Service location anomaly

Abstract

The disclosed techniques include using machine learning to detect electrical anomalies in an electric power distribution system. As an example, the method includes accessing voltage measurements measured at an electric meter device. The method further includes calculating, from the voltage measurements, a corresponding average voltage and a corresponding minimum voltage for each time window of the set of time windows. The method further includes applying a machine learning model to the average voltages and the minimum voltages. The machine learning model is trained to identify one or more predefined electrical anomalies from the voltages. The method further includes receiving, from the machine learning model, a classification indicative of the identified anomalies. The method further includes sending an alert to a utility based on the classification.

Description

Technical Field

[0001] This application relates to machine learning for detecting anomalies in a power distribution system.

[0002] Cross-reference of related applications This application claims the benefit of U.S. Provisional Application No. 63 / _{216,375}, filed June 29, 2021, and U.S. Patent Application No. 17 / _{732,788}, filed April 29, 2022, both of which are incorporated herein by reference in their entirety.

Background Art

[0003] Electricity is supplied to consumers through a power distribution system. The power distribution system is complex, and the availability of electricity is important to customers. Therefore, anomalies in the power distribution system, if left unaddressed, can increase downtime, wear out components, and drive up service costs.

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, adding additional equipment to the power distribution system to detect anomalies can be costly. Therefore, it is desirable to detect anomalies in the power distribution system using existing components such as electricity meters.

Means for Solving the Problems

[0005] The disclosed technology includes using machine learning to detect electrical anomalies in a power distribution system. As an example, one method includes accessing voltage measurements taken by an electric metering device. The method further includes calculating the corresponding average voltage and corresponding minimum voltage for each time window in a set of time windows from the voltage measurements. The method further includes applying a machine learning model to the average voltage and minimum voltage. The machine learning model is trained to identify one or more predetermined electrical anomalies from the voltages. The method further includes receiving a classification from the machine learning model indicating the identified anomalies. The method further includes sending an alert to a utility operator based on the classification.

[0006] In another embodiment, the system for detecting anomalies in a resource distribution system includes a headend system and an electric metering device. The headend system includes a computing device and a machine learning model. Each electric metering device includes a sensor. Each electric metering device is electrically connected to a distribution transformer upstream of its respective electric metering device. Each electric metering device is configured to acquire its respective voltage measurement values ​​from its respective sensor. Each electric metering device is configured to provide its respective set of voltage measurement values ​​to the headend system. The headend system includes a machine learning model. The headend system is configured to acquire its respective set of voltage measurement values ​​from each of the electric metering devices. The headend system is further configured to access a first set of voltage measurement values ​​measured by the first electric metering device. The headend system is further configured to calculate a first corresponding average voltage and a first corresponding minimum voltage for each of the first time windows from the first set of voltage measurement values. The headend system is further configured to apply the machine learning model to the first average voltage and the first minimum voltage. The machine learning model is learned to identify a first voltage signature corresponding to an electrical anomaly from the voltage measurement values. The headend system is further configured to receive a first classification from a machine learning model indicating a first faulty connection. Based on the first classification, the headend system may be further configured to send a first alarm to a power company operator. The first alarm identifies a first electric meter device. The headend system is further configured to access a second set of voltage measurements taken by a second electric meter device. The headend system is further configured to calculate a second corresponding average voltage and a second corresponding minimum voltage for each of a second time window from the second set of voltage measurements. The second set of time windows may occur before a plurality of first time windows. The headend system is further configured to apply a machine learning model to the first average voltage, the first minimum voltage, the second average voltage, and the second minimum voltage.The headend system is further configured to receive a second classification from a machine learning model that identifies a second voltage signature indicating a second faulty connection. The headend system is further configured to send a second alarm to the utility operator based on the second classification. The second alarm identifies a second electric metering device.

[0007] In another embodiment, the method includes accessing a first set of voltage measurements taken by an electric metering device. The method further includes calculating a first corresponding average voltage and a first corresponding minimum voltage for each of a first time window from the first set of voltage measurements. The method further includes applying a machine learning model to the first average voltage and the first minimum voltage. The machine learning model is trained to identify a first voltage signature corresponding to an electrical anomaly from the voltage measurements. The method further includes receiving a first classification from the machine learning model indicating a first bad contact connection. The method further includes transmitting a first alarm to a utility operator based on the first classification. The method further includes calculating a second corresponding average voltage and a second corresponding minimum voltage for each of a second set of time windows from a second set of voltage measurements. The second set of time windows may occur before a plurality of first time windows. The method further includes applying a machine learning model to the first average voltage, the first minimum voltage, the second average voltage, and the second minimum voltage. The method further includes receiving a second classification from a machine learning model that identifies a second voltage signature indicating a second poor contact connection. The method further includes transmitting a second alarm to a utility operator based on the second classification.

[0008] These exemplary embodiments are not described to limit or define the Disclosure, but are provided to give examples that aid in understanding the Disclosure. Additional embodiments and further descriptions are described in the "Detailed Description".

[0009] These and other features, aspects, and advantages of this disclosure will be better understood by reading the following detailed description with reference to the accompanying drawings. [Brief explanation of the drawing]

[0010] [Figure 1] An exemplary communication network topology of a power distribution system according to one aspect of this disclosure is shown. [Figure 2] An exemplary power distribution network relating to one aspect of this disclosure is shown. [Figure 3] A flowchart illustrating an exemplary process for detecting anomalies using a machine learning model, according to one aspect of this disclosure, is shown. [Figure 4] This shows how statistical voltage data obtained from a metering device is calculated according to one aspect of this disclosure. [Figure 5] A flowchart illustrating an exemplary process for detecting anomalies using a machine learning model, according to one aspect of this disclosure, is shown. [Figure 6] A flowchart illustrating an exemplary process for training a machine learning model using supervised learning to detect anomalies, relating to aspects of this disclosure, is provided. [Figure 7] This is a graph showing voltage measurements related to poorly connected connections, according to one aspect of this disclosure. [Figure 8] This is a graph showing voltage measurements related to seasonal overload, according to one aspect of this disclosure. [Figure 9] This graph shows voltage measurements related to excessive voltage drop due to a long secondary transmission line, as disclosed herein. [Figure 10] An exemplary computing device relating to one aspect of this disclosure is shown. [Modes for carrying out the invention]

[0011] Detailed explanation Aspects of the present invention relate to the use of machine learning to detect electrical anomalies in an electrical system by learning and identifying voltage patterns (signatures) in voltage measurements obtained by a metering device (meter) installed in an end-user's home. Anomalies include, but are not limited to, poor contact in the connection between the meter and meter socket in the end-user's home, seasonal overloads (e.g., overloads that occur only seasonally), and long secondary lines (e.g., connections from distribution transformers to the end-user's home). Each of these anomalies may result in an identifiable voltage signature within the end-user's home.

[0012] Therefore, the advantages of certain embodiments include early identification of electrical anomalies, which helps in failure avoidance, improved system efficiency, and improved system reliability in the form of improved Mean Interruption Frequency Index (SAIFI) or Mean Interruption Time Index (SAIDI) scores. For example, once an anomaly is identified, proactive action can be taken to resolve the cause of the anomaly and avoid unplanned power outages, resulting in these advantages.

[0013] In addition, or alternatively, after identifying an electrical anomaly, the disclosed system can retrospectively analyze measurement data from one or more meters to determine additional patterns indicating the anomaly. An untrained machine learning model that is unaware of the voltage signatures that identify anomalies may not be able to identify such patterns, but a machine learning model that is aware of voltage signatures can be trained to identify such voltage signatures. In this respect, the disclosed solution can alert to the occurrence of electrical problems earlier. For example, a trained machine learning model with the ability to identify voltage signatures that match a poor contact can identify patterns from the corresponding electric meter data months or even years before the poor contact becomes a serious problem.

[0014] The following non-limiting example is provided for illustrative purposes. Voltage measurements are collected by a metering device at a specific frequency (e.g., every 15 minutes). Suitable examples of metering devices include smart meters or advanced measurement infrastructure (AMI) meters. The metering device transmits voltage data to a headend system via a communication network, either together with or separately from metering data such as power consumption.

[0015] Continuing this example, the headend system receives voltage measurements from a metering device and derives statistical data from these measurements. For example, statistical data such as the average daily voltage or the lowest daily voltage is calculated over a certain period (e.g., one month). This statistical data is provided to a machine learning model. From the voltage data and / or derived statistics, the machine learning model, which has been pre-trained to detect one or more anomalies such as poor contact or seasonal overload, determines the presence of an anomaly.

[0016] Next, looking at the drawings, Figure 1 shows an exemplary communication network topology of a power distribution system according to an embodiment. Figure 1 includes a headend system 100, network connection 108, collection device 110, network connections 112, 114, 116, 118, metering devices 122, 124, 126, 128, and end-user premises 132, 134, 136, 138. The communication network topology shown in Figure 1 illustrates an example of how various devices are interconnected in a communication network. The communication network topology differs from the power distribution method illustrated in Figure 2.

[0017] In the example shown in Figure 1, one or more of the metering devices 122, 124, 126, and 128 measure one or more parameters such as voltage, current, phase, actual power consumption (watt-hours), and actual power (watts). These parameters are then provided to the data acquisition device 110 via network connections 112, 114, 116, and 118. The data acquisition device 110 then provides the measurement data to the headend system 100. The headend system 100 then processes the data to detect anomalies using machine learning.

[0018] The head - end system 100 includes a computing device 102 and a machine - learning model 104. An example of the computing device is illustrated with respect to FIG. 9. The computing device 102 can derive statistical values from voltage data collected by the meter devices 122, 124, 126, and 128, and / or can provide the statistical values to the machine - learning model 104. The machine - learning model 104 can identify one or more electrical anomalies in the meter devices 122, 124, 126, and 128. Examples of anomalies that can be detected by the machine - learning model 104 include poor contact connections (e.g., of the meter socket), seasonal overloads, and excessive voltage drops due to long secondary distribution lines. Exemplary voltage signatures indicating these states are illustrated in FIGS. 6, 7, and 8, respectively. <​​​​

[0020] The functions described with respect to the head - end system 100 can be implemented in any combination of the head - end system 100 (e.g., the computing device 102), one or more of the meter devices 122, 124, 126, and 128, and a cloud - based system (i.e., servers connected via a persistent network connection). Examples of systems include an AMI system and a meter data management (MDM) system.

[0021] For example, while the network connection 108 is illustrated as a wired network connection, the network connections 112, 114, 116, 118 are illustrated as wireless network connections. Examples of typical networks include wireless (Wi - Fi, Bluetooth®, mesh, cellular, etc.) and wired (Ethernet, power - line communication, etc.) networks. However, various configurations are possible. For example, the meter devices 122, 124, 126, and 128 can communicate with each other and / or with the head - end system 100 using any communication network. Communication can also be sent from the head - end system 100 to the meter devices 122, 124, 126, and 128. In some configurations, the collection device 110 does not exist, and the meter devices 122, 124, 126, and 128 can communicate directly with the head - end system 100.

[0022] FIG. 2 shows an exemplary distribution network according to an aspect of the present disclosure. In the example illustrated in FIG. 2, the distribution system 200 includes a power source 202, a transformer 204 of a distribution substation, a feeder line 208, distribution transformers 210 and 212, primary distribution lines 240 and 241, secondary lines 242, 244, 246, and 248, meter devices 222, 224, 226, and 228, and end - user premises 232, 234, 236, and 238.

[0023] An example of a power source 202 is a reduced representation of a bulk power system to which a transformer 204 of a distribution substation is connected, which includes a sub-power transmission network, a transmission network, and a power source (e.g., a power plant, solar panels, or wind turbine generator). The distribution substation 204 transforms the voltage output from power source 202 to a level suitable for feed line 208. Feed line 208 supplies power to distribution lines 240 and 241. The distribution substation 210 transforms the voltage of distribution line 240 to different voltages for secondary lines 242, 244, and 246. Similarly, the distribution transformer 212 transforms the voltage of distribution line 241 to different voltages for secondary line 248.

[0024] As shown in the diagram, meter devices 222, 224, 226, and 228 correspond to end-user premises 232, 234, 236, and 238, respectively. Meter devices 222, 224, 226, and 228 may correspond to meter devices 122, 124, 126, and 128, respectively. End-user premises 232, 234, 236, and 238 may correspond to end-user premises 132, 134, 136, and 138, respectively.

[0025] Metering device 222 measures the parameters of secondary line 242, metering device 224 measures the parameters of secondary line 244, metering device 226 measures the parameters of secondary line 246, and metering device 228 measures the parameters of secondary line 248. Since the length of each distribution line may vary, in some cases, an anomaly may occur due to the secondary line being too long. For example, if secondary line 248 is longer than typical, the voltage drop between the distribution transformer 212 and the end-user premises 238 may become abnormally high. This anomaly can be identified by machine learning model 104.

[0026] In some examples, the metering device may be associated with a secondary distribution line distributing one or more phases of a multiphase distribution system. For example, the secondary distribution line 242 may include two phases of a three-phase power generation and distribution system. In this configuration, using the voltage measurements obtained by the metering device 222, the headend system 100 can determine that different phases have different average voltages and / or minimum voltages by providing the measured voltages and / or derived statistics therefrom to a machine learning model 104.

[0027] Figure 3 shows a flowchart of an exemplary process 300 for detecting anomalies using a machine learning model, according to an aspect of this disclosure. For exemplary purposes, the process 300 is discussed as being performed by a headend system 100. However, the process 300 may be performed in any computing device, such as a metering device.

[0028] In block 302, process 300 includes accessing a set of voltage measurements taken over a period of time by a metering device. For example, the headend system 100 accesses voltage measurements obtained from the metering device 122. Accessing voltage measurements may include sending requests to the electric metering device and receiving voltage measurements returned from the electric metering device. In another example, the electric metering device may periodically transmit voltage measurements to the headend system. Process 300 is illustrated with reference to Figure 4 as an example.

[0029] Figure 4 illustrates the calculation of statistical voltage data obtained from a metering device according to one aspect of the present disclosure. Figure 4 illustrates the data flow 400, which shows how average voltage data and minimum voltage data are derived for various periods from data obtained from a single metering device. The data flow 400 includes voltage data 410 acquired over a period 420, a time window 430a-n, statistical data 440a-n, and a machine learning model 104. Periods and time windows of any length can be used.

[0030] Continuing this example, the computing device 102 accesses the voltage data 410 corresponding to the time period 420 and separates the voltage data 410 into multiple time windows 430a-n. Each of the time windows 430a-n is smaller than the time period 420. Each of the time windows 430a-n contains multiple voltages. For example, if a particular time window is one day and the voltage is acquired every 15 minutes, one time window will contain 96 voltage measurements. In one example, the period 420 is one month, and each of the time windows 430a-n is one day. For example, time window 430a is day 1, time window 430b is day 2, and so on.

[0031] Returning to Figure 3, in block 304, process 300 includes calculating the corresponding average voltage and corresponding minimum voltage for each time window in a set of time windows from a set of voltage measurements. Continuing the example, the calculator 102 calculates statistical data 440a-n from time windows 430a-n. More specifically, the calculator 102 calculates statistical data 440a, which includes the average and minimum values ​​from time window 430a, and statistical data 440b, which includes the average and minimum values ​​from time window 430b, and so on.

[0032] While the mean and minimum values ​​are discussed in relation to Figure 4, other statistical metrics such as the median, mode, and maximum values ​​derived from voltage or other parameters can also be used by the machine learning model 104. In this case, the machine learning model 104 is trained using these statistical data.

[0033] In block 306, process 300 includes applying a machine learning model to the average and minimum voltages over a time period. Continuing this example, the computer 102 provides statistical data 440a-n to the machine learning model 104.

[0034] The machine learning model 104 is trained to identify one or more electrical anomalies from the voltage, as will be further explained with reference to Figure 6. In some cases, one or more of the average voltage and minimum voltage are converted into one or more feature vectors. A feature vector is a vector containing multiple elements about the object in question (e.g., an electric meter). Thus, a feature vector may include the average voltage, minimum voltage, or other statistics. One or more feature vectors are provided to the machine learning model 104 at a time.

[0035] In block 308, process 300 includes receiving a classification from the machine learning model indicating the identified electrical anomaly. Continuing the example, machine learning model 104 outputs a classification that identifies an electrical anomaly. Examples of anomalies include poor contact to power lines within an end-user premises, seasonal consumption overloads, and voltage drops caused by long secondary lines (connections from distribution transformers to end-user premises).

[0036] One example of an abnormality is a poor contact related to an electric meter device, as will be further explained with reference to Figure 6. A poor contact can be represented by a first decrease in the lowest voltage over a period of time and a second decrease in the average voltage over the same period, in which case the second decrease is smaller than the first decrease.

[0037] Another example of an anomaly, as further illustrated with reference to Figure 7, is a seasonal overload caused by power consumption measured by an electric metering device. Such anomalies can be represented by one or more correlations between one or more peaks or valleys of the minimum voltage and one or more peaks or valleys of the average voltage.

[0038] The machine learning model 104 can output one or more specific anomaly classifications. In other cases, the machine learning model 104 can output determined probabilities for one or more classifications. For example, the machine learning model 104 could output an 80% probability that the voltage indicates a poor connection and a 20% probability that there is no poor connection. In other cases, a set of probabilities may be generated such that there is a 50% chance of a poor connection, a 40% chance of a long secondary line, and a 10% chance that neither anomaly exists. A machine learning model that generates specific classifications (e.g., positive or negative) may be a different type of model than a machine learning model that generates a range of probabilities.

[0039] The output of the machine learning model 104 is provided to the computing device 102. In some cases, the computing device 102 can decide that a corresponding classification is used when the probability exceeds a certain threshold. In other cases, the computing device 102 can decide that no classification is performed when the probability falls below the threshold.

[0040] In block 310, process 300 includes sending an alarm to a utility operator or adjusting one or more parameters of the distribution system based on classification. For example, the computing device 102 may send an alarm to a utility operator or to an engineer visiting a customer's home to perform repairs such as replacing a faulty meter socket or meter. As a result, the balance of the electrical load on the line may be changed, equipment may be added, or equipment (e.g., transformers) may be replaced.

[0041] In another example, in block 310, the computing unit 102 can cause the headend system 100 to adjust one or more parameters of the power distribution system. Examples of parameters that can be adjusted include line voltage, phase, load, reactance, capacitance, etc. In some cases, such adjustments may be performed remotely via a communication network, for example, by communicating with a resource adjustment device, which then performs the adjustments.

[0042] In one embodiment, data from multiple metering devices can be analyzed by the machine learning model 104. For example, process 300 can be executed multiple times, once for each metering device. Alternatively, blocks 302-304 can be executed multiple times, once for each metering device, and then the data from multiple metering devices can be analyzed in an aggregated format by the machine learning model 104, for example, in block 308. This analysis may be performed in real time or after a threshold amount of data has been buffered.

[0043] As an example, the headend system 100 accesses voltage measurements from two or more metering devices. The headend system 100 calculates statistical metrics for each metering device, for each voltage measurement, and for each time window. The headend system 100 then applies a machine learning model 104 to the statistical metrics. In some cases, the application adjusts the learning of the machine learning model. Continuing the example, the headend system 100 identifies one or more anomalies in one or more electric metering devices.

[0044] In some cases, machine learning models can use topology information of the electric metering system. For example, given topology information (such as that illustrated in Figure 2), machine learning model 104 can determine whether an anomaly exists in one metering system but not in another. From this determination, machine learning model 104 can identify problems in other components, such as distribution transformers. For example, if a particular anomaly is displayed in metering systems 222, 224, and 226 but not in metering system 228, there may be an anomaly in the distribution transformer 210 or the distribution line 240.

[0045] As further explained with respect to Figure 9, another example of an anomaly is an excessive voltage drop due to a long distribution line associated with an electric metering device. In some cases, determining this excessive voltage drop requires collecting measurements from additional electric meters. For example, such an electrical network topology may involve one or more distribution transformers located upstream of a subset of electric metering devices, electrically connected via the distribution line. One identifiable anomaly can be identified by the difference between a set of sub-average voltages of electric meters and the daily average voltage of one electric meter exceeding a threshold.

[0046] Figure 5 shows a flowchart of an exemplary process 500 for detecting anomalies using a machine learning model according to an aspect of the present disclosure. In relation to the process 300 discussed with respect to Figure 3, process 500 includes using a first identification of an anomaly (optionally including relevant voltage data) in combination with machine learning to identify a second identification of an anomaly. The second identification of an anomaly can correspond to anomalies that occurred earlier in time in the data than the first identification.

[0047] For example, process 500 can detect a bad connection in a first set of voltage data, and use the detected voltage signature in the same or a different machine learning model to identify a second voltage signature in statistical data (e.g., voltage data) originating from the same or a different electric meter. For example, a bad connection may have been present for some time but not detected until the initial classification. With the initial classification, earlier identification may be possible retrospectively and / or in a different dataset.

[0048] In block 502, process 500 includes receiving a first classification from a machine learning model that indicates a first bad contact connection based on a first set of voltage measurements. In block 502, process 500 includes operation similar to that described with respect to blocks 302-308 of process 300. The classification may include a voltage signature. Process 500 may optionally include sending an alarm to a utility operator based on the first classification, as discussed with respect to block 310 of process 300.

[0049] In block 504, process 500 includes calculating a second corresponding average voltage and a second corresponding minimum voltage for each of the second set of time windows from a second set of voltage measurements. In block 504, process 500 includes the same operation as discussed for block 304 of process 300. The second set of voltages may come from the same electric meter from which the voltage data used in block 502 originates, and / or from a different electric meter(s). The second set of voltages may occur before the first set of time windows (for example, as discussed for block 304 of process 300).

[0050] In block 506, process 500 includes applying a machine learning model to a first voltage signature, a second average voltage, and a second minimum voltage. Continuing the example, the computer 102 provides the machine learning model 104 with one or more of the first voltage signature (e.g., identified in block 502), the first average and minimum voltages (e.g., identified in block 502), the second average voltage, and the second minimum voltage. In some cases, a different machine learning model than the one used in block 502 may be used. The machine learning model can be trained to identify one or more electrical anomalies from the voltage, as will be further illustrated with reference to Figure 6.

[0051] In block 508, process 500 includes receiving a second classification from a machine learning model that identifies a second voltage signature indicating a second poor contact connection. Continuing the example, computing device 102 receives a second classification that identifies a second voltage signature. The second voltage signature may be identical, similar, or different from the first voltage signature.

[0052] In block 510, process 500 includes sending an alarm to a utility operator based on a second classification. In block 510, process 500 includes the same operation as discussed with respect to block 310 of process 300.

[0053] The machine learning model 104 is trained to detect one or more predetermined voltage signatures, each corresponding to one or more anomalies. Various learning techniques can be used, such as supervised learning (e.g., labeled training data), unsupervised learning (e.g., unlabeled data), and reinforcement learning. Learning can be performed by the headend system 100, for example, the computing device 102, or another computing system. If learning is performed by another computing system, the machine learning model 104 can be provided to the headend system 100 (e.g., downloaded) and / or updated as needed.

[0054] In some cases, learning can be performed at runtime. For example, an operator can provide feedback to the computing device 102 indicating whether the classification or prediction of anomalies is correct or incorrect, and the computing device 102 updates the machine learning model 104 accordingly.

[0055] During training, the machine learning model learns an algorithm to identify electrical anomalies. Figure 6 illustrates supervised learning, but other learning techniques can also be used. Furthermore, the learned machine learning algorithm can be improved over time, for example, through additional training at runtime.

[0056] Figure 6 shows a flowchart of an exemplary process 600 for training a machine learning model using supervised learning to detect anomalies, according to an aspect of this disclosure. In the supervised learning approach, a determined probability or classification is computed and compared to an expected or known probability or classification. A loss function is computed from the comparison. Based on the computed loss function, the machine learning model 104 adjusts its internal parameters to minimize the loss function. Examples of suitable machine learning models include neural networks, classifiers, and decision trees.

[0057] In block 602, process 600 includes accessing a set of training data pairs. Each training data pair includes statistics (e.g., a set of average voltages and a set of minimum voltages over a certain period) and a predicted classification indicating one or more electrical anomalies. Each training data pair includes previously identified data as part of a positive training set (i.e., voltage data corresponding to previously identified voltage signatures) and / or a negative training set (i.e., voltage data not corresponding to previously identified voltage signatures). Process 600 is described as being performed by computing device 102. However, training may be performed by any computing system.

[0058] In block 604, process 600 includes providing one of the sets of training data pairs to the machine learning model. Continuing the example, computing device 102 provides the machine learning model 104 with a set of training data pairs.

[0059] In block 606, process 600 includes receiving the determined classification from the machine learning model. Subsequently, the exemplary computing device 102 receives the determined classification from the machine learning model 104.

[0060] In block 608, process 600 includes calculating a loss function by comparing the determined classification with the expected classification. Continuing the example, computing device 102 calculates a loss function by comparing the determined classification (i.e., received in block 606) with the expected classification (i.e., contained in the training data set accessed in block 604).

[0061] In block 610, process 600 includes tuning the intrinsic parameters of the machine learning model to minimize the loss function. Continuing the example, the computer 102 appropriately tunes the machine learning model 104.

[0062] In block 612, process 600 checks whether training is complete. If training is complete, process 600 proceeds to block 614, where training ends. If training is not complete, process 600 returns to 604 and continues training the machine learning model using a different set of training data. Completion of training can be indicated by the end of the training data, the loss function being minimized below a threshold level, or other conditions.

[0063] In one embodiment, the machine learning model 104 is trained to detect multiple voltage signatures. In this case, the machine learning model 104 can identify electrical signatures indicating multiple electrical anomalies. For example, the machine learning model 104 can identify socket contact failure and seasonal overload. In this case, process 600 can be completed one or more times for each electrical anomaly.

[0064] In a further embodiment, the machine learning model 104 is tested to ensure sufficient accuracy in classifying electrical anomalies. Typically, the data used to test the machine learning model 104 is not included in the dataset (i.e., training data pair) used to train the machine learning model.

[0065] In another embodiment, the machine learning model 104 can be trained in an unsupervised manner using historical data from metering equipment. The historical data may include data from a period prior to the initial identification of the defect. For example, after a particular metering device is identified as malfunctioning (e.g., by analyzing the voltage from the metering device as described in process 300), the computing device 102 and / or the machine learning model 104 can analyze additional data from the metering device to determine whether additional voltage signatures exist. For example, a signature may be identified in the historical data, and then the time prior to the occurrence of the defect may be identified. Using this approach, the machine learning model 104 can identify one or more precursor patterns in the collected voltage data potentially earlier than would have been previously detectable using supervised techniques (e.g., process 600).

[0066] In a more detailed example, process 300 is used to identify five meter devices with loose connections. For example, one month's worth of historical voltage data is processed by process 300. The five meter devices are secured by a service technician tightening the screws on the base of the meter devices. This process of identifying meter devices with loose or poorly connected connections can be continued.

[0067] Continuing this example, after several months, 20 meters are identified and improved. However, for each meter, there is one year of historical data. By training the machine learning model 104 with this historical data, the machine learning model 104 identifies one or more additional precursor signatures common to all 20 meters. The signatures thus identified can complement and / or replace the signatures identified in process 500. In this way, learning can be continued over time, either unsupervised or supervised. The advantage of this continuous learning approach is that it includes the identification of voltage signatures that may occur at unpredictable or unexpected times, thereby improving the machine learning model 104.

[0068] Figure 7 shows a graph illustrating voltage measurements related to poor contact connections according to an aspect of this disclosure. Graph 700 shows the average daily voltage 710 and the minimum daily voltage 720 measured by a metering device within the end user's premises. The data corresponding to Graph 700 is shown in Table 1 below.

[0069] As seen in Graph 700, there can be a significant difference at some point between the average daily voltage of 710 and the lowest daily voltage of 720. These differences indicate arcing of the connection or frequent disconnections of the connection. Combining the measurements in Graph 700, a voltage signature indicating poor socket contact can be formed. For example, the magnitude of the dip in the lowest voltage of 720 relative to the average voltage of 710 indicates poor contact. The machine learning model 104 is trained to identify these signatures as anomalies.

[0070] [Table 1]

[0071] Figure 8 shows a graph illustrating voltage measurements related to seasonal overload according to an aspect of this disclosure. Graph 800 illustrates the average daily voltage 810 and the minimum daily voltage 820 measured by a metering device within the end-user's premises. The data for Graph 800 is shown in Table 2 below, where the period used is in years and the time period is in months.

[0072] The data shows voltage measurements from January to December. As can be seen, the load is higher from April to September, as indicated by the decrease in voltage. Such voltage drops can form a sign that the transformer is overloaded. For example, the relative position of the peak and valley of the lowest voltage of 820 to the peak and valley of the average voltage of 810 could indicate seasonal overload. The machine learning model 104 is trained to identify these signatures as anomalies.

[0073] [Table 2]

[0074] Figure 9 shows a graph illustrating voltage measurements related to excessive voltage drop due to a long secondary line, according to an aspect of the present disclosure. Graph 900 illustrates the daily average voltage 910 of all meters behind the service transformer (e.g., metering devices 222, 224, and 226) compared to the daily average voltage 920 measured by a metering device (metering device 226) within the end-user premises. Graph 900 is shown for measurements over a one-month period, as shown in Table 3 below.

[0075] As seen in Graph 900, there is a large gap between the daily average voltages of all meters behind the service transformer, compared to the daily average voltage of meters with long secondary lines. Some topological information is needed to recognize this signature. This topological information associates the metering equipment with the transformer for the purpose of using voltage to recognize this particular pattern. The machine learning model 104 can recognize this signature by comparing the average voltage of all metering equipment behind the transformer with the daily average of each metering equipment.

[0076] [Table 3]

[0077] Figure 10 shows an exemplary computing device relating to this disclosure. Any suitable computing system may be used to perform the operations described herein. An example of the illustrated computing device 1000 includes a processor 1002 communicatively coupled to one or more memory devices 1004. The processor 1002 executes computer executable program code 1030 stored in the memory devices 1004, accesses data 1020 stored in the memory devices 1004, or both. Examples of the processor 1002 include a microprocessor, an application-specific integrated circuit ("ASIC"), a field-programmable gate array ("FPGA"), or any other suitable processing device. The processor 1002 may include a single processing device or any number of processing devices or cores. The functions of the computing device may be implemented in hardware, software, firmware, or a combination thereof.

[0078] In some embodiments, the computing device 1000 may include at least one sensor configured to measure parameters related to resources in a resource distribution network. For example, in a power distribution system, the sensor may measure power consumption, voltage, current, etc. In some embodiments, the computing device 1000 may include multiple sensors. For example, the computing device 1000 may include both a power sensor and a temperature sensor.

[0079] The memory device 1004 includes any suitable non-temporary computer-readable medium for storing data, program code, or both. The computer-readable medium may include any electronic, optical, magnetic, or other storage device that can provide computer-readable instructions or other program code to the processor. Non-limiting examples of computer-readable medium include flash memory, ROM, RAM, ASIC, or other medium from which the processing unit can read instructions. The instructions may include processor-specific instructions generated by a compiler or interpreter from code written in any suitable computer programming language, such as C, C++, C#, Visual Basic, Java, or a scripting language.

[0080] The computing device 1000 may also include numerous external and internal devices, such as input and output devices. For example, the computing device 1000 is shown to have one or more input / output ("I / O") interfaces 1008. The I / O interface 1008 can receive input from input devices and provide output to output devices. One or more buses 1006 are also included in the computing device 1000. The buses 1006 communicately connect one or more components of each of the computing devices 1000.

[0081] The computing device 1000 executes program code 1030 that configures the processor 1002 to perform one or more operations described herein.

[0082] The computing device 1000 also includes a network interface device 1010. The network interface device 1010 includes any device or group of devices suitable for establishing wired or wireless data connections to one or more data networks. The network interface device 1010 may be a wireless device and have an antenna 1014. The computing device 1000 can use the network interface device 1010 to communicate with one or more other computing devices implementing computing or other functions via the data network.

[0083] The computing device 1000 may also include a display device 1012. The display device 1012 may be an LCD, LED, touchscreen, or other device capable of displaying information about the computing device 1000. For example, the information may include the operating status of the computing device, the network status, and so on.

[0084] While the subject matter has been described in detail with respect to certain aspects, those skilled in the art will understand that, with the foregoing understanding, modifications, variations, and equivalents to such aspects can be readily created. Therefore, it should be understood that this disclosure is presented for illustrative purposes only, not limitation, and does not preclude the inclusion of modifications, variations, and / or additions to the subject matter that would be readily apparent to those skilled in the art.

Claims

1. A method of using machine learning to detect electrical anomalies in a power distribution system, The aforementioned method, Accessing a first set of voltage measurements taken by an electric metering device, From the first plurality of voltage measurements, calculate the first corresponding average voltage and the first corresponding minimum voltage for each of the first plurality of time windows, The method involves applying a machine learning model to a first average voltage and a first minimum voltage, wherein the machine learning model is trained to identify a first voltage signature corresponding to an electrical anomaly from the voltage measurement values. The machine learning model receives a first classification indicating a first poorly connected connection, Based on the first classification described above, a first alarm is transmitted to the utility operator, Calculating a second corresponding average voltage and a second corresponding minimum voltage for each of a second set of time windows from a second set of voltage measurements, wherein the second set of time windows occurs before the first set of time windows, and the second set of time windows occurs before the first set of time windows. The machine learning model is applied to the first average voltage, the first minimum voltage, the first voltage signature, the second average voltage, and the second minimum voltage. The machine learning model receives a second classification that identifies a second voltage signature indicating a second poor contact connection. A method comprising transmitting a second alarm to a utility operator based on the second classification described above.

2. The aforementioned second set of voltage measurements are measured by an additional electric metering device. The method according to claim 1.

3. The first voltage signature includes a first decrease in minimum voltage over a certain period and a second decrease in average voltage over the same period. The second decrease is smaller than the first decrease. The method according to claim 1.

4. The aforementioned second set of voltage measurements are measured by an electric metering device. The method according to claim 1.

5. The machine learning model is further applied to topology information relating the electric meter device to one or more distribution transformers electrically connected to the electric meter device via a distribution line. The method according to claim 1.

6. Accessing the first plurality of voltage measurements includes sending a request to the electric meter device and receiving the first plurality of voltage measurements from the electric meter device. The method according to claim 1.

7. The aforementioned method, Training a machine learning model by accessing a set of training data pairs, wherein each training data pair is: (i) A set of average voltages to learn and a set of minimum voltages, (ii) A learning set of the average voltage of all electric metering devices connected to the distribution transformer, or (iii) A learning set of average voltages from one electric meter unit behind a distribution transformer, and an expected classification indicating one or more electrical anomalies. Including one or more of the following, The machine learning model is provided with one of the training data pairs from the set of training data pairs. The machine learning model receives the determined classification, This involves comparing the determined classification with the expected classification to calculate the loss function, and The intrinsic parameters of the aforementioned machine learning model are adjusted to minimize the loss function, Further including, The method according to claim 1.

8. The set of training data pairs further includes topology information relating one or more distribution transformers to one or more metering devices. The method according to claim 7.

9. A non-temporary computer-readable storage medium storing computer-executable program instructions, wherein when executed by a processing device, the computer-executable program instructions cause the processing device to perform a predetermined operation. The aforementioned operation is, Accessing a first set of voltage measurements taken by an electric metering device, From the first plurality of voltage measurements, calculate the first corresponding average voltage and the first corresponding minimum voltage for each of the first plurality of time windows, The method involves applying a machine learning model to a first average voltage and a first minimum voltage, wherein the machine learning model is trained to identify a first voltage signature corresponding to an electrical anomaly from the voltage measurement values. From the aforementioned machine learning model, a first classification indicating a first anomaly is received, Based on the first classification described above, a first alarm is transmitted to the utility operator, The method involves calculating a second corresponding average voltage and a second corresponding minimum voltage for each of a second set of time windows from a second set of voltage measurements, wherein the second set of time windows occurs before the first set of time windows. The machine learning model is applied to the first average voltage, the first minimum voltage, the second average voltage, and the second minimum voltage. The machine learning model receives a second classification that identifies a second voltage signature indicating a second anomaly, Based on the second classification described above, a second alarm is transmitted to the utility operator, Non-temporary computer-readable storage media, including [specific data / information].

10. One or more of the first and second abnormalities are related to a contact failure in the electric meter device. A non-temporary computer-readable storage medium according to claim 9.

11. One or more of the first and second abnormalities are represented by a first decrease in the minimum voltage over a certain period and a second decrease in the average voltage over that period. The second decrease is smaller than the first decrease. A non-temporary computer-readable storage medium according to claim 9.

12. One or more of the first and second abnormalities are represented by one or more correlations between one or more peaks or valleys of the first minimum voltage and one or more peaks or valleys of the first average voltage. A non-temporary computer-readable storage medium according to claim 9.

13. When executed by the aforementioned processing unit, a computer-executable program instruction causes the processing unit to apply a machine learning model to topology information relating one or more distribution transformers electrically connected to the electric meter device via a distribution line to the electric meter device. A non-temporary computer-readable storage medium according to claim 9.

14. Accessing the first plurality of voltage measurement values ​​includes sending a request to the electric meter device and receiving the plurality of voltage measurement values ​​from the electric meter device. A non-temporary computer-readable storage medium according to claim 9.

15. When executed by the aforementioned processing unit, the computer executable program instruction causes the processing unit to train a machine learning model in the following manner, and the manner is: Accessing a set of training data pairs, where each training data pair is: (i) A set of average voltages and a set of minimum voltages, (ii) A set of average voltages of multiple electric meter devices connected to a distribution transformer, (iii) A set of average voltages from one electric meter device behind a distribution transformer, and an expected classification indicating one or more electrical anomalies, Including one or more of the following, The machine learning model is provided with a set of training data pairs, The machine learning model receives the determined classification, This involves comparing the determined classification with the expected classification to calculate the loss function, and The intrinsic parameters of the aforementioned machine learning model are adjusted to minimize the loss function, including, A non-temporary computer-readable storage medium according to claim 9.

16. A system for detecting electrical anomalies in a resource allocation system, The aforementioned system, A headend system including computing devices and machine learning models, Equipped with multiple electric meter devices, Each of the aforementioned electric meter devices is equipped with a sensor, Each of the aforementioned electric meter devices is electrically connected to a distribution transformer located upstream of each of the aforementioned electric meter devices. Each of the aforementioned electric meter devices is, Multiple voltage measurement values ​​are obtained from each sensor of the aforementioned electric meter device. Each of the multiple voltage measurements is configured to provide the headend system. The headend system includes a machine learning model, Access the first set of voltage measurement values ​​measured by the first of the set of electricity metering devices, From the first plurality of voltage measurements, a first corresponding average voltage and a first corresponding minimum voltage are calculated for each of the first plurality of time windows. The machine learning model is configured to be applied to a first average voltage and a first minimum voltage, and the machine learning model is trained to identify a first voltage signature corresponding to an electrical anomaly from the voltage measurement. The headend system is From the machine learning model, a first classification indicating a first poorly connected connection is received. Based on the first classification, it is configured to send a first alarm to the utility operator, the first alarm identifies a first electric meter device, The headend system is Access the second set of voltage measurements taken by the second electric meter device among the set of electric meter devices, The system is configured to calculate a second corresponding average voltage and a second corresponding minimum voltage for each of the second multiple time windows from the second multiple voltage measurements, wherein the second multiple time windows occur before the first multiple time windows. The headend system is The machine learning model is applied to the first average voltage, the first minimum voltage, the second average voltage, and the second minimum voltage. From the machine learning model, a second classification is received that identifies a second voltage signature indicating a second poor contact connection. Based on the second classification, it is configured to send a second alarm to the utility operator, the alarm identifying a second electric meter device. system.

17. The headend system is further configured to train a machine learning model with multiple voltage measurements from at least one electric metering device. The system according to claim 16.

18. The headend system is further configured to apply topology information to a machine learning model that associates multiple electric metering devices with one or more distribution transformers electrically connected upstream of the multiple electric metering devices via distribution lines. The first abnormality is that the difference between the average voltage of the multiple electric meters and the daily average voltage of one of the multiple electric meters is greater than or equal to a threshold. The system according to claim 16.

19. The first voltage signature includes a first decrease in the lowest voltage over a certain period and a second decrease in the average voltage over a certain period. The second decrease is smaller than the first decrease. The system according to claim 16.

20. Receiving the aforementioned multiple voltage measurement values ​​includes transmitting a request to each electric meter device and receiving the multiple voltage measurement values ​​from each electric meter device. The system according to claim 16.

Images