CN117916973A

CN117916973A - Service location anomaly

Info

Publication number: CN117916973A
Application number: CN202280045718.5A
Authority: CN
Inventors: C·I·加尔扎
Original assignee: Landis+gail Technology Co ltd
Current assignee: Landis+gail Technology Co ltd
Priority date: 2021-06-29
Filing date: 2022-06-06
Publication date: 2024-04-19

Abstract

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

Description

Service location anomaly

Cross Reference to Related Applications

The present application claims the benefit of U.S. provisional application No. 63/216,375 filed on 29 at 2021 and U.S. patent application No. 17/732,788 filed on 29 at 2022, which are incorporated herein by reference in their entirety.

Technical Field

The present application relates to using machine learning to detect anomalies in a power distribution system.

Background

Power is distributed to users through a power distribution system. Power distribution systems are complex and the availability of power is important to customers. Thus, anomalies in the power distribution system may increase downtime, component wear, and increase service costs if not repaired. But adding additional equipment to the power distribution system to detect anomalies can be expensive. Accordingly, it is desirable to detect anomalies in a power distribution system by using existing components such as electricity meters.

Disclosure of Invention

In another example, a system for detecting anomalies in a resource allocation system includes a headend system and an electric metering device. The headend system includes a computing device and a machine learning model. Each electrical metering device includes a sensor. Each electrical metering device is electrically connected to a distribution transformer upstream of the respective electrical metering device. Each electrical metering device is configured to obtain a respective set of voltage measurements from a respective sensor of the electrical metering device. Each electrical metering device is configured to provide a respective set of voltage measurements to the headend system. The headend system includes a machine learning model. The headend system is configured to obtain a respective set of voltage measurements from each of the electrical metering devices. The headend system is also configured to access a first set of voltage measurements measured at the first electrical metering device. The headend system is also configured to calculate a first corresponding average voltage and a first corresponding minimum voltage from the first set of voltage measurements and for each time window in the first set of time windows. The headend system is also configured to apply 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 characteristic (signature) corresponding to the electrical anomaly from the voltage measurements. The headend system is also configured to receive a first classification from the machine learning model indicating a first loose connection. The headend system may also be configured to send a first alert to the utility operator based on the first classification. The first alarm identifies the first electrical metering device. The headend system is also configured to access a second set of voltage measurements measured at a second electrical metering device. The headend system is also configured to calculate a second corresponding average voltage and a second corresponding minimum voltage from the second set of voltage measurements and for each time window in the second set of time windows. The second set of time windows may occur before the first plurality of time windows. The headend system is also 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 also configured to receive a second classification from the machine learning model that identifies a second voltage characteristic indicative of a second loose connection. The headend system is also configured to send a second alert to the utility operator based on the second classification. The second alarm identifies a second electrical metering device.

In another example, a method includes accessing a first set of voltage measurements measured at an electrical metering device. The method further includes calculating a first corresponding average voltage and a first corresponding minimum voltage from the first set of voltage measurements and for each time window in the first set of time windows. The method also 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 characteristic corresponding to the electrical anomaly from the voltage measurements. The method also includes receiving a first classification from the machine learning model indicating a first loose connection. The method also includes sending a first alert to the utility operator based on the first classification. The method further includes calculating a second corresponding average voltage and a second corresponding minimum voltage from the second set of voltage measurements and for each time window in the second set of time windows. The second set of time windows may occur before the first plurality of time windows. The method also 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 also includes receiving a second classification from the machine learning model identifying a second voltage characteristic indicative of a second loose connection. The method also includes sending a second alert to the utility operator based on the second classification.

These illustrative examples are mentioned not to limit or define the disclosure, but to provide examples to aid in understanding the disclosure. Additional examples and further description are provided in the detailed description.

Drawings

These and other features, aspects, and advantages of the present disclosure will become better understood when the following detailed description is read with reference to the accompanying drawings in which:

Fig. 1 illustrates an exemplary communication network topology of a power distribution system in accordance with aspects of the present disclosure.

Fig. 2 illustrates an exemplary power distribution network in accordance with aspects of the present disclosure.

FIG. 3 depicts a flowchart of an example process for detecting anomalies using a machine learning model, according to aspects of the present disclosure.

Fig. 4 depicts calculating statistical voltage data obtained from a metering device in accordance with aspects of the present disclosure.

FIG. 5 depicts a flowchart of an example process for detecting anomalies using a machine learning model, according to aspects of the present disclosure.

FIG. 6 depicts a flowchart of an example process for training a machine learning model to detect anomalies using supervised learning, in accordance with aspects of the present disclosure.

Fig. 7 depicts a graph indicating voltage measurements associated with loose connections, in accordance with an aspect of the present disclosure.

Fig. 8 depicts a graph indicating voltage measurements associated with seasonal overload in accordance with an aspect of the present disclosure.

Fig. 9 depicts a graph indicating voltage measurements associated with excessive voltage drops due to long secondary lines, in accordance with an aspect of the present disclosure.

FIG. 10 illustrates an exemplary computing device in accordance with aspects of the present disclosure.

Detailed Description

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 or features in voltage measurements obtained by metering devices or meters located at end-user premises. Anomalies may include, but are not limited to, loose connections between meters and meter sockets at the end user premises, seasonal overloads (e.g., only seasonal overloads), and long secondary lines (e.g., connections from distribution transformers to the end user premises). Each of these anomalies can produce identifiable voltage characteristics at the end-user premises.

Accordingly, advantages of certain aspects include early identification of electrical anomalies, which may help to avoid failures, improve system efficiency, and improve system reliability in the form of an improved System Average Interrupt Frequency Index (SAIFI) or System Average Interrupt Duration Index (SAIDI) score. For example, once an anomaly is identified, action may be proactively taken to address the source of the anomaly and avoid unexpected interruptions, thereby bringing about these advantages.

Additionally or alternatively, after identifying an electrical anomaly, the disclosed system may retrospectively analyze metering data from one or more meters and determine additional patterns indicative of the anomaly. While an untrained machine learning model that does not know the voltage characteristics that identify anomalies may not recognize such patterns, a machine learning model that knows voltage characteristics may be trained to identify such voltage characteristics, in which regard the disclosed solution may provide early warning of developing electrical problems. For example, equipped with the ability to identify voltage characteristics that match a loose connection, a trained machine learning model may identify patterns in data from a corresponding meter months or years before the loose connection becomes a serious problem.

The following non-limiting examples are provided for illustrative purposes. The voltage measurements are collected at a particular frequency (e.g., every fifteen minutes) at the metering device. Examples of suitable metering devices are smart meters or Advanced Metering Infrastructure (AMI) meters. The metering device transmits the voltage data to the head-end system over a communication network, either together with or separately from metering data such as power consumption.

Continuing with the example, the headend system receives the voltage measurements from the metering device and derives statistics from the voltage measurements. For example, statistics such as daily average voltage and daily minimum voltage are calculated over a period of time (e.g., one month). The statistical data is provided to a machine learning model. Machine learning models previously trained to detect one or more anomalies from voltage data and/or derived statistical data determine the presence of anomalies, such as loose connections or seasonal overloads.

Turning now to the drawings, FIG. 1 illustrates an exemplary communication network topology of a power distribution system according to an embodiment. Fig. 1 includes a headend system 100; a network connection 108; a collector 110; network connections 112, 114, 116, and 118; metering devices 122, 124, 126, and 128; and end user premises 132, 134, 136 and 138. The communication network topology shown in fig. 1 depicts an example of how various devices may be interconnected in a communication network. The communication network topology differs from how power is allocated as shown in fig. 2.

In the example shown in fig. 1, one or more of metering devices 122, 124, 126, and 128 measure one or more parameters, such as voltage measurements, current measurements, phase measurements, active power consumption (watt hours), active power (watt), and provide parameters to collector 110 via network connections 112, 114, 116, and 118. The collector 110 then provides the metering data to the head-end system 100. The headend system 100 then processes the data to detect anomalies using machine learning.

Headend system 100 includes computing device 102 and machine learning model 104. An example of a computing device is depicted with respect to fig. 9. The computing device 102 may derive statistical data from the voltage data collected by the metrology devices 122, 124, 126, and 128 and/or provide the statistical data to the machine learning model 104, which in turn may identify one or more electrical anomalies at the metrology devices 122, 124, 126, and 128. Examples of anomalies that the machine learning model 104 may detect include loose connections (e.g., of meter sockets), seasonal overloads, and excessive voltage drops due to long secondary distribution lines. Exemplary voltage characteristics that may represent these cases are shown in fig. 6, 7, and 8, respectively.

Metering devices 122, 124, 126, and 128 are located at end user premises 132, 134, 136, and 138, respectively. Metering devices 122, 124, 126, and 128 may each include computing devices, memory, other storage capabilities, and one or more network connections. The metering devices 122, 124, 126, and 128 may detect any parameter related to the distribution of power, such as voltage, load, current, power consumption, or volt-ampere reactive (VAR) power. In some cases, each metering device 122, 124, 126, and 128 may push voltage measurements to the head-end system 100. In other cases, the headend system may pull voltage measurements from the metering devices 122, 124, 126, and 128. In another aspect, metering devices 122, 124, 126, and 128 may be configured to push voltage measurements to head-end system 100 at a particular frequency, which may be different than the frequency of transmission of metering data (e.g., power consumption data). For example, the metering devices 122, 124, 126, and 128 may measure voltage at certain intervals (such as every 15 minutes or every hour), but may push power consumption data daily or monthly.

The functionality discussed with respect to headend system 100 may be implemented on any combination of headend system 100 (e.g., computing device 102), one or more of metering devices 122, 124, 126, and 128, and a cloud-based system (i.e., a server connected via a persistent network connection). Examples of systems include AMI systems and Meter Data Management (MDM) systems.

For purposes of example, network connection 108 is depicted as a wired network connection, while network connections 112, 114, 116, and 118 are depicted as wireless network connections. Examples of typical networks include wireless (e.g., wiFi, bluetooth, mesh, or cellular) and wired (e.g., ethernet or power line communication) networks. But different configurations are possible. For example, metering devices 122, 124, 126, and 128 may communicate with each other and/or with headend system 100 using any communication network. Communications may also be sent from head-end system 100 to metering devices 122, 124, 126, and 128. In some configurations, collector 110 may be absent and metering devices 122, 124, 126, and 128 are in direct communication with headend system 100.

Fig. 2 illustrates an exemplary power distribution network in accordance with aspects of the present disclosure. In the example shown in fig. 2, the power distribution system 200 includes a power source 202; a distribution substation transformer 204; a feeder 208; distribution transformers 210 and 212; distribution primary lines 240 and 241, secondary lines 242, 244, 246, and 248; metering devices 222, 224, 226, and 228; and end user premises 232, 234, 236 and 238.

One example of a power source 202 is a simplified representation of a high capacity power system to which a distribution substation transformer 204 is connected, including a power sub-transmission network, a power transmission network, and a power generation source (e.g., a power plant, solar panel, or wind turbine generator). The distribution substation 204 converts the voltage output from the power source 202 to an appropriate level for the feeder 208. The feeder 208 in turn feeds the distribution lines 240 and 241. The distribution substation 210 converts the voltage of the distribution line 240 into different voltages in the secondary lines 242, 244 and 246. Similarly, distribution transformer 212 converts the voltage of distribution line 241 into a different voltage in secondary line 248.

As shown, metering devices 222, 224, 226, and 228 correspond to end user premises 232, 234, 236, and 238, respectively. Metering devices 222, 224, 226, and 228 may correspond to metering 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.

Metering device 222 measures a parameter on secondary line 242, metering device 224 measures a parameter on secondary line 244, metering device 226 measures a parameter on secondary line 246, and metering device 228 measures a parameter on secondary line 248. Because the length of each distribution line may be different, in some cases, anomalies may occur due to long secondary lines. For example, secondary line 248 may be longer than a typical line, resulting in an abnormally high voltage drop between distribution transformer 212 and end user premises 238. The anomaly may be identified by the machine learning model 104.

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

FIG. 3 depicts a flowchart of an example process 300 for detecting anomalies using a machine learning model, according to aspects of the present disclosure. For purposes of example, process 300 is discussed as being performed by headend system 100. Process 300 may be implemented on any computing device, such as in a metering device.

At block 302, the process 300 involves accessing a set of voltage measurements measured at a metering device over a period of time. For example, headend system 100 accesses voltage measurements obtained from metering device 122. Accessing the voltage measurement may include sending a request to the electrical metering device and receiving a back voltage measurement from the electrical metering device. In another example, the electrical metering device may periodically send the voltage measurements to the head-end system. For purposes of example, process 300 is discussed with respect to fig. 4.

Fig. 4 depicts calculating statistical voltage data obtained from a metering device in accordance with aspects of the present disclosure. Fig. 4 depicts a data flow 400 that depicts how average voltage data and minimum voltage data for various time periods are derived from data obtained by a single metering device. The data stream 400 includes voltage data 410, time windows 430a-n, statistics 440a-n, and the machine learning model 104 obtained over a time period 420. Any length of time period and any length of time window may be used.

Continuing with the example, computing device 102 accesses voltage data 410 corresponding to time period 420 and divides voltage data 410 into a plurality of time windows 430a-n. Time windows 430a-n are less than time period 420. Each of the time windows 430a-n includes a plurality of voltages. For example, if the particular time window is one day and the frequency at which voltages are obtained is every fifteen minutes, one time window will include ninety-six voltage measurements. In an example, the time period 420 is one month and the time windows 430a-n are each one day. For example, time window 430a is the first day, time window 430b is the second day, and so on.

Returning to fig. 3, at block 304, process 300 involves calculating a corresponding average voltage and a corresponding minimum voltage from the set of voltage measurements and for each time window in a set of time windows. Continuing with the example, computing device 102 calculates statistics 440a-n from time windows 430 a-n. More specifically, the computing device 102 calculates statistics 440a including the average and minimum values from the time window 430a, statistics 440b including the average and minimum values from the time window 430b, and so on.

Although average and minimum values are discussed with respect to fig. 4, the machine learning model 104 may use other statistical metrics derived from voltages or other parameters, such as median, mode, maximum, etc. In this case, the machine learning model 104 is trained using these statistics.

At block 306, the process 300 involves applying a machine learning model to the average voltage and the minimum voltage over a period of time. Continuing with the example, computing device 102 provides statistics 440a-n to machine learning model 104.

The machine learning model 104 is trained to identify one or more electrical anomalies from the voltages, as further discussed with respect to fig. 6. In some cases, one or more of the average voltage and the minimum voltage are converted into one or more feature vectors. A feature vector is a vector that contains a plurality of elements about an object (e.g., an electricity meter). Thus, the feature vector may include average voltage, minimum voltage, or other statistical data. The feature vectors are then provided to the machine learning model 104 one or more at a time.

At block 308, the process 300 involves receiving a classification from the machine learning model indicative of the identified electrical anomaly. Continuing with the example, the machine learning model 104 outputs a classification that identifies an electrical anomaly. Examples of anomalies include loose connections to the power line at the end user premises, seasonal overloads of consumption, and voltage drops caused by long secondary lines (connections from distribution transformers to the end user premises).

One example of an anomaly is a loose connection associated with an electrical metering device, as discussed further with respect to fig. 6. The loose connection may be represented by a first decrease in the minimum voltage over the period of time and a second decrease in the average voltage over the period of time, wherein the second decrease is less than the first decrease.

Another example of an anomaly is seasonal overload caused by power consumption measured by an electrical metering device, as discussed further with respect to fig. 7. Such anomalies may 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.

The machine learning model 104 may output a classification of one or more particular anomalies. In other cases, the machine learning model 104 may output a determined probability of one or more classifications. For example, the machine learning model 104 may output a first probability that the voltage indicates 80% of loose connections and a probability that no loose connections are present of 20%. In other cases, a set of probabilities may be generated, such as a probability of 50% loose connections, a probability of 40% long secondary lines, and a probability of 10% of both anomalies not present. The machine learning model that generates a particular classification (e.g., positive or negative) may be a different type of model than the machine learning model that generates the probability range.

The output of the machine learning model 104 is provided to the computing device 102. In some cases, computing device 102 may determine that when the probability is above a certain threshold, the corresponding classification is used. In some cases, computing device 102 may determine not to classify when the probability is below a threshold.

At block 310, the process 300 involves sending an alert to a utility operator or adjusting one or more parameters of the power distribution system based on the classification. For example, the computing device 102 may send an alert to a utility operator or an alert to an engineer to access a customer premises to perform a repair, such as replacing a bad meter socket or meter. Thus, the electrical load on the line may be rebalanced, additional equipment may be added, or equipment (e.g., transformers) may be replaced.

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

In an aspect, data from a plurality of metrology devices may be analyzed by the machine learning model 104. For example, the process 300 may be performed multiple times, once for each metering device. Alternatively, blocks 302-304 may be performed multiple times, once for each metering device, and then data from the multiple metering devices is analyzed in aggregate by the machine learning model 104, for example, at block 308. The analysis may occur in real time or after a threshold amount of data has been buffered.

In 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 and each voltage measurement and for each time window. The headend system 100 then applies the machine learning model 104 to the statistical metrics. In some cases, the application may result in an adjustment of the training of the machine learning model. Continuing with the example, headend system 100 identifies one or more anomalies using one or more electrical metering devices.

In some cases, the machine learning model may use topology information of the electric metering device. For example, given terrain information (e.g., as shown in fig. 2), the machine learning model 104 may determine whether an anomaly is present on one metrology device and not on another metrology device. From this determination, the machine learning model 104 may identify a problem on another component, such as a distribution transformer. For example, if a particular anomaly is shown on metering devices 222, 224, and 226 but not on metering device 228, then an anomaly may exist on distribution transformer 210 or distribution line 240.

As discussed further with respect to fig. 9, another example of an anomaly is an excessive voltage drop due to a long distribution line associated with the electrical metering device. In some cases, determining the excess voltage drop involves collecting measurements from an additional electricity meter. For example, the topology of such an electrical power network may be that one or more distribution transformers are electrically connected to and upstream of a subset of the electrical metering devices via distribution lines. One identifiable anomaly may be identified by a difference between an average voltage of a subset of the electrical metering devices and a daily average of the individual electrical metering devices being above a threshold.

FIG. 5 depicts a flowchart of an example process 500 for detecting anomalies using a machine learning model, in accordance with an aspect of the present disclosure. With respect to process 300 discussed with respect to fig. 3, process 500 involves using a first identification of an anomaly (optionally including associated voltage data) in conjunction with machine learning to identify a second identification of the anomaly. The second identification of anomalies may correspond to anomalies occurring earlier in time within the data than the first identification.

For example, the process 500 may detect loose connections in a first set of voltage data and then use the detected voltage characteristics in the same or a different machine learning model to identify second voltage characteristics in statistical data (e.g., voltage data) derived from the same or a different meter. For example, a loose connection may already exist for a period of time, but not detected until the first classification. Given the first classification, earlier identification can be performed retrospectively and/or on different data sets.

At block 502, the process 500 involves learning a model from a machine and receiving a first classification indicative of a first loose connection based on a first set of voltage measurements. At block 502, process 500 involves operations similar to those discussed with respect to blocks 302-308 of process 300. The classification may include voltage characteristics. The process 500 may optionally involve sending an alert to the utility operator based on the first classification, as discussed with respect to block 310 of the process 300.

At block 504, the process 500 involves calculating a second corresponding average voltage and a second corresponding minimum voltage from the second set of voltage measurements and for each time window in the second set of time windows. At block 504, process 500 involves operations similar to those discussed with respect to block 304 of process 300. The second set of voltages may be from the same meter from which the voltage data used at block 502 originated and/or may be from a different meter. The second set of voltages may occur before the first set of time windows (e.g., as discussed with respect to block 304 of process 300).

At block 506, the process 500 involves applying a machine learning model to the first voltage characteristic, the second average voltage, and the second minimum voltage. Continuing with the example, the computing device 102 provides the first voltage characteristic (e.g., identified at block 502), the first average voltage (e.g., identified at block 502), and one or more of the first minimum voltage, the second average voltage, and the second minimum voltage to the machine learning model 104. In some cases, a different machine learning model may be used than the machine learning model used at block 502. The machine learning model may be trained to identify one or more electrical anomalies from the voltages, as further discussed with respect to fig. 6.

At block 508, the process 500 involves receiving a second classification from the machine learning model identifying a second voltage characteristic indicative of a second loose connection. Continuing with the example, computing device 102 receives a second classification identifying a second voltage characteristic. The second voltage characteristic may be the same as, similar to, or different from the first voltage characteristic.

At block 510, the process 500 involves sending an alert to the utility operator based on the second classification. At block 510, process 500 involves operations similar to those discussed with respect to block 310 of process 300.

The machine learning model 104 is trained to detect one or more predetermined voltage characteristics that each correspond to one or more anomalies. Different learning techniques may be used, such as supervised learning (e.g., with labeled training data), unsupervised learning (e.g., without labels), or reinforcement learning. Training may be performed by headend system 100 (e.g., computing device 102) or a different computing system. Where training is performed by a different computing system, the machine learning model 104 may be provided (e.g., downloaded) to the head-end system 100 and/or updated as needed.

In some cases, training may be performed at runtime. For example, the operator may indicate that the classification or prediction of the anomaly is correct or incorrect and provide feedback to the computing device 102, which in turn, updates the machine learning model 104 accordingly.

During training, the machine learns a model learning algorithm to identify electrical anomalies. Although fig. 6 is discussed with respect to supervised learning, other learning techniques may be used. Additionally, the trained machine learning algorithm may be improved over time, for example by additional training at runtime.

FIG. 6 depicts a flowchart of an example process 600 for training a machine learning model to detect anomalies using supervised learning, in accordance with aspects of the present disclosure. In a supervised learning approach, a determined probability or class is calculated and compared to an expected or known probability or class. The loss function is calculated based on the comparison. Based on the calculated loss function, the machine learning model 104 adjusts internal parameters of the machine learning model to minimize the loss function. Examples of suitable machine learning models include neural networks, classifiers, and decision trees.

At block 602, process 600 involves accessing a set of training data pairs. Each training data pair includes statistical data (e.g., a set of average voltages and a set of minimum voltages over a period of time) and an expected classification indicative of one or more electrical anomalies. Each training data pair includes data previously identified as part of a positive training set (i.e., corresponding to voltage data corresponding to previously identified voltage features) and/or a negative training set (i.e., not corresponding to previously identified voltage features). Process 600 is discussed as being performed by computing device 102. However, training may be performed by any computing system.

At block 604, the process 600 involves providing one of the set of training data pairs to a machine learning model. Continuing with the example, computing device 102 provides one of the training data pairs to machine learning model 104.

At block 606, the process 600 involves receiving the determined classification from the machine learning model. Continuing with the example, computing device 102 receives the determined classification from machine learning model 104.

At block 608, the process 600 involves calculating a loss function by comparing the determined classification to an expected classification. Continuing with the example, computing device 102 calculates a loss function by comparing the determined classification (i.e., the classification received at block 606) to an expected classification (i.e., comprising the training data pair accessed at block 604).

At block 610, process 600 involves adjusting internal parameters of the machine learning model to minimize the loss function. Continuing with the example, computing device 102 adjusts machine learning model 104 appropriately.

At block 612, process 600 involves checking whether training is complete. If training is complete, process 600 continues to block 614 where training ends at block 614. If training is not complete, process 600 returns to 604 and continues to train the machine learning model using another training data pair. Completion of training may be indicated by the end of training data, the loss function being minimized below a threshold level, or some other condition.

In an aspect, the machine learning model 104 is trained to detect a plurality of voltage characteristics. In this case, the machine learning model 104 may identify electrical features indicative of more than one electrical anomaly. For example, the machine learning model 104 may identify loose receptacle connections and seasonal overloads. In this case, process 600 may be completed one or more times for each electrical anomaly.

In another aspect, the machine learning model 104 is tested to ensure sufficient accuracy in the classification of electrical anomalies. Typically, the data for testing the machine learning model 104 is not included in the data set (i.e., training data pairs) used to train the machine learning model.

In another aspect, the machine learning model 104 may be trained in an unsupervised manner using historical data from the metrology device. The historical data may include data for a period of time prior to a period of time during which the defect was initially identified. For example, after a particular metering device is identified as exhibiting an anomaly (e.g., by analyzing a voltage from the metering device, as described in process 300), computing device 102 and/or machine learning model 104 may analyze additional data from the metering device to determine whether any additional voltage characteristics are present. For example, features may be identified in the historical data, and then an earlier time at which the defect occurred may be identified. Using this approach, the machine learning model 104 may identify one or more precursor patterns in the collected voltage data that may be earlier than previously detectable using the supervised technique (e.g., process 600).

In a more detailed example, process 300 is used to identify five metering devices having loose connections. For example, process 300 processes historical voltage data for one month. The five metering devices are secured by a service person tightening screws on the base of the metering devices. The process of identifying metering devices having loose connections may continue.

Continuing with this example, after a few months, twenty meters are identified and repaired. But for each metering device historical data for each meter over a period of one year is available. By training the machine learning model 104 using this historical data, the machine learning model 104 identifies one or more additional precursor features that are common to all twenty meters. The features so identified may supplement and/or replace the features identified in process 500. In this way, training may continue over time in an unsupervised or supervised manner. Benefits of such a continuous training method include identifying voltage characteristics that may occur at unpredictable or unexpected times, thereby improving the machine learning model 104.

Fig. 7 depicts a graph indicating voltage measurements associated with loose connections, in accordance with an aspect of the present disclosure. Graph 700 depicts a daily average voltage 710 and a daily minimum voltage 720 measured by a metering device at an end user premises. The data corresponding to graph 700 is shown in table 1 below.

As can be seen in graph 700, a large difference between daily average voltage 710 and daily minimum voltage 720 may exist at some points. These differences indicate the presence of an arc at the junction or frequent disconnection of the junction. The measurements in graph 700 may together form a voltage signature indicative of a loose receptacle connection. For example, the magnitude of the decrease in minimum voltage 720 relative to average voltage 710 may indicate a loose connection. The machine learning model 104 is trained to identify these features as anomalies.

Fig. 8 depicts a graph indicating voltage measurements associated with seasonal overload in accordance with an aspect of the present disclosure. Graph 800 depicts a daily average voltage 810 and a daily minimum voltage 820 measured by a metering device at an end user premises. The data for graph 800 is shown in table 2 below. Here, the period of time used is a year, and the period of time is a month.

The data shows voltage measurements from one month to december. It can be seen that during the period of 4 months to 9 months, there is a higher load, as indicated by a lower voltage. These voltage drops may form a feature that indicates overload of the transformer. For example, the relative alignment of the peaks and valleys in the minimum voltage 820 with respect to the peaks and valleys of the average voltage 810 may indicate seasonal overload. The machine learning model 104 is trained to identify these features as anomalies.

Fig. 9 depicts a graph indicating voltage measurements associated with excessive voltage drops due to long secondary lines, in accordance with an aspect of the present disclosure. Graph 900 depicts a daily average voltage 910 of all meters (e.g., metering devices 222, 224, and 226) behind the service transformer as compared to a daily average voltage 920 measured by a metering device (metering device 226) at the end user premises. A graph 900 of the measurement results for one month is shown as shown in table 3 below.

As can be seen from graph 900, there is a large separation between the daily average voltages in all meters behind the service transformer compared to the daily average voltage of meters with long secondary lines. In order to discern this feature, some topology information is required. The topology information associates the metering device with the transformer in order to use the voltage to discern the particular mode. The machine learning model 104 may discern the characteristic by comparing the average voltage of all metering devices behind the transformer to the daily average for each metering device.

/>

FIG. 10 illustrates an exemplary computing device in accordance with aspects of the present disclosure. Any suitable computing system may be used to perform the operations described herein. The depicted example of 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 device 1004, accesses data 1020 stored in the memory device 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 any number of processing devices or cores, including a single processing device. The functionality of the computing device may be implemented in hardware, software, firmware, or a combination thereof.

In some aspects, the computing device 1000 may include at least one sensor configured to measure a parameter related to a resource of the resource allocation network. For example, in a power distribution system, sensors may measure power consumption, voltage, current, and the like. In some aspects, computing device 1000 may include a plurality of sensors. For example, computing device 1000 may include both power and temperature sensors.

Memory device 1004 includes any suitable non-transitory 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 a processor. Non-limiting examples of computer readable media include flash memory, ROM, RAM, ASIC, or any other medium from which a processing device may read instructions. The instructions may include processor-specific instructions generated by a compiler or interpreter from code written in any suitable computer programming language, including, for example, C, C ++, c#, visual Basic, java, or scripting languages.

Computing device 1000 may also include a number of external or internal devices, such as input or output devices. For example, computing device 1000 is shown with one or more input/output ("I/O") interfaces 1008. The I/O interface 1008 may receive input from an input device or provide output to an output device. One or more buses 1006 are also included in computing device 1000. Bus 1006 communicatively couples one or more components of a respective one of computing devices 1000.

Computing device 1000 executes program code 1030, program code 1030 configuring processor 1002 to perform one or more operations described herein.

Computing device 1000 also includes a network interface device 1010. The network interface device 1010 includes any device or group of devices adapted to establish a wired or wireless data connection to one or more data networks. The network interface device 1010 may be a wireless device and has an antenna 1014. Computing device 1000 may communicate with one or more other computing devices implementing computing devices or other functions via a data network using network interface device 1010.

Computing device 1000 may also include a display device 1012. The display device 1012 may be an LCD, LED, touch screen, or other device operable to display information about the computing device 1000. For example, the information may include an operational state of the computing device, a network state, and the like.

While the subject matter has been described in detail with respect to specific aspects thereof, it will be appreciated that those skilled in the art, upon attaining an understanding of the foregoing, may readily produce alterations to, variations of, and equivalents to such aspects. It should therefore be understood that the present disclosure has been presented for purposes of example and not limitation, and that such modifications, variations and/or additions to the subject matter are not excluded as would be obvious to a person of ordinary skill in the art.

Claims

1. A method of detecting electrical anomalies in a power distribution system using machine learning, the method comprising:

Accessing a first plurality of voltage measurements measured at an electrical metering device;

Calculating a first corresponding average voltage and a first corresponding minimum voltage from the first plurality of voltage measurements and for each time window of a first plurality of time windows;

Applying a machine learning model to the first average voltage and the first minimum voltage, wherein the machine learning model is trained to identify first voltage features corresponding to the electrical anomaly from voltage measurements;

receiving a first classification from the machine learning model indicative of a first loose connection;

based on the first classification, sending a first alert to a utility operator;

calculating a second corresponding average voltage and a second corresponding minimum voltage from a second plurality of voltage measurements and for each time window of a second plurality of time windows, wherein the second plurality of time windows occurs before the first plurality of time windows;

applying the machine learning model to the first average voltage, the first minimum voltage, the first voltage characteristic, the second average voltage, and the second minimum voltage;

Receiving a second classification from the machine learning model identifying a second voltage characteristic indicative of a second loose connection; and

Based on the second classification, a second alert is sent to the utility operator.

2. The method of claim 1, wherein the second plurality of voltage measurements are measured at an additional electrical metering device.

3. The method of claim 1, wherein the first voltage characteristic comprises a first decrease in minimum voltage over a period of time and a second decrease in average voltage over the period of time, and wherein the second decrease is less than the first decrease.

4. The method of claim 1, wherein the second plurality of voltage measurements are measured at the electrical metering device.

5. The method of claim 1, further comprising applying the machine learning model to topology information that associates the electrical metering device with one or more distribution transformers electrically connected to the electrical metering device via a distribution line.

6. The method of claim 1, wherein accessing the first plurality of voltage measurements comprises sending a request to the electrical metering device and receiving the first plurality of voltage measurements from the electrical metering device.

7. The method of claim 1, further comprising training the machine learning model by:

A set of training data pairs is accessed, wherein each training data pair includes one or more of: (i) a training set of average voltages and a set of minimum voltages, (ii) a training set of average voltages of all electrical metering devices connected to the distribution transformer, or (iii) a training set of average voltages of one electrical metering device behind the distribution transformer and an expected classification indicative of one or more electrical anomalies;

Providing one of the set of training data pairs to the machine learning model;

receiving a determined classification from the machine learning model;

Calculating a loss function by comparing the determined classification with the expected classification; and

Internal parameters of the machine learning model are adjusted to minimize the loss function.

8. The method of claim 7, wherein the set of training data pairs further comprises topology information that associates one or more metering devices with one or more distribution transformers.

9. A non-transitory computer-readable storage medium storing computer-executable program instructions, wherein the computer-executable program instructions, when executed by a processing device, cause the processing device to perform operations comprising:

applying a machine learning model to the first average voltage and the first minimum voltage, wherein the machine learning model is trained to identify first voltage features corresponding to electrical anomalies from voltage measurements;

Receiving a first classification from the machine learning model indicative of a first anomaly;

based on the first classification, sending a first alert to a utility operator;

Applying the machine learning model to the first average voltage, the first minimum voltage, the second average voltage, and the second minimum voltage;

receiving a second classification from the machine learning model identifying a second voltage characteristic indicative of a second anomaly; and

10. The non-transitory computer-readable storage medium of claim 9, wherein one or more of the first anomaly and the second anomaly relate to a loose connection associated with the electrical metering device.

11. The non-transitory computer-readable storage medium of claim 9, wherein one or more of the first anomaly and the second anomaly are represented by a first decrease in a minimum voltage over a period of time and a second decrease in an average voltage over the period of time, and wherein the second decrease is less than the first decrease.

12. The non-transitory computer-readable storage medium of claim 9, wherein one or more of the first anomaly and the second anomaly 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.

13. The non-transitory computer readable storage medium of claim 9, wherein the computer executable program instructions, when executed by a processing device, cause the processing device to apply the machine learning model to topology information that associates the electrical metering device with one or more distribution transformers electrically connected to the electrical metering device via a distribution line.

14. The non-transitory computer-readable storage medium of claim 9, wherein accessing the first plurality of voltage measurements includes sending a request to the electrical metering device and receiving the plurality of voltage measurements from the electrical metering device.

15. The non-transitory computer readable storage medium of claim 9, wherein the computer executable program instructions, when executed by a processing device, cause the processing device to train the machine learning model by:

A set of training data pairs is accessed, wherein each training data pair includes one or more of: (i) a set of average voltages and a set of minimum voltages, (ii) a set of average voltages of an electrical metering device connected to a distribution transformer, or (iii) a set of average voltages of one electrical metering device behind the distribution transformer and an expected classification indicative of one or more electrical anomalies;

Providing one of the set of training data pairs to the machine learning model;

receiving a determined classification from the machine learning model;

16. A system for detecting an electrical anomaly in a resource allocation system, the system comprising:

a headend system comprising a computing device and a machine learning model; and

A plurality of electrical metering devices, each electrical metering device comprising a sensor, each electrical metering device electrically connected to a distribution transformer upstream of a respective electrical metering device, each electrical metering device configured to:

Obtaining a respective plurality of voltage measurements from respective sensors of the electrical metering device; and

Providing the respective plurality of voltage measurements to the headend system, wherein the headend system includes a machine learning model and is configured to:

Accessing a first plurality of voltage measurements measured at a first one of the plurality of electrical metering devices;

Based on the first classification, sending a first alert to a utility operator, wherein the first alert identifies the first electrical metering device;

accessing a second plurality of voltage measurements measured at a second one of the plurality of electrical metering devices;

Calculating a second corresponding average voltage and a second corresponding minimum voltage from the second plurality of voltage measurements and for each time window of a second plurality of time windows, wherein the second plurality of time windows occurs before the first plurality of time windows;

Based on the second classification, a second alert is sent to the utility operator, wherein the alert identifies the second electrical metering device.

17. The system of claim 16, wherein the headend system is further configured to further train the machine learning model with the plurality of voltage measurements from at least one electrical metering device.

18. The system of claim 16, wherein the headend system is further configured to:

Applying topology information to the machine learning model, the topology information associating the plurality of electrical metering devices with one or more distribution transformers electrically connected to and upstream of the plurality of electrical metering devices via distribution lines, and

Wherein the first anomaly is a difference between an average voltage of the plurality of electrical metering devices and a daily average of the one of the plurality of electrical metering devices being above a threshold.

19. The system of claim 16, wherein the first voltage characteristic comprises a first decrease in a minimum voltage over a period of time and a second decrease in an average voltage over the period of time, and wherein the second decrease is less than the first decrease.

20. The system of claim 16, wherein receiving the plurality of voltage measurements comprises sending a request to a respective electrical metering device and receiving the plurality of voltage measurements from a respective electrical metering device.