A method and system for load event detection based on conditional autoencoder

Abstract

The application discloses a load event detection method and system based on a conditional autoencoder, and comprises the following steps: acquiring active power sequences of each power consumption device in a power grid; performing load event detection on each active power sequence according to a preset load event detection model to obtain a load event detection result; wherein the load event detection model is obtained by training a classification network of a preset algorithm according to event labels and hidden layer features of time points of each load event in historical power data; the historical power data are historical power data of each power consumption device in the power grid; and the hidden layer features are obtained by performing feature extraction on the historical power data according to a conditional autoencoder network of a long short-term memory network structure. The method and system provided by the application can more accurately detect load events.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power load detection, and particularly relates to a load event detection method and system based on a conditional autoencoder. BACKGROUND

[0002] In a power system, monitoring and controlling loads is very important. Load switching event detection is a key task in load monitoring. At present, load switching event detection mainly adopts a threshold judgment-based method and a machine learning method. The threshold judgment-based method usually sets a threshold according to the change amplitude of current and voltage signals to determine whether a load switching event occurs. The machine learning method trains a large amount of power system data using a supervised learning algorithm to establish a load switching event classification model.

[0003] In the prior art of load event detection using a machine learning method, since some machine learning methods extract redundant load event features, the redundant features are not conducive to load event detection learning, and thus there are false positives and false negatives in load event detection. Therefore, it is urgent to propose a method that can more accurately detect load events. SUMMARY

[0004] The present application provides a load event detection method and system based on a conditional autoencoder, which realizes more accurate detection of load events.

[0005] To solve the above technical problems, the present application provides a load event detection method based on a conditional autoencoder, comprising:

[0006] obtaining active power sequences of each electrical equipment in a power grid;

[0007] performing load event detection on each active power sequence according to a preset load event detection model to obtain a load event detection result;

[0008] The load event detection model is obtained by training a classification network of a preset algorithm according to event labels and hidden layer features of time points of each load event in historical power data; the historical power data are historical power data of each electrical equipment in the power grid; and the hidden layer features are obtained by performing feature extraction on the historical power data according to a conditional autoencoder network of a long short-term memory network structure.

[0009] It can be understood that, compared with the prior art, the method provided by the application can use a lightweight conditional autoencoder to detect whether a load event occurs in a power sequence. By marking the time point at which the load event occurs, the historical power data is preprocessed to facilitate subsequent analysis and modeling for load event detection. By using a conditional autoencoder network with a long short-term memory network structure, the historical power data is mapped to a hidden layer feature space, and then the historical power data is reconstructed according to the hidden layer feature space. The network not only can retain the time sequence characteristics of the historical power data, but also can extract the hidden layer features of the historical power data, and thus can more accurately complete the work of load event detection. By using event labels to back-propagate the training classification network, whether a load event occurs can be predicted according to the labels of the actual occurrence of the load event. More accurate load event detection is realized, which can effectively avoid false detection and missed detection of load events and assist in carrying out load identification work.

[0010] Further, the hidden layer feature is obtained by using a conditional autoencoder network with a long short-term memory network structure to extract features from the historical power data, and specifically includes:

[0011] Deleting or replacing abnormal values in the historical power data;

[0012] Aligning the time stamps in the historical power data of the respective electrical equipment, and normalizing the historical power data using a preset normalization method;

[0013] A conditional autoencoder network with a long short-term memory network structure is constructed; wherein the conditional autoencoder network includes an encoder network and a decoder network; the encoder network uses two long short-term memory network blocks to extract first features of load events in the historical power data; the decoder network uses two long short-term memory network blocks to reconstruct the historical power data according to the first features, to obtain a reconstructed sequence.

[0014] The conditional autoencoder network is trained using a preset optimizer, and the training is stopped when the average absolute error loss of the historical power data and the reconstructed sequence is less than or equal to a first preset value, and the first features corresponding to the stopping of the training of the encoder network are taken as the hidden layer features.

[0015] It can be understood that the method provided by the application processes the abnormal values in the historical power data, so as to prevent the abnormal data from affecting the accuracy of the model. Meanwhile, the load event detection needs to process the power data of different power equipment, and the data has different time stamps, power ranges and units, so it is necessary to adopt the alignment and normalization manner to pre-process the data, so as to facilitate subsequent analysis and modeling. By using the conditional auto-encoder network with the long short-term memory network structure, the historical power data is mapped to the hidden layer feature space, and then the historical power data is reconstructed according to the hidden layer feature space. The network not only can retain the time sequence characteristics of the historical power data, but also can extract the hidden layer features of the historical power data, so that the work of load event detection can be more accurately completed.

[0016] Further, the encoder network uses two long short-term memory network blocks to extract the first features of the load events in the historical power data, and the specific formula is:

[0017] F = φ (W e x) ; wherein, F is the first feature of the load event; W e is an encoder network parameter; φ is a preset activation function; and x is the historical power data.

[0018] It can be understood that the method provided by the application realizes the coding of the high-dimensional input historical power data into low-dimensional features by using the two long short-term memory network blocks of the encoder, so that the load event detection model can learn the features with the largest amount of information.

[0019] Further, the decoder network uses two long short-term memory network blocks to reconstruct the historical power data according to the first features, so as to obtain a reconstructed sequence, and the specific formula is:

[0020] wherein, is the reconstructed sequence; F is the first feature of the load event; W d is a decoder network parameter; φ is a preset activation function; and x is the historical power data.

[0021] It can be understood that the method provided by the application restores the low-dimensional features to the initial high-dimensional input historical power data by using the decoder with the two long short-term memory network blocks, and if the reconstructed sequence is consistent with the initial sequence, the extracted low-dimensional features achieve the effect of de-noising.

[0022] Further, the load event detection model is obtained by training a classification network through a preset algorithm according to the event labels and the hidden layer features of the time points of each load event in the historical power data, and specifically includes:

[0023] determine occurrence of the load event in the historical power data according to whether there is an event label at the current time point;

[0024] If there is the event label at the current time point, the hidden layer features are taken as inputs and transmitted into a classification network composed of two full connection layers for training, and the training is stopped when a loss function of the classification network is less than or equal to a second preset value, and a trained classification network is obtained.

[0025] It can be understood that the method provided by the application can predict whether a load event occurs according to a label of a real occurrence of the load event by backpropagation training of a classification network, and can provide a basis for subsequent load identification work.

[0026] Correspondingly, the application also provides a load event detection system based on a conditional autoencoder, which comprises:

[0027] a data acquisition module configured to acquire active power sequences of each electrical equipment in a power grid;

[0028] an event detection module configured to perform load event detection on each active power sequence according to a preset load event detection model to obtain a load event detection result;

[0029] The load event detection model is trained by a classification network of a preset algorithm according to event labels and hidden layer features of time points of each load event in historical power data, the historical power data are historical power data of each electrical equipment in the power grid, and the hidden layer features are obtained by feature extraction on the historical power data by a conditional autoencoder network with a long short-term memory network structure.

[0030] It can be understood that, compared with the prior art, the system provided by the application performs preprocessing on historical power data by marking time points of load events, which is beneficial to load event detection and facilitates subsequent analysis and modeling. The conditional autoencoder network with a long short-term memory network structure can extract hidden layer feature information and time sequence information in the historical power data, and establish a load event detection model with stronger generalization. Meanwhile, the power sequence of the power user is monitored in real time by using the conditional autoencoder with a long short-term memory network structure and the classification network, the accuracy of load event detection is improved, and a basis is provided for subsequent load identification work.

[0031] Further, the event detection module specifically comprises:

[0032] a data preprocessing submodule configured to delete or replace outliers in the historical power data, align time stamps in the historical power data of each electrical equipment, and normalize the historical power data by using a preset normalization method.

[0033] The conditional encoder construction submodule is configured to construct a conditional autoencoder network using a long short-term memory network structure; wherein the conditional autoencoder network comprises an encoder network and a decoder network; the encoder network is configured to extract first features of load events in the historical power data using two long short-term memory network blocks; and the decoder network is configured to reconstruct the historical power data using two long short-term memory network blocks according to the first features to obtain a reconstructed sequence.

[0034] The hidden layer feature extraction submodule is configured to train the conditional autoencoder network using a preset optimizer, and stop training when a mean absolute error loss of the historical power data and the reconstructed sequence is less than or equal to a first preset value, and take the first features corresponding to the stop of the training of the encoder network as hidden layer features.

[0035] It can be understood that the system provided by the present application processes abnormal values in historical power data to prevent abnormal data from affecting the accuracy of the model. At the same time, load event detection needs to process power data of different electrical equipment, and the data has different timestamps, power ranges and units, so it is necessary to use alignment and normalization to preprocess the data for subsequent analysis and modeling. Through the conditional autoencoder network with a long short-term memory network structure, the hidden layer feature information and time sequence information in the historical power data can be extracted, and a load event detection model with stronger generalization is established.

[0036] Further, the conditional encoder construction submodule further comprises:

[0037] The encoder network submodule is configured to extract first features of load events in the historical power data using two long short-term memory network blocks, and the specific formula is: F = φ(W e x); wherein F is the first feature of the load event; W e is an encoder network parameter; φ is a preset activation function; and x is the historical power data.

[0038] It can be understood that the system provided by the present application realizes the encoding of high-dimensional input historical power data into low-dimensional features through the encoder of two long short-term memory network blocks, so that the load event detection model can learn the features with the largest amount of information.

[0039] Further, the conditional encoder construction submodule further comprises:

[0040] The decoder network submodule is configured to reconstruct the historical power data using two long short-term memory network blocks according to the first features to obtain a reconstructed sequence, and the specific formula is: wherein, is the reconstructed sequence; F is the first feature of the load event; W d is the decoder network parameter; φ is a preset activation function; x is the historical power data.

[0041] It can be understood that the system provided by the present application restores the low-dimensional features to the initial high-dimensional input historical power data through the decoder of the two long short-term memory network blocks, and if the reconstructed sequence is consistent with the initial sequence, the extracted low-dimensional features achieve the effect of denoising.

[0042] Further, the event detection module further comprises:

[0043] An event label determination submodule is configured to determine the occurrence of the load event in the historical power data according to whether the event label exists at the current time point.

[0044] A classification network training submodule is configured to, if the event label exists at the current time point, input the hidden layer feature as an input into a classification network composed of two full connection layers for training, stop the training when a loss function of the classification network is less than or equal to a second preset value, and obtain the trained classification network.

[0045] It can be understood that the system provided by the present application trains the classification network through the event label back propagation, can predict whether the load event occurs according to the label of the real occurrence of the load event, and can provide a basis for subsequent load identification work. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 is a step flow chart of a load event detection method based on a conditional autoencoder provided by an embodiment of the present application;

[0047] Figure 2 is a step flow chart of constructing a load event detection model in a load event detection method based on a conditional autoencoder provided by an embodiment of the present application;

[0048] Figure 3 is a load event detection model structure diagram provided by an embodiment of the present application;

[0049] Figure 4 is a structure diagram of a load event detection system based on a conditional autoencoder provided by an embodiment of the present application;

[0050] Figure 5 is an internal structure diagram of an event detection module in a load event detection system based on a conditional autoencoder provided by an embodiment of the present application;

[0051] Figure 6: This is a schematic diagram of the internal structure of the condition encoder construction submodule in a load event detection system based on a condition autoencoder provided in an embodiment of the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Example 1

[0054] Please refer to Figure 1 The flowchart below shows a load event detection method based on a conditional autoencoder provided in an embodiment of the present invention, which includes the following steps S101-S102, each of which is detailed below.

[0055] S101: Obtain the active power sequence of each electrical device in the power grid.

[0056] S102: Based on the preset load event detection model, load event detection is performed on each of the active power sequences to obtain load event detection results.

[0057] In this embodiment, the load event detection model is trained using a classification network with a preset algorithm based on the event labels and hidden features of each load event at the time point in the historical power data; the historical power data refers to the historical power data of each electrical device in the power grid; the hidden features are obtained by extracting features from the historical power data using a conditional autoencoder network with a long short-term memory network structure.

[0058] Please refer to Figure 2 The flowchart below shows the steps for constructing a load event detection model in a load event detection method based on a conditional autoencoder provided in this embodiment of the invention, including the following steps S201-S206, each step of which is as follows:

[0059] S201: Delete or replace outliers in historical power data. Align the timestamps in the historical power data of each electrical device and normalize the historical power data using a preset normalization method.

[0060] Understandably, the method provided by this invention processes outliers in historical power data to prevent them from affecting the accuracy of the model. Meanwhile, load event detection requires processing power data from different electrical devices, which vary in timestamps, power ranges, and units. Therefore, alignment and normalization are necessary for data preprocessing to facilitate subsequent analysis and modeling.

[0061] S202: Use two long short-term memory network blocks to extract the first feature of load events from the historical power data.

[0062] In this embodiment, the encoder network uses two long short-term memory network blocks to extract the first feature of load events from the historical power data, specifically using the formula: F = φ(W e ·x); where F is the first characteristic of the load event; W e φ represents the encoder network parameters; φ represents the preset activation function; and x represents the historical power data.

[0063] It is understood that the method provided by the present invention encodes high-dimensional input historical power data into low-dimensional features through encoders of two long short-term memory network blocks, thereby enabling the load event detection model to learn the features with the most information content.

[0064] S203: Based on the first feature, the historical power data is reconstructed using two long short-term memory network blocks to obtain a reconstructed sequence.

[0065] In this embodiment, the decoder network, based on the first feature, uses two long short-term memory network blocks to reconstruct the historical power data, obtaining a reconstructed sequence, specifically as follows:

[0066] in, The reconstructed sequence is defined as F; the first characteristic of the load event is defined as W. d φ represents the decoder network parameters; φ represents the preset activation function; and x represents the historical power data.

[0067] It is understood that the method provided by the present invention uses the decoder of two long short-term memory network blocks to restore the low-dimensional features to the historical power data of the initial high-dimensional input. If the reconstructed sequence is consistent with the initial sequence, then the extracted low-dimensional features achieve the effect of denoising.

[0068] S204: Train the conditional autoencoder network using a preset optimizer. Stop training when the average absolute error loss of the historical power data and the reconstructed sequence is less than or equal to a first preset value. Use the first feature corresponding to when the encoder network stops training as the hidden layer feature.

[0069] As a preferred solution, the ADAM optimizer is used to train the load event detection model to reduce the mean absolute error loss, and the specific formula of the mean error loss is: Wherein, L e is the error loss; x is the historical power data; x is the reconstructed sequence.

[0070] It can be understood that the method provided by the application maps the historical power data to the hidden layer feature space through the conditional auto-encoder network with the long short-term memory network structure, and then reconstructs the historical power data according to the hidden layer feature space. The network not only can retain the time sequence characteristics of the historical power data, but also can extract the hidden layer features of the historical power data, so that the work of load event detection can be more accurately completed.

[0071] S205: Determine the occurrence of the load event in the historical power data according to whether there is an event label at the current time point.

[0072] S206: If there is the event label at the current time point, input the hidden layer feature as an input into the classification network composed of two full connection layers for training, stop training when the loss function of the classification network is less than or equal to a second preset value, and obtain the trained classification network.

[0073] In this embodiment, the hidden layer feature obtained by inputting the historical power data into the encoder network is classified, and then it is judged whether there is a load event in the input historical power data. The loss function in the training process is: Wherein, y is a true label, the value of y is 0 or 1, 0 represents that there is no load event in the input historical power data, and 1 represents that there is a load event in the input historical power data. The predicted result is represented as y The value of y

[0074] It can be understood that the method provided by the application can predict whether the load event occurs according to the label of the real occurrence of the load event by backward propagation training of the classification network, and can provide a basis for subsequent load identification work.

[0075] Specifically, please refer to Figure 3 , the load event detection model structure diagram provided by the embodiment of the application. The original power sequence is input into the conditional auto-encoder network composed of an encoder and a decoder network respectively by two long short-term memory network blocks, to obtain the encoding feature in the original power sequence. The encoding feature is input into a preset classifier to determine whether there is a load event in the original power sequence.

[0076] As a preferred solution, the application takes practical application scenarios as examples, and realizes load event detection by using the method provided by the embodiments of the application.

[0077] As a preferred solution, taking the BLUED public load data set as an example, all data is divided into training and test data sets according to a ratio of 8:2. The conditional autoencoder network parameters are initialized, and the BLUED event data set is used for training. The trained model is tested, the total power sequence of the test data set is input into the trained model, so as to detect the load event in the test machine.

[0078] As a preferred solution, for the evaluation index of the load decomposition result, the true positive percentage TPP, the false positive percentage FPP and the false negative percentage FNP are adopted, and the specific expressions are as follows: Wherein, E is the number of real events; TP is the number of true positive events; FP is the number of false positive events; and FN is the number of false negative events.

[0079] Table 1 load event detection result

[0080]

[0081] As a preferred solution, the result of load event detection by using the method of the application is shown in Table 1. It can be seen that the method provided by the embodiments of the application can accurately detect the occurrence of load events.

[0082] The method provided by the application can use a lightweight conditional autoencoder to detect whether a load event occurs in a power sequence. By marking the time point of the load event, the historical power data is preprocessed to facilitate subsequent analysis and modeling. By using the conditional autoencoder network with the long short-term memory network structure, the historical power data is mapped to the hidden layer feature space, and then the historical power data is reconstructed according to the hidden layer feature space. The network not only can retain the time sequence characteristics of the historical power data, but also can extract the hidden layer features of the historical power data, so as to more accurately complete the work of load event detection. By using the event label to back-propagate the training classification network, whether the load event occurs can be predicted according to the label of the real occurrence of the load event. More accurate load event detection is realized, which can effectively avoid false detection and missed detection of load events, and assist in carrying out load identification work.

[0083] Embodiment two

[0084] Please refer to Figure 4 Fig. 1 is a structural schematic diagram of a load event detection system based on a conditional autoencoder provided by an embodiment of the application, which comprises a data acquisition module 301 and an event detection module 302.

[0085] The data acquisition module 301 is configured to acquire active power sequences of each power utilization device in the power grid.

[0086] The event detection module 302 is configured to perform load event detection on each active power sequence according to a preset load event detection model to obtain a load event detection result.

[0087] In this embodiment, the load event detection model is obtained by training a classification network of a preset algorithm according to event labels and hidden layer features of time points of each load event in historical power data; the historical power data are historical power data of each power utilization device in the power grid; and the hidden layer features are obtained by performing feature extraction on the historical power data according to a long short-term memory network structure condition auto-encoder network.

[0088] In this embodiment, please refer to Figure 5 FIG. 2 is a schematic diagram of an internal structure of an event detection module of a load event detection system based on a condition auto-encoder according to an embodiment of the present application, which specifically includes a data preprocessing submodule 401, a condition auto-encoder construction submodule 402, a hidden layer feature extraction submodule 403, an event label determination submodule 404, and a classification network training submodule 405.

[0089] The data preprocessing submodule 401 is configured to delete or replace abnormal values in the historical power data; align time stamps in the historical power data of each power utilization device; and normalize the historical power data using a preset normalization method.

[0090] It can be understood that the system provided by the present application processes abnormal values in the historical power data to prevent abnormal data from affecting the accuracy of the model. Meanwhile, load event detection needs to process power data of different power utilization devices, and the data time stamps, power ranges, and units are different, so it is necessary to use the alignment and normalization method to preprocess the data for subsequent analysis and modeling.

[0091] The condition auto-encoder construction submodule 402 is configured to construct a condition auto-encoder network using a long short-term memory network structure; wherein the condition auto-encoder network includes an encoder network and a decoder network; the encoder network uses two long short-term memory network blocks to extract first features of load events in the historical power data; and the decoder network uses two long short-term memory network blocks to reconstruct the historical power data according to the first features to obtain a reconstructed sequence.

[0092] In this embodiment, please refer to Figure 6An internal structure schematic diagram of a conditional encoder construction submodule in a load event detection system based on a conditional autoencoder provided by an embodiment of the present application, the conditional encoder construction submodule 402 comprises an encoder network submodule 501 and a decoder network submodule 502.

[0093] The encoder network submodule 501 is configured to extract first features of the load event in the historical power data by using two long short-term memory network blocks, and a specific formula is F=φ(W e x); wherein F is the first feature of the load event, W e is an encoder network parameter, φ is a preset activation function, and x is the historical power data.

[0094] It can be understood that the system provided by the present application realizes the encoding of high-dimensional input historical power data into low-dimensional features by using the encoder of two long short-term memory network blocks, so that the load event detection model can learn the features with the largest amount of information.

[0095] The decoder network submodule 502 is configured to reconstruct the historical power data by using two long short-term memory network blocks according to the first features, to obtain a reconstructed sequence, and a specific formula is:

[0096] wherein x is the reconstructed sequence, F is the first feature of the load event, W d is a decoder network parameter, φ is a preset activation function, and x is the historical power data.

[0097] It can be understood that the system provided by the present application restores the low-dimensional features to the initial high-dimensional input historical power data by using the decoder of two long short-term memory network blocks, and if the reconstructed sequence is consistent with the initial sequence, the extracted low-dimensional features achieve the effect of de-noising.

[0098] The hidden layer feature extraction submodule 403 is configured to train the conditional autoencoder network by using a preset optimizer, and stop training when the average absolute error loss of the historical power data and the reconstructed sequence is less than or equal to a first preset value, and take the first feature corresponding to the stop of the encoder network training as a hidden layer feature.

[0099] The event label determination submodule 404 is configured to determine the occurrence of the load event in the historical power data according to whether there is an event label at a current time point.

[0100] The classification network training submodule 405 is configured to, if the event label exists at the current time point, input the hidden layer feature as an input into a classification network composed of two full connection layers for training, stop the training when a loss function of the classification network is less than or equal to a second preset value, and obtain a trained classification network.

[0101] It can be understood that the system provided by the present application can predict whether a load event occurs according to a label of a real occurrence of the load event by back propagation training of the classification network, and can provide a basis for subsequent load identification.

[0102] The system provided by the present application can preprocess historical power data by marking time points at which load events occur, which is beneficial to load event detection, and facilitates subsequent analysis and modeling. The conditional autoencoder network with a long short-term memory network structure can extract hidden layer feature information and time sequence information in the historical power data, and establish a load event detection model with stronger generalization. Meanwhile, the power sequence of the power user is monitored in real time by using the conditional autoencoder with a long short-term memory network structure and the classification network, thereby improving the accuracy of load event detection and providing a basis for subsequent load identification.

[0103] The above-described specific embodiments further illustrate the purpose, technical solutions and beneficial effects of the present application. It should be understood that the above-described specific embodiments are merely examples of the present application and are not intended to limit the protection scope of the present application. It is particularly pointed out that any modification, equivalent replacement, improvement, etc. made by those skilled in the art within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A load event detection method based on a conditional autoencoder, characterized in that, include: Obtain the active power sequence of each electrical device in the power grid; Based on the preset load event detection model, load event detection is performed on each of the active power sequences to obtain load event detection results; The load event detection model is trained using a classification network with a preset algorithm based on the event labels and hidden features of each load event at a specific time point in the historical power data. The historical power data refers to the historical power data of each electrical device in the power grid. The hidden features are obtained by extracting features from the historical power data using a conditional autoencoder network with a long short-term memory structure. The hidden layer features are obtained by extracting features from the historical power data using a conditional autoencoder network with a long short-term memory (LSM) network structure. Specifically, this includes: deleting or replacing outliers in the historical power data; aligning the timestamps in the historical power data of each electrical device and normalizing the historical power data using a preset normalization method; constructing a conditional autoencoder network using an LSM network structure; wherein the conditional autoencoder network includes an encoder network and a decoder network; the encoder network uses two LSM network blocks to extract a first feature of load events from the historical power data; the decoder network, based on the first feature, uses two LSM network blocks to reconstruct the historical power data to obtain a reconstructed sequence; training the conditional autoencoder network using a preset optimizer, stopping training when the average absolute error loss between the historical power data and the reconstructed sequence is less than or equal to a first preset value, and using the first feature corresponding to the stoppage of the encoder network as the hidden layer feature.

2. The load event detection method based on conditional autoencoder as described in claim 1, characterized in that, The encoder network uses two long short-term memory network blocks to extract the first feature of load events from the historical power data, specifically using the following formula: ;in, This is the first characteristic of the load event; For encoder network parameters; This is the preset activation function; The historical power data is referred to here.

3. The load event detection method based on conditional autoencoder as described in claim 1, characterized in that, The decoder network, based on the first feature, uses two long short-term memory network blocks to reconstruct the historical power data, obtaining a reconstructed sequence, specifically as follows: ;in, The reconstructed sequence; This is the first characteristic of the load event; These are the decoder network parameters; This is the preset activation function; The historical power data is referred to here.

4. The load event detection method based on conditional autoencoder as described in claim 1, characterized in that, The load event detection model is trained using a pre-defined algorithm classification network based on the event labels and hidden features of each load event at a specific time point in historical power data. Specifically, it includes: Based on whether an event tag exists at the current time point, determine the occurrence of load events in the historical power data; If the event label exists at the current time point, the hidden layer features are used as input and fed into a classification network consisting of two fully connected layers for training. Training stops when the loss function of the classification network is less than or equal to a second preset value, and the trained classification network is obtained.

5. A load event detection system based on a conditional autoencoder, characterized in that, The system includes: The data acquisition module is used to obtain the active power sequence of each electrical device in the power grid; The event detection module is used to perform load event detection on each of the active power sequences according to the preset load event detection model, and obtain the load event detection results. The load event detection model is trained using a classification network with a preset algorithm based on the event labels and hidden features of each load event at a specific time point in the historical power data. The historical power data refers to the historical power data of each electrical device in the power grid. The hidden features are obtained by extracting features from the historical power data using a conditional autoencoder network with a long short-term memory structure. The hidden layer features are obtained by extracting features from the historical power data using a conditional autoencoder network with a long short-term memory (LSM) network structure. Specifically, this includes: deleting or replacing outliers in the historical power data; aligning the timestamps in the historical power data of each electrical device and normalizing the historical power data using a preset normalization method; constructing a conditional autoencoder network using an LSM network structure; wherein the conditional autoencoder network includes an encoder network and a decoder network; the encoder network uses two LSM network blocks to extract a first feature of load events from the historical power data; the decoder network, based on the first feature, uses two LSM network blocks to reconstruct the historical power data to obtain a reconstructed sequence; training the conditional autoencoder network using a preset optimizer, stopping training when the average absolute error loss between the historical power data and the reconstructed sequence is less than or equal to a first preset value, and using the first feature corresponding to the stoppage of the encoder network as the hidden layer feature.

6. The load event detection system based on a conditional autoencoder as described in claim 5, characterized in that, The conditional autoencoder construction submodule further includes: The encoder network submodule is used to extract the first feature of load events from the historical power data using two long short-term memory network blocks, specifically as follows: ;in, This is the first characteristic of the load event; For encoder network parameters; This is the preset activation function; The historical power data is referred to here.

7. The load event detection system based on a conditional autoencoder as described in claim 5, characterized in that, The conditional autoencoder construction submodule also includes: The decoder network submodule is used to reconstruct the historical power data based on the first feature using two long short-term memory network blocks to obtain a reconstructed sequence, specifically using the following formula: ;in, The reconstructed sequence; This is the first characteristic of the load event; These are the decoder network parameters; This is the preset activation function; The historical power data is referred to here.

8. The load event detection system based on a conditional autoencoder as described in claim 5, characterized in that, The event detection module further includes: The event tag determination submodule is used to determine the occurrence of load events in the historical power data based on whether an event tag exists at the current time point; The classification network training submodule is used to input the hidden layer features as input to the classification network consisting of two fully connected layers if the event label exists at the current time point. Training is stopped when the loss function of the classification network is less than or equal to a second preset value, and the trained classification network is obtained.

Images