Load prediction of electrical equipment using machine learning

Abstract

Several embodiments are disclosed for predicting, by a processor circuit, a load parameter value for a future time for an electrical device based on at least one machine learning model and a plurality of load parameter values, the plurality of load parameter values comprising a predefined number of load parameter values extracted from a time series data stream of load parameter values obtained for the electrical device. Thereafter, an overload capacity for the future time is calculated by the processor circuit based on the predicted load parameter value, and at least one parameter associated with the electrical device at a current time is changed by the processor circuit based on the calculated overload capacity for the future time.

Description

Technical Field

[0001] This disclosure relates to the analysis of electrical equipment (such as high-voltage transformers). In particular, this disclosure relates to the use of machine learning for load forecasting of electrical equipment. Background Technology

[0002] Conventional load forecasting techniques for electrical equipment (such as transformers) typically employ sophisticated statistical tools to estimate short- and long-term forecasts of load, hotspot temperatures, overload capacity, and other parameters. These tools (such as Autoregressive Integrated Moving Average (ARIMA)) decompose time series of historical load data into several components, such as trend components, seasonal variation components, and stochastic variation components, to derive a complex model that aggregates these components into a final forecast. To reduce forecast errors, these tools may also incorporate additional relevant parameters, such as daily average temperature, seasonality, holidays, and sporting events. This adds further complexity and may require specialized tools and data for different times and regions (e.g., in countries or regions with different cultures, holidays, climates, etc.).

[0003] For example, in the case of transformers, accurate estimates of load for future times (such as one or two hours in advance) are desirable in order to plan for load shedding, unexpected overload demands, and the eventual removal of transformers from operation based on their service life and remaining lifespan, or the ever-changing needs of the power system. However, due to the complexity of existing solutions, many types of electrical equipment lack the computational power required to accurately estimate future loads. Summary of the Invention

[0004] According to some embodiments, a method includes: a processor circuit predicting load parameter values ​​for a future time for an electrical device based on at least one machine learning model and a plurality of load parameter values, said plurality of load parameter values ​​including a predefined number of load parameter values ​​extracted from a time-series data stream of load parameter values ​​obtained for the electrical device. The method further includes: the processor circuit calculating an overload capacity for a future time based on the predicted load parameter values. The method further includes: the processor circuit changing at least one parameter associated with the electrical device at the current time based on the calculated overload capacity for the future time.

[0005] According to some embodiments, the at least one machine learning model is trained based on multiple defined relationships between a predefined number of load parameter values ​​and at least one subsequent load parameter value from a time-series data stream obtained within a predetermined time period.

[0006] According to some embodiments, the plurality of determined relationships are verified based on a comparison of at least one expected load parameter value derived from a predefined number of load parameter values ​​with the at least one subsequent load parameter value.

[0007] According to some embodiments, a predefined number of load parameter values ​​are used as the input set of the at least one machine learning model to successively predict load parameter values ​​for future times. This input set is generated successively from a time-series data stream of load parameter values ​​using a moving window technique.

[0008] According to some embodiments, the plurality of load parameter values ​​includes a set of at least five load parameter values ​​iteratively extracted from a stream of load parameter values ​​obtained from electrical equipment.

[0009] According to some embodiments, the predicted load parameter values ​​are for a future time of at least one hour after the prediction.

[0010] According to some embodiments, the at least one machine learning model predicts the load parameter values ​​of electrical equipment for future times based solely on the plurality of load parameter values.

[0011] According to some embodiments, the at least one machine learning model predicts the load parameter values ​​of the electrical equipment for future time based on the plurality of load parameter values ​​and at least one temperature parameter associated with the electrical equipment.

[0012] According to some embodiments, the method further includes: predicting, by processor circuitry, at least one hot spot temperature value for a component of an electrical device for a future time based on predicted load parameter values.

[0013] According to some embodiments, the method further includes: a processor circuit predicting at least one overload capacity value of an electrical device based on predicted load parameter values, the at least one overload capacity value being associated with at least one time period after a future time.

[0014] According to some embodiments, the electrical equipment includes a transformer, and the method further includes operating the transformer based at least in part on the at least one modified parameter.

[0015] According to some embodiments, operating the transformer includes operating at least one cooling component of the transformer in response to predicted load parameter values ​​to change the temperature of at least one component of the electrical equipment before a future time.

[0016] According to some embodiments, a monitoring device includes processor circuitry and a memory including machine-readable instructions. When executed by the processor circuitry, the machine-readable instructions cause the processor circuitry to predict load parameter values ​​of an electrical device for a future time based on at least one machine learning model and a plurality of load parameter values, said plurality of load parameter values ​​including a predefined number of load parameter values ​​extracted from a time-series data stream of load parameter values ​​obtained for the electrical device. The machine-readable instructions further cause the processor circuitry to calculate an overload capacity for a future time based on the predicted load parameter values. The machine-readable instructions further cause the processor circuitry to change at least one parameter associated with the electrical device at the current time based on the calculated overload capacity for the future time.

[0017] According to some embodiments, the at least one machine learning model is trained based on multiple defined relationships between a predefined number of load parameter values ​​and at least one subsequent load parameter value from a time-series data stream obtained within a predetermined time period.

[0018] According to some embodiments, the plurality of load parameter values ​​includes a set of at least five load parameter values ​​iteratively extracted from a stream of load parameter values ​​obtained from electrical equipment.

[0019] According to some embodiments, the predicted load parameter values ​​are for a future time of at least one hour after the prediction.

[0020] According to some embodiments, a non-transitory computer-readable medium includes instructions that, when executed by processor circuitry, cause the processor circuitry to predict load parameter values ​​of an electrical device for a future time based on at least one machine learning model and a plurality of load parameter values, said plurality of load parameter values ​​including a predefined number of load parameter values ​​extracted from a time-series data stream of load parameter values ​​obtained for the electrical device. These instructions further cause the processor circuitry to calculate an overload capacity for a future time based on the predicted load parameter values. These instructions further cause the processor circuitry to change at least one parameter associated with the electrical device at the current time based on the calculated overload capacity for the future time.

[0021] According to some embodiments, the at least one machine learning model is trained based on multiple defined relationships between a predefined number of load parameter values ​​and at least one subsequent load parameter value from a time-series data stream obtained within a predetermined time period.

[0022] According to some embodiments, the plurality of load parameter values ​​includes a set of at least five load parameter values ​​iteratively extracted from a stream of load parameter values ​​obtained from electrical equipment.

[0023] According to some embodiments, the predicted load parameter values ​​are for a future time of at least one hour after the prediction. Attached Figure Description

[0024] The accompanying drawings illustrate certain non-limiting embodiments of the inventive concept. These drawings are included to provide a further understanding of the present disclosure and are incorporated to form a part of this application. In the drawings:

[0025] Figure 1 The illustration depicts a planar file according to some embodiments, which is provided for use by machine learning applications for iterative application of multivariate algorithms;

[0026] Figure 2A and Figure 2B The illustration shows, according to some embodiments, the transformation of a univariate dataset representing historical load data into a flattened dataset for use with machine learning algorithms to predict future loads on electrical equipment;

[0027] Figure 3 The illustration depicts, according to some embodiments, the operation of training and selecting a machine learning model to predict the future load of electrical equipment based on historical load data;

[0028] Figure 4 The illustrations depict some embodiments. Figure 3 Visualization of operations on training and validation datasets for machine learning models;

[0029] Figures 5A-5C This is a graph illustrating a comparison of the actual load on electrical equipment over time with the predicted load over time using a machine learning model, according to some embodiments.

[0030] Figure 6 This is a flowchart, according to some embodiments, of an operation for predicting load parameter values ​​for future times based on historical load data;

[0031] Figure 7 The illustrations are based on some embodiments of use. Figure 6 Charts showing historical load over time, projected load for future times, and determined future overload capacity; and

[0032] Figure 8 This is a block diagram illustrating a load forecasting system for performing operations according to some embodiments. Detailed Implementation

[0033] The inventive concept will now be described more fully below with reference to the accompanying drawings, which illustrate examples of embodiments of the inventive concept. However, the inventive concept can be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the inventive concept to those skilled in the art. It should also be noted that these embodiments are not mutually exclusive. It may be tacitly assumed that components from one embodiment are present / used in another embodiment.

[0034] The following description presents various embodiments of the disclosed subject matter. These embodiments are presented as illustrative examples and are not to be construed as limiting the scope of the disclosed subject matter. For example, certain details of the described embodiments may be modified, omitted, or expanded without departing from the scope of the described subject matter.

[0035] Transformer overload capacity for the current time can be estimated based on existing dynamic thermal models and current ambient temperature and load conditions (which allow for continuous calculation of hotspot temperatures and the resulting lifespan degradation). To calculate overload capacity for a given future time interval, the model can assume that the hotspot temperature will not exceed given limits (e.g., 110°C for thermally upgraded kraft paper and 98°C for regular kraft paper) to avoid unnecessarily shortening the transformer's operating life. Based on the transformer's current condition, an optimal "k" load factor can be calculated, allowing the transformer to be overloaded at different levels for different corresponding times without sacrificing lifespan.

[0036] Accurately predicting future load values ​​can facilitate effective load management of electrical equipment and transmission / distribution lines in a power system. For example, in a power system, if load-bearing equipment has overload capacity in response to equipment failure in parallel distribution lines, load management can be accomplished by reallocating the load. It can also be used to prepare electrical equipment effectively for future conditions, such as providing sufficient cooling before anticipated overload conditions to extend the amount of time the equipment can be safely overloaded, or reducing cooling in anticipation of decreased demand. This is useful in larger transformers, where cooling can be a time-consuming process.

[0037] According to embodiments of this disclosure, a relatively small set of historical load data can be used to accurately estimate the future load of a transformer to train a machine learning model and efficiently select the machine learning model with the highest accuracy for different electrical devices. In some embodiments, future loads can be accurately estimated without any other external information, such as holidays, sporting events, time of year, etc. One advantage of this approach is that the complexity of the model is greatly reduced, allowing a wider variety of electrical and / or computing devices to adopt the model and reducing reliance on large-power computing devices that are remotely located and difficult to access.

[0038] Many machine learning applications (e.g., supervised learning problems) work on a flat file or sampled data stream. In this respect, Figure 1 An example of a flat file 100 is illustrated, which is used by machine learning applications to iteratively apply a multivariate regression algorithm to each row 1–m (using columns x1–xn as input and the last column as the target output) to determine and improve the algorithm over time.

[0039] As by Figure 2A and Figure 2B As shown, a univariate dataset 200 (in this example, a time series of loads against time) can be transformed into a multivariate flattened dataset 202. In this simplified example, the univariate dataset 200 provides fourteen consecutive load values ​​204 over times t1–t14, and each row 206 of the flattened dataset 202 uses a sequence of five consecutive load values ​​204 as input 208 to model the next load value in that sequence (i.e., the target value 210), such that each row 206 of the flattened dataset 202 uses a sequence of six consecutive load values ​​204 (e.g., a “moving window”). However, it should be understood that the choice of the number of variables (i.e., predictors) will depend on each individual problem. The number of variables in each row 206 can be determined and optimized based on additional testing (e.g., for sensitivity, model accuracy, hardware and software constraints, and other parameters).

[0040] To transform the univariate dataset 200 into a flattened dataset 202, the first row 206 of the flattened dataset is filled with loading values ​​204 from the univariate dataset 200 at times t1–t6, the second row 206 is filled with loading values ​​204 at times t2–t7, and so on, filling nine rows 206 of the flattened dataset, with the ninth row 206 filled with the last six loading values ​​204 at times t9–t14. In this way, the vertical univariate dataset 200 is transformed into a tabular multivariate flattened dataset suitable for use with many types of machine learning models.

[0041] One advantage of this data transformation technique is that it transforms univariate datasets (e.g., load versus time) into multivariate problems, which facilitates the use of many machine learning models suitable for regression or classification applications. The power of such machine learning models lies in the fact that they can "learn" from large datasets containing a large number of cases (or examples) and also a large number of features (or predictors or independent variables). In this example, the machine learning model performs regression-type predictions to predict a real number as the target value 210, but in other examples, it can perform classification-type predictions to predict categorical targets (such as "good or bad," "yes or no," level 1, 2, 3, etc.).

[0042] One advantage of using these and other machine learning techniques with transformer load data is that these techniques provide highly accurate predictions of future loads based on a relatively small univariate dataset of historical loads over time, without requiring any other external parameters such as temperature, holidays, events, etc.

[0043] In some embodiments, flattened data is used to train many different machine learning models (e.g., linear and nonlinear algorithms), and these results are compared to determine the machine learning model with the highest accuracy. Many different metrics can be used to determine accuracy, such as root mean square error (RMSE), mean absolute error (MAE), etc. Examples of suitable linear machine learning models may include general linear regression, logistic regression (e.g., for classification), linear discriminant analysis, etc. Examples of suitable nonlinear machine learning models may include classification and regression trees, Naive Bayes, K-nearest neighbors, support vector machines, etc. Examples of suitable ensemble machine learning models may include random forests, tree-bagging, extreme gradient boosting machines, artificial neural networks, etc.

[0044] In an illustrative example, a large dataset of approximately 120,000 load data points is used to train and test different machine learning models. This dataset represents hourly load data collected over a period of several years. In this example, the load data is divided into a set of training data (containing 80% of the data) and validation data (containing the remaining 20%). The training data is transformed into a flat file with 25 columns (i.e., 24 columns for input variables and one column for the target variable). The training data is then used to train various machine learning models. After training, the validation data is similarly transformed into a flat file with 25 columns, where the target column is removed. After predicting the target value for the validation data, the predicted values ​​are compared with the actual target values. This process is then repeated so that differences in how the machine learning models handle the data can be incorporated into the results, enabling high-precision predictions of short-term loads over the next 1-2 hours.

[0045] In one example Figure 3 The diagram illustrates the operation 300 for training and selecting a machine learning model to predict future loads on electrical equipment based on historical load data. See also: Figure 4 Its illustration shows Figure 3 Visualization of Operation 300 on the training and validation dataset 402 used for machine learning models.

[0046] Figure 3 Operation 300 includes: breaking down the complete dataset into training data and visualization data (box 302). For example, in Figure 4 In this dataset 402, it is divided into training data 404 (e.g., 80%) and validation data 406 (e.g., 20%). These operations further include: decomposing the training data into K random fragments (boxes 304). For example, Figure 4 The training data 404 is broken down into K fragments 408 of equal size (e.g., 10 fragments for the purposes of this example).

[0047] Next, the machine learning model is trained using K-1 fragments of the training data (box 306), and validated using the remaining fragments of the training data (box 308). The model's accuracy is annotated (box 310), and this process is repeated for several different machine learning models (box 312). For example, in... Figure 4 In this process, nine of the ten fragments 408 (e.g., fragments 1-9) are used to train each machine learning model, and the last fragment 408 (e.g., fragment 10) is used to test the accuracy of each trained machine learning model, wherein each fragment 408 of the training data 404 is reserved for testing each machine learning model once.

[0048] Then, for each group of K-1 fragments, repeat operations 306-310 (box 312), performing a total of K training and testing operations for each machine learning model (e.g., for...). Figure 4 Each of the 9 fragments in the 408 combination is performed 10 times for each machine learning model. Then, operations 304-314 (box 316) are repeated for a different set of K random fragments of the training data to obtain independent results. For example, for Figure 4 Repeating operations 306-310 a total of three times on training data 404 will result in a total of 30 training operations in this example. Operation 300 further includes using validation data (e.g., Figure 4 The validation data (406) is used to validate the trained machine learning model and to select the most accurate machine learning model (or multiple models) to predict future loads.

[0049] Figure 3These operations 300 have the advantage of minimizing "overtraining," where a given model processes the same data many times and assigns excessive weights to the training data, making its output almost 100% correct when faced with the training data again. This means the model is almost "unbiased" about the training set and is considered "overtrained." When faced with new, unseen data, an overtrained model may provide poor results because it can only perfectly represent the training data and not the new, unseen data. Generally speaking, trained machine learning often expects strong generalization capabilities, even at the cost of perfect accuracy with respect to the training data.

[0050] Now for reference Figures 5A-5C Charts 500A-500C illustrate the use of... Figure 3 Operation 300 trains and selects a machine learning model to compare the actual load values ​​502A-502C (solid lines) of electrical equipment over time with the predicted load values ​​504A-504C (dashed lines) over time. In this example, the datasets of actual load values ​​502A-502C from three different 1000-hour periods are selected from a large validation dataset covering thousands of hours of load data for transformers. Each dataset of actual load values ​​502A-502C is used to train the machine learning model to predict the predicted load values ​​504A-504C over time with high accuracy for each corresponding set of actual load values ​​502A-502C. It is worth noting that each set of predicted load values ​​504A-504C in this example is derived independently from each corresponding set of actual load values ​​502A-502C, without retraining or otherwise combining results from other training examples. Despite large variations between the different sets of actual load values ​​502A-502C, all three sets of predicted load values ​​504A-504C exhibit high accuracy. For example, Figure 5A The actual load value of 502A varies between approximately 11250MW and 21250MW. Figure 5B The actual load value of 502A varies between approximately 10,000MW and 17,000MW, and Figure 5C The actual load value 502C varied between approximately 12,500 MW and 22,500 MW. Despite the high differences between the actual load values ​​502A-502C across different groups, the corresponding predicted load 504A all provided an average prediction error between 1.13% and 1.26%, with 90% of all predicted load values ​​504A-504C providing an error rate of less than 2.55%. As a result, using the embodiments disclosed herein (including, for example...) Figure 3 The machine learning model trained and selected by the operation 300 can provide high-precision predicted load values ​​504A-504C, especially short-term predicted loads 506A-506C.

[0051] Now for reference Figure 6 The diagram illustrates a flowchart of operation 600 for predicting load parameter values ​​for future times based on historical load data. See also: Figure 7 The diagram shows the load over time 702, the predicted load parameter value for future time 706, and the use of... Figure 6 The future overload capacity determined by the operation is shown in Figure 700 for 710A-710C.

[0052] Operation 600 includes: predicting load parameter values ​​for a future time for an electrical device based on at least one machine learning model and multiple load parameter values ​​(box 602). In this example, the multiple load parameter values ​​are obtained by sampling the values ​​using a continuously moving window on a time-series data stream of load parameter values. Therefore, the multiple load parameter values ​​comprise a predefined number of load parameter values ​​extracted from the time-series data stream of load parameter values ​​obtained for the electrical device (e.g., a predetermined window size, such as those mentioned above). Figure 2A and Figure 2B The description refers to a moving window with five consecutive load values ​​of 204. For example, as described by... Figure 7 As shown, one or more machine learning models (e.g., using the above) Figure 3 The model (trained and selected using operations) is used to predict the load at the current time t0 for a future time t1 (e.g., one to two hours after the current time t0). The prediction in this example is based on a set of historical load values ​​704. In some examples, the load parameter values ​​can be the electrical and / or thermal load values ​​of electrical equipment. Electrical load can be measured by the load current value. Furthermore, in a transformer, the secondary current can be measured, and the electrical load of the transformer can be derived from the measured secondary current.

[0053] Thermal load can be measured by temperature values ​​corresponding to the heat generated by electrical losses within the electrical equipment. In a transformer, for example, temperature values ​​can be obtained from different components, such as, for example, the highest oil temperature, and can be used, together with the current load and thermal fingerprint of the individual transformer, to estimate future electrical loads and / or future associated hot spot temperatures that can indicate the contribution of excess heat to transformer aging. Other parameters can also be used to estimate operating temperatures, including output power, primary current, etc. Therefore, in some embodiments, operation 600 may further include: predicting at least one hot spot temperature value for a component of the electrical equipment for a future time based on predicted load parameter values ​​(box 604).

[0054] Operation 600 further includes: calculating the overload capacity for a future time based on the predicted load parameter values ​​(box 606). In some examples, calculating the overload capacity includes predicting at least one overload capacity value for the electrical equipment for a period of time after the future time based on the predicted load parameter values ​​(box 608). For example, as... Figure 7 As shown, different overload capacities 710A-710C can be calculated based on load parameter values ​​and / or other factors (such as ambient temperature, maximum oil level, thermal fingerprints, etc.). For example, based on a predicted load parameter value 706 for a future time t1, the transformer may be able to operate safely at more than full capacity 712 (i.e., 100%) for different time periods. For example, based on the predicted load parameter value 706, the transformer may have an overload capacity 710A of 135% for 30 minutes (e.g., until time t2), an overload capacity 710B of 120% for one hour (e.g., until time t3), or an overload capacity 710C of 105% for two hours (e.g., until time t3). In this way, overload capacity can be determined as a percentage indicating an additional load that can be borne by a particular electrical appliance for a specific time period without causing abnormal and / or extensive aging or damage to the electrical appliance.

[0055] Unlike some conventional overload capacity calculations (where overload capacity is based on the load directly measured at the current time, e.g., Figure 7 These and other embodiments allow future overload capacity (e.g., overload capacity 710A-710C) to be calculated based on highly accurate predicted load (e.g., predicted load parameter value 706 for future time t1) before the actual load for future time is known. This, in turn, provides operators and asset managers with more time to plan for and respond to contingencies.

[0056] Operation 600 further includes: changing at least one parameter associated with the electrical equipment at the current time based on overload capacity calculated for a future time (box 610). Operation 600 may also include: operating the electrical equipment at least in part based on said at least one changed parameter (box 612). For example, as discussed above, the cooling components of a transformer may be operated in response to a predicted load parameter value to change the temperature of at least one component of the electrical equipment before a future time. Changing parameters and / or operating the electrical equipment may also include communicating with the electrical equipment and / or providing instructions to the electrical equipment.

[0057] In some examples, the at least one machine learning model is trained based on multiple defined relationships between a predefined number of load parameter values ​​and at least one subsequent load parameter value from a time-series data stream obtained over a predetermined time period. In this example, training is performed before deploying the machine learning model and may also occur continuously during operation.

[0058] In some examples, the multiple determined relationships are verified based on a comparison of at least one expected load parameter value derived from a predefined number of load parameter values ​​with the at least one subsequent load parameter value. For example, as described above regarding... Figure 2A-2B The discussed method utilizes a predefined number of load parameter values ​​as input to the at least one machine learning model to sequentially predict load parameter values ​​for future times. This input set can be sequentially generated from a time-series data stream of load parameter values ​​using a moving window technique. For example, the plurality of load parameter values ​​may include a set of at least five load parameter values ​​iteratively extracted from a stream of load parameter values ​​obtained from electrical equipment. The future time for the predicted load parameter values ​​may also be at least one hour or more after the prediction.

[0059] In some examples, the at least one machine learning model predicts the load parameter values ​​of the electrical equipment for a future time based solely on the plurality of load parameter values. For example, as discussed above, the machine learning model can accurately predict future loads using a relatively small set of historical load data without any additional input. In some examples, the machine learning model may also take into account known variations of the electrical equipment, such as anticipating overload conditions to perform cooling operations. For example, in some embodiments, the at least one machine learning model predicts the load parameter values ​​of the electrical equipment for a future time based on the plurality of load parameter values ​​and at least one temperature parameter associated with the electrical equipment.

[0060] The at least one machine learning model is trained accordingly with a data stream of load parameter values, which may include actually measured load values, temperature values, etc., for prediction. The at least one machine learning model may be included within a device (e.g., a controller or relay associated with electrical equipment) that receives data stream values ​​from sensors connected to the electrical equipment and is capable of controlling or coordinating other devices to perform at least one operation associated with the electrical equipment or included in a central monitoring system deployed in a substation or distribution management system for managing a power system. For example, based on predictions of future load parameter values ​​and based on processing implemented to determine overload conditions at future times, the device (e.g., controller / relay) may anticipate future overload conditions and perform / coordinate cooling operations.

[0061] Figure 8 This is a block diagram of a transformer load forecasting system 800, which is configured to perform the operations disclosed herein, such as, for example... Figure 3 Operation 300 and / or Figure 6 Operation 600. In Figure 8 In some embodiments, the transformer monitoring system 30 of the load forecasting system 800 can monitor one or more transformers 10A, 10B. In some embodiments, the transformer monitoring system 30 is integrated within transformer 10A, which is provided as a means for monitoring and load forecasting and can be enabled to monitor only transformer 10A. In other embodiments, the transformer monitoring system 30 may be integrated with transformer 10A to monitor transformer 10A and optionally also monitor or receive data from one or more adjacent electrical devices (e.g., transformer 10B or another power or current transformer or circuit breaker) or connected transmission / distribution lines. In yet another embodiment, the transformer monitoring system 30 is separate from the monitored transformers 10A, 10B.

[0062] The transformer monitoring system 30 includes a processor circuit 34, a communication interface 32 connected to the processor circuit, and a memory 36 connected to the processor circuit 34. The memory 36 includes machine-readable computer program instructions that, when executed by the processor circuit 34, cause the processor circuit 34 to perform the operations described and illustrated herein (such as, as above, for example...). Figure 6 Some of the operations described in operation 600).

[0063] As shown in the figure, the load forecasting system 800 includes a communication interface 32 (also referred to as a network interface) configured to provide communication with other devices, such as communication with sensors 20 in transformers 10A and 10B via wired or wireless communication channel 14. The transformer monitoring system 30 can receive signals from the sensors 20 indicating physical parameters of transformers 10A and 10B (e.g., voltage, current, oil temperature, ambient temperature, etc. associated with transformers 10A and 10B). One advantage of some embodiments is that the transformer monitoring system 30 can be a resource-constrained device because the machine learning algorithms require less data and / or parameters to perform accurate load forecasting and prediction.

[0064] In this example, the transformer monitoring system 30 is depicted as a separate monitoring device communicating with the circuits of transformers 10A and 10B via communication channel 14, for example in a server-client model, a cloud-based platform, a substation automation system used in a substation, a distribution management system for power system management, or other network arrangements. One advantage of a client-server configuration is that overall optimization of the load in the power system and / or substation can be achieved based on load forecasts and overload capacity calculations for multiple individual devices (such as transformers 10A and 10B). For example, load management in a power system may include redistributing loads across different electrical devices based on predicted load capacities. However, it should also be understood that in other embodiments, the transformer monitoring system 30 may be part of transformers 10A, 10B, or other electrical devices as needed, or it may operate in a client-server model where the client associates with and measures the electrical device, and the server runs machine learning algorithms and calculates overload capacity.

[0065] In another embodiment of the server-client model, the transformer monitoring system may have a device (e.g., a client) associated with the monitored transformer, wherein the device includes a machine learning model for load forecasting, and a central system (e.g., a server) is configured to monitor multiple electrical devices / transformers. The server may also include instances of the machine learning model included in the device associated with the transformer. The machine learning model in the server can be continuously trained for load forecasting using data received from the transformer and / or multiple electrical devices, wherein the server provides information / data (e.g., model coefficients) to tune / adapt the machine learning model in the device, which is derived from the continuously learned machine learning model in the server. The server may also be able to perform simulations or advanced processing to forecast / simulate conditions in the transformer (e.g., hotspot determination based on electrical / thermal load information provided by devices or sensors connected to the transformer) and provide information related to such determinations to the device connected to the transformer (e.g., the client) so that the device can change at least one parameter (e.g., cooling) associated with the transformer (or other electrical device). According to various embodiments, the transformer monitoring system 30 may include electronic, computing, and communication hardware and software for predicting load parameter values ​​and performing at least one activity associated with the transformer.

[0066] The transformer monitoring system 30 also includes a processor circuit 34 (also referred to as a processor) and a memory circuit 36 ​​(also referred to as a memory) coupled to the processor circuit 34. According to other embodiments, the processor circuit 34 may be defined to include the memory, such that a separate memory circuit is not required.

[0067] As discussed herein, other aspects of the operation of the transformer monitoring system 30 and the load forecasting system 800 may be performed by the processor circuitry 34 and / or the communication interface 32. For example, the processor circuitry 34 may control the communication interface 32 to transmit communications to one or more other devices and / or receive communications from one or more other devices via a network interface. Furthermore, modules may be stored in memory 36, and these modules may provide instructions such that when the instructions of a module are executed by the processor circuitry 34, the processor circuitry 34 performs corresponding operations (e.g., the operations discussed herein with respect to exemplary embodiments). For example, modules may be further configured to manage fault detection as needed, generate updated probabilities for different nodes, provide interfaces (e.g., application programming interfaces (APIs)) for management, configuration, and / or modification of fault tree structures by customers or other users, etc.

[0068] Transformers 10A and 10B (which may be, for example, high-voltage transformers) include sensors 20 that measure various quantities associated with transformers 10A and 10B (such as operating load, ambient temperature, moisture and / or oxygen content) and transmit these measurements to transformer monitoring system 30 via communication channel 14. Transformers 10A and 10B may also include subsystems such as active sections 22 connected to power line 26 (e.g., overhead power transmission line), cooling systems 24 (e.g., for transformers or reactors), etc., which may be operated by, for example, processor circuitry 34 or in response to instructions from that processor circuitry. In some examples, similar monitoring systems may be associated with power line 26 or other components of load forecasting system 800. In this and other examples, embodiments are described in the context of transformers for simplicity, but it should be understood that many other types of electrical equipment subjected to load conditions can benefit from the embodiments described herein, such as reactors, transmission lines, instrument transformers, generators, etc., and all such electrical equipment should also be contemplated within the scope of this disclosure.

[0069] The transformer monitoring system 30 can use these measured quantities to detect and / or determine the presence of faults in various components or subsystems of transformers 10A and 10B and / or the general fault condition of transformers 10A and 10B. The communication channel 14 may include a wired or wireless link, and in some embodiments may include a wireless local area network (WLAN) or a cellular communication network, such as a 4G or 5G communication network.

[0070] The load forecasting system 800 can receive online or offline measurements of operating load, temperature, moisture, oxygen content, etc., from transformers 10A and 10B and process these measurements to detect and / or determine the presence of faults. The load forecasting system 800 can be implemented in a server, server cluster, cloud-based remote server system, and / or standalone unit. Sensor data can be obtained by the load forecasting system 800 from one transformer and / or multiple transformers.

[0071] The load forecasting system 700 described herein can be implemented in many different ways. For example, the transformer monitoring system 30 according to some embodiments may receive online / offline data, and machine learning techniques configured in the device may be used to learn and classify the received data to identify different patterns that can be considered for estimation / simulation as described in the various embodiments. The device may be able to connect to one or more transformers 10 to receive measurement data.

[0072] In the above description of various embodiments of the inventive concept, it will be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the inventive concept. Unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the inventive concept pertains. It will be further understood that terms (such as those defined in common dictionaries) should be interpreted as having the same meaning as they have in the context of this specification and related art.

[0073] When an element is referred to as “connected,” “linked,” “responding,” or a variation thereof to another element, it may be directly connected, linked, or responding to another element, or an intervening element may be present. In contrast, when an element is referred to as “directly connected,” “directly linked,” “directly responding,” or a variation thereof to another element, no intervening element is present. Similar numbers refer to similar elements throughout. Furthermore, “linked,” “connected,” “responding,” or variations thereof as used herein may include wireless links, links, or responses. As used herein, the singular forms “a,” “an,” and “the” are intended to also include the plural forms unless the context explicitly indicates otherwise. For brevity and / or clarity, well-known functions or constructions may not be described in detail. The term “and / or” includes any and all combinations of one or more of the associated listed items. The phrase “at least one of A and B” means “A or B” or “A and B.”

[0074] It will be understood that although the terms first, second, third, etc., may be used herein to describe various elements / operations, these elements / operations should not be limited by these terms. These terms are used only to distinguish one element / operation from another. Therefore, without departing from the teachings of the inventive concept, a first element / operation in some embodiments may be referred to as a second element / operation in other embodiments. The same reference numerals or the same reference indicators throughout the specification denote the same or similar elements.

[0075] As used herein, the terms “comprise,” “comprising,” “include,” “have,” “has,” or variations thereof are open-ended and include one or more of the stated features, integers, elements, steps, components, or functions, but do not exclude the presence or addition of one or more other features, integers, elements, steps, components, functions, or groups thereof.

[0076] Example embodiments are described herein with reference to block diagrams and / or flowcharts of computer-implemented methods, apparatus (systems and / or devices), and / or computer program products. It should be understood that blocks in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented by computer program instructions executed by one or more computer circuits. These computer program instructions can be provided to processor circuitry of general-purpose computer circuitry, processor circuitry of special-purpose computer circuitry, and / or other programmable data processing circuitry to produce a machine, such that instructions executed via a processor of a computer and / or other programmable data processing apparatus transform and control transistors, values ​​stored in memory locations, and other hardware components within such circuitry systems to implement the functions / actions specified in the block diagrams and / or one or more flowchart blocks, thereby creating means (functions) and / or structures for implementing the functions / actions specified in the block diagrams and / or (multiple) flowchart blocks.

[0077] These computer program instructions may also be stored in a tangible computer-readable medium that can instruct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable medium produce an article of writing comprising instructions that implement the functions / actions specified in block diagrams and / or one or more flowchart blocks. Therefore, embodiments of the inventive concept may be embodied in hardware and / or software (including firmware, resident software, microcode, etc.) running on a processor (such as a digital signal processor), which may be collectively referred to as a “circuit system,” a “module,” or a variation thereof.

[0078] It should also be noted that in some alternative embodiments, the functions / actions indicated in the boxes may not occur in the order shown in the flowchart. For example, two boxes shown consecutively may actually be performed substantially simultaneously, or these boxes may sometimes be performed in reverse order, depending on the functions / actions involved. Furthermore, the function of a given box in a flowchart and / or block diagram may be divided into multiple boxes, and / or the functions of two or more boxes in a flowchart and / or block diagram may be at least partially integrated. Finally, other boxes may be added / inserted between the illustrated boxes without departing from the scope of the inventive concept, and / or boxes / actions may be omitted. Additionally, although some figures include arrows on communication paths to indicate the primary communication direction, it will be understood that communication may occur in the opposite direction to the depicted arrows.

[0079] Many variations and modifications may be made to the embodiments without substantially departing from the principles of the inventive concept. All such variations and modifications are intended to be included herein within the scope of the inventive concept. Therefore, the subject matter disclosed above is to be considered illustrative rather than limiting, and the examples of embodiments are intended to cover all such modifications, enhancements, and other embodiments falling within the spirit and scope of the inventive concept. Thus, to the fullest extent permitted by law, the scope of the inventive concept will be determined by the broadest permissible interpretation of this disclosure (including examples of embodiments and their equivalents) and should not be construed as limited or restricted by the foregoing detailed description.

Claims

1. A method for using a machine learning model for prediction, comprising: The processor circuitry predicts load parameter values ​​for future times for an electrical device based on at least one machine learning model and multiple load parameter values, said multiple load parameter values ​​comprising a predefined number of load parameter values ​​extracted from a time-series data stream of load parameter values ​​obtained for the electrical device. Specifically, the at least one machine learning model is trained and selected based on historical load data using the following operations, the operations including: Step S1: Decompose the historical load data dataset into training data and validation data. Step S2: Decompose the training data into K random segments. Step S3: Train a machine learning model using K-1 fragments of the training data. Step S4: Validate the machine learning model using the remaining fragments of the training data. Step S5: Annotate the accuracy of the machine learning model. Step S6: Repeat steps S3 to S5 for multiple different machine learning models. Step S7: Repeat steps S3 to S5 for each group of K-1 fragments, performing a total of K training and testing operations for each machine learning model. Step S8: Repeat steps S2 to S7 for a different set of K random fragments of the training data. Step S9: Use the validation data to validate the trained machine learning model, and Step S10: Select at least one of the most accurate machine learning models for prediction; The processor circuitry calculates the overload capacity for the future time based on the predicted load parameter values; and The processor circuitry alters at least one parameter associated with the electrical device at the current time based on its overload capacity calculated for the future time. The plurality of load parameter values ​​constitute a univariate dataset obtained from the electrical equipment, and the at least one machine learning model is configured to predict load parameter values ​​for the future time based solely on the univariate dataset without using sensor data from outside the electrical equipment as input.

2. The method according to claim 1, wherein, The at least one machine learning model is trained based on multiple defined relationships between a predefined number of load parameter values ​​and at least one subsequent load parameter value from a time-series data stream obtained within a predetermined time period.

3. The method according to claim 2, wherein, The multiple defined relationships are verified based on a comparison between at least one expected load parameter value derived from the predefined number of load parameter values ​​and the at least one subsequent load parameter value.

4. The method according to any one of claims 1 to 3, wherein, Using the predefined number of load parameter values ​​as the input set of the at least one machine learning model, the load parameter values ​​for the future time are predicted sequentially, and The input set is generated sequentially from the time-series data stream of load parameter values ​​using a moving window technique.

5. The method according to any one of claims 1 to 3, wherein, The plurality of load parameter values ​​include a set of at least five load parameter values ​​extracted iteratively from the load parameter value stream obtained from the electrical equipment.

6. The method according to any one of claims 1 to 3, wherein, The predicted future time for the load parameter values ​​is at least one hour after the prediction.

7. The method according to any one of claims 1 to 3, further comprising: The processor circuitry predicts at least one hot spot temperature value for a component of the electrical equipment for the future time based on the predicted load parameter values.

8. The method according to any one of claims 1 to 3, further comprising: The processor circuitry predicts at least one overload capacity value of the electrical equipment based on predicted load parameter values, the at least one overload capacity value being associated with at least one time period following the future time.

9. The method according to any one of claims 1 to 3, wherein, The electrical equipment includes a transformer, and the method further includes: The transformer is operated at least in part based on the at least one modified parameter.

10. The method according to claim 9, wherein, Operating the transformer includes operating at least one cooling component of the transformer in response to a predicted load parameter value to change the temperature of at least one component of the electrical equipment before the future time.

11. A monitoring device that uses a machine learning model for prediction, comprising: Processor circuitry; as well as The memory includes machine-readable instructions that, when executed by the processor circuitry, cause the processor circuitry to: The system predicts load parameter values ​​for future times for electrical equipment based on at least one machine learning model and multiple load parameter values, wherein the multiple load parameter values ​​include a predefined number of load parameter values ​​extracted from a time-series data stream of load parameter values ​​obtained for the electrical equipment. Specifically, the at least one machine learning model is trained and selected based on historical load data using the following operations, the operations including: Step S1: Decompose the historical load data dataset into training data and validation data. Step S2: Decompose the training data into K random segments. Step S3: Train a machine learning model using K-1 fragments of the training data. Step S4: Validate the machine learning model using the remaining fragments of the training data. Step S5: Annotate the accuracy of the machine learning model. Step S6: Repeat steps S3 to S5 for multiple different machine learning models. Step S7: Repeat steps S3 to S5 for each group of K-1 fragments, performing a total of K training and testing operations for each machine learning model. Step S8: Repeat steps S2 to S7 for a different set of K random fragments of the training data. Step S9: Use the validation data to validate the trained machine learning model, and Step S10: Select at least one of the most accurate machine learning models for prediction; Calculate the overload capacity for the future time based on the predicted load parameter values; and Based on the overload capacity calculated for the future time, at least one parameter associated with the electrical equipment at the current time is changed. The plurality of load parameter values ​​constitute a univariate dataset obtained from the electrical equipment, and the at least one machine learning model is configured to predict load parameter values ​​for the future time based solely on the univariate dataset without using sensor data from outside the electrical equipment as input.

12. The monitoring device according to claim 11, wherein, The at least one machine learning model is trained based on multiple defined relationships between a predefined number of load parameter values ​​and at least one subsequent load parameter value from a time-series data stream obtained within a predetermined time period.

13. The monitoring device according to claim 11, wherein, The plurality of load parameter values ​​include a set of at least five load parameter values ​​extracted iteratively from the load parameter value stream obtained from the electrical equipment.

14. The monitoring device according to any one of claims 11 to 13, wherein, The predicted future time for the load parameter values ​​is at least one hour after the prediction.

15. A non-transitory computer-readable medium using a machine learning model for prediction, comprising instructions that, when executed by processor circuitry, cause the processor circuitry to: The system predicts load parameter values ​​for future times for electrical equipment based on at least one machine learning model and multiple load parameter values, wherein the multiple load parameter values ​​include a predefined number of load parameter values ​​extracted from a time-series data stream of load parameter values ​​obtained for the electrical equipment. in, The at least one machine learning model is trained and selected based on historical load data using the following operations, the operations including: Step S1: Decompose the historical load data dataset into training data and validation data. Step S2: Decompose the training data into K random segments. Step S3: Train a machine learning model using K-1 fragments of the training data. Step S4: Validate the machine learning model using the remaining fragments of the training data. Step S5: Annotate the accuracy of the machine learning model. Step S6: Repeat steps S3 to S5 for multiple different machine learning models. Step S7: Repeat steps S3 to S5 for each group of K-1 fragments, performing a total of K training and testing operations for each machine learning model. Step S8: Repeat steps S2 to S7 for a different set of K random fragments of the training data. Step S9: Use the validation data to validate the trained machine learning model, and Step S10: Select at least one of the most accurate machine learning models for prediction; Calculate the overload capacity for the future time based on the predicted load parameter values; and Based on the overload capacity calculated for the future time, at least one parameter associated with the electrical equipment at the current time is changed. The plurality of load parameter values ​​constitute a univariate dataset obtained from the electrical equipment, and the at least one machine learning model is configured to predict load parameter values ​​for the future time based solely on the univariate dataset without using sensor data from outside the electrical equipment as input.

16. The computer-readable medium of claim 15, wherein, The at least one machine learning model is trained based on multiple defined relationships between a predefined number of load parameter values ​​and at least one subsequent load parameter value from a time-series data stream obtained within a predetermined time period.

17. The computer-readable medium of claim 15, wherein, The plurality of load parameter values ​​include a set of at least five load parameter values ​​extracted iteratively from the load parameter value stream obtained from the electrical equipment.

18. The computer-readable medium according to any one of claims 15 to 17, wherein, The predicted future time for the load parameter values ​​is at least one hour after the prediction.

Images