. The method is applied to a feeder to monitor the feeder and identify early transient faults. The probes are in place to constantly monitor the feeder and it is anticipated that the probes will be installed permanently since early transient faults could arise at any time. Although events may start to be detected within hours or days of installing the probes, it is after they have been installed for a longer period, for example a few months, that the data becomes particularly useful. This is because events of interest which have occurred at different times, but have been located to the same place on the feeder, can be viewed and analysed. The aggregate data may provide a very good indication of where real early transient faults have occurred, and therefore where problems with the feeder cable are likely to be developing.

[0063] Figure 3 shows a map, in this case of a fairly typical “high street” with mainly small shops I commercial buildings. Overlaid on the map is a series of lines showing the electricity feeder serving the buildings on the street. The feeder in this case is a radial feeder (i.e. there is a substation I power source only at one end) but two probes have been installed, marked on the map as 40a, 40b.

[0064] After the probes have been installed over a period of a few months, many events of interest have been identified and located as candidate faults. These have been plotted on the map where they appear as spots. Multiple events have been located around the area marked 42a and a smaller number, but still a few events occurring at different times have been located around the area marked 42b. Further analysis of the events located to these areas is required as they could have a number of causes. To facilitate that further analysis, the recorded waveform data from individual events can be viewed. This is shown in Figure 4 where, for each of two events, the voltage waveforms recorded by each of the two probes 40a, 40b are shown on the left hand side. On the right hand side, the corresponding output from the ensemble of neural networks is shown as a histogram on a diagram of the feeder.

[0065] The method of the invention locates candidate events, which include early transient faults. Although there may be false positives when it comes to the detection of early transient faults, by locating the event, whatever it is, accurately, a follow-on decision can be made as to how to act on the candidate fault. For example, if the event is located to a place where there is known to be a restaurant, and the event always occurs at about the same time of day which is just before the restaurant opens, it is more likely to be caused by a large load (e.g. a commercial oven) being turned on, than by a fault in the feeder. However, if the event occurs at different times of day, but more often on days when it is raining than dry days, that may be a very good indication that what is being seen is an early transient fault.

[0066] Because the fault is identified early, it can be repaired before it has caused any serious problem for customers. Because its location is accurately identified, it can be repaired quickly without large scale long duration road closures, and in many cases without causing even a short power outage during the repair.

[0067] The embodiments described above are provided by way of example only, and various changes and modifications will be apparent to persons skilled in the art without departing from the scope of the present invention as defined by the appended claims.

Claims

CLAIMS1 . A method of detecting and locating an early transient fault in a feeder from an electricity substation, the method comprising: monitoring a voltage and / or current waveform by at least one probe at at least one point on the feeder; detecting an event of interest based on one or more characteristic(s) of the voltage and / or current waveform exceeding a predetermined threshold; recording a time series of samples of the voltage and / or current waveform associated with the event; applying the time series of samples to the input of an artificial neural network, the output of the artificial neural network corresponding to estimated distance of a candidate fault from the probe; and locating the candidate fault according to the output of the artificial neural network.

2. A method as claimed in claim 1 , in which at least two probes are provided to monitor the voltage and / or current waveform at at least two different points on the feeder, and in which the output of the artificial neural network corresponds to estimated distance of the candidate fault from each probe.

3. A method as claimed in claim 1 or claim 2, in which an event of interest is detected based on a current exceeding a threshold magnitude for a predetermined duration.

4. A method as claimed in claim 3, in which the threshold magnitude is an offset from a measured moving average current over a time window which is greater than the predetermined duration.

5. A method as claimed in any of the preceding claims, in which an event of interest is detected based on a dip in voltage of more than a threshold magnitude and lasting for longer than a threshold duration.

6. A method as claimed in claim 5 when dependent on claim 2, in which an event of interest is detected based on a dip in voltage of more than a thresholdmagnitude and lasting for longer than a threshold duration, detected by more than one probe within a predetermined time period.

7. A method as claimed in any of the preceding claims, in which an ensemble of neural networks is provided, each neural network in the ensemble being provided with the time series of samples as the input, and each neural network in the ensemble having an output corresponding to estimated distance of a candidate fault from the or each probe.

8. A method as claimed in any of the preceding claims, in which the or each artificial neural network is trained using training data generated by a simulation model of the feeder.

9. A method as claimed in any of the preceding claims, in which the or each artificial neural network is trained using training data generated by multiple different simulation models of the feeder, each of the simulation models having different parameters.

10. A method as claimed in claim 7, in which each network in the ensemble is trained using the same training data, and in which each network in the ensemble has different hyperparameters.

11. A method as claimed in any of the preceding claims, in which the neural network is verified against verification data, the verification data comprising waveform measurements taken from the feeder while a simulated transient fault is generated on the feeder at a known location.

12. A method as claimed in claim 7, in which the outputs from the ensemble of neural networks are plotted on a histogram13. A method as claimed in claim 7, further comprising the step of determining a range of estimated locations of the candidate fault according to the outputs from the ensemble of neural networks, and identifying a probable fault in the case that the range of estimated locations is less than a predetermined threshold range.

14. A method as claimed in any of the preceding claims, in which the time series of samples is pre-processed before applying it as an input to the neural network(s), pre-processing the time series including subtracting a reference waveform.

15. A method as claimed in any of the preceding claims, in which the neural network(s) are convolutional neural network(s).

16. Apparatus adapted to carry out the method according to any of claims 1 to 15.

17. Non-transient computer-readable media having instructions which when executed on a processor coupled with probe(s) on a feeder, carries out the method according to any of claims 1 to 15.

Images