Virtual power plant distributed data anomaly diagnosis method based on deep neural network

Abstract

The invention provides a virtual power plant distributed data anomaly diagnosis method based on a deep neural network, and relates to the technical field of artificial intelligence, and the method comprises the steps: obtaining the multi-modal data of each distributed energy of a virtual power plant, and obtaining the processed distributed data; modeling each node of the virtual power plant and the relationship thereof into a graph structure to obtain a multi-time-scale spatio-temporal feature vector; pre-training is carried out by constructing a self-supervised learning task, and effective features between different time scales and modal data are extracted; generating a normal data distribution model by using a generative adversarial network, and summarizing abnormal data samples into a comprehensive data set; obtaining a multi-task optimization model through sharing characteristics among tasks; a final optimized anomaly response system is obtained, and then the task of virtual power plant distributed data anomaly diagnosis is completed through the anomaly response system. According to the method, efficient anomaly diagnosis and response processing of the distributed data of the virtual power plant are realized, and the method has remarkable technical advantages and practical application value.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method for diagnosing anomalies in distributed data of a virtual power plant based on a deep neural network. Background Art

[0002] With the continuous optimization of energy structure and the rapid development of renewable energy, virtual power plants, as a new type of power system dispatching and management mode, have received extensive attention. Virtual power plants can improve energy efficiency and system stability by integrating distributed energy resources (such as wind power, photovoltaics, and energy storage equipment) for unified dispatching and optimized management.

[0003] However, there are many challenges in the operation of existing virtual power plants, especially for the anomaly detection and response processing of distributed data, which relies on traditional monitoring and alarm systems. These systems have the following problems: High data complexity: The amount of distributed energy data is large and the types are complex. Traditional methods are difficult to process multimodal and heterogeneous data; High real-time requirements: The power system requires real-time monitoring and rapid response, and existing technologies are difficult to meet high real-time requirements; Low detection accuracy: Traditional anomaly detection technologies mostly use simple threshold judgments or rules based on experience, which are difficult to accurately capture complex abnormal patterns. Existing technologies generally detect anomalies by setting simple alarm thresholds or using rules based on experience. These methods are powerless when faced with massive and complex data, and are prone to false alarms and missed alarms, affecting the safety and reliability of the power system. Therefore, there is an urgent need for a more intelligent and efficient anomaly diagnosis method to solve this problem. Summary of the invention

[0004] The purpose of the present invention is to provide a virtual power plant distributed data anomaly diagnosis method based on deep neural network to improve the above problems. In order to achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0005] In a first aspect, the present application provides a virtual power plant distributed data anomaly diagnosis method based on a deep neural network, comprising:

[0006] Obtain multimodal data of each distributed energy source of the virtual power plant, perform data cleaning and normalization on the multimodal data, and obtain processed distributed data, wherein the multimodal data includes power output, meteorological data, and equipment operating status;

[0007] The processed distributed data is decomposed into three time scales: short-term, medium-term and long-term. The nodes of the virtual power plant and their relationships are modeled as a graph structure, and GNN is used to capture the spatiotemporal dependencies to obtain multi-time-scale spatiotemporal feature vectors. The nodes include wind farms, photovoltaic power stations and energy storage equipment. The short-term, medium-term and long-term time scales are data from minutes to hours, days to weeks, and months to years, respectively.

[0008] Based on the feature vector, pre-training is performed by constructing a self-supervised learning task to extract deep features. At the same time, the contrastive learning method is used to extract effective features between data of different time scales and modalities.

[0009] Based on effective features, a normal data distribution model is generated using a generative adversarial network. New normal sample data is analyzed from the normal data distribution model. By comparing the new normal sample data with the actual sample, anomalies that deviate from the normal distribution are detected. Based on the anomalies, abnormal data samples for enhancing model training are generated, and the abnormal data samples are aggregated into a comprehensive data set.

[0010] A multi-task learning framework is constructed based on the comprehensive data set, and multi-task learning of anomaly detection, load forecasting and fault diagnosis is carried out simultaneously. Through the shared features between tasks, a multi-task optimization model is obtained. Based on the multi-task optimization model, reinforcement learning is used to design the optimal abnormal response strategy. Through online learning and adaptive adjustment, the anomaly detection model is optimized in real time to obtain the final optimized abnormal response system that adapts to dynamic changes. Then, the task of distributed data anomaly diagnosis of the virtual power plant is completed through the abnormal response system.

[0011] Preferably, the processed distributed data is decomposed into three time scales: short-term, medium-term and long-term, including: short-term data is used to capture rapidly changing dynamic information in the power system, wherein the dynamic information includes instantaneous power fluctuations and instantaneous state changes of equipment; medium-term data is used to capture daily and periodic changes in the power system, wherein daily and periodic changes include daytime load fluctuations and weekend power consumption patterns; long-term data is used to capture seasonal and annual trends in the power system, wherein seasonal and annual trends include seasonal load changes and annual energy consumption patterns.

[0012] Preferably, the nodes of the virtual power plant and their relationships are modeled as a graph structure, which includes:

[0013] Model each energy node, load node, and substation node in the virtual power plant as a node in the graph;

[0014] The physical connections, power flows, and geographic location relationships between nodes are modeled as edges in a graph, where the weights of the edges are set according to the power transmission capacity, distance, and transmission efficiency, thereby forming a network graph containing all nodes and edges, where the network graph includes the topological structure and operating status of the virtual power plant.

[0015] Preferably, the GNN is used to capture the spatiotemporal dependency to obtain a spatiotemporal feature vector of multiple time scales, including:

[0016] Based on three time scales and graph structures, we use graph neural networks to perform spatiotemporal convolution operations, and perform graph convolution and time convolution processing on the data of each node at different time scales.

[0017] Combined with time convolution, it captures the dependency of nodes in the time dimension and obtains the graph feature vector containing spatiotemporal features;

[0018] The feature vector of the graph is subjected to feature extraction and fusion processing, which includes feature fusion of the feature vector of each node at different time scales to obtain the first fusion feature, and the feature vectors of each node at different time scales are integrated based on the first fusion feature to obtain the final multi-time-scale spatiotemporal feature vector.

[0019] Preferably, the method is performed by constructing a self-supervised learning task based on the feature vector to perform pre-training, extract deep features, and use a contrastive learning method to extract effective features between different time scales and modal data, including:

[0020] Based on the spatiotemporal feature vectors of multiple time scales, a self-supervised learning task is constructed, which includes pre-training the autoencoder with time scale and modality data, training the encoder and decoder for the feature vector of each node, reconstructing the input data and minimizing the reconstruction error to obtain deep features;

[0021] Based on deep features, the contrastive learning method is used to learn to compare feature vectors from different time scales and modal data. For each node, a twin network is constructed to calculate the similarity loss of positive and negative samples to obtain optimized deep features.

[0022] According to the optimized deep features, the feature vectors of each node in the short-term, medium-term and long-term time scales are integrated. Through the feature fusion technology, the effective feature vectors of different time scales and modal data are integrated, and finally a multi-time scale feature representation for anomaly detection and prediction tasks is obtained.

[0023] Preferably, based on the effective features, a normal data distribution model is generated by using a generative adversarial network, new normal sample data is obtained from the normal data distribution model by analysis, and abnormal points deviating from the normal distribution are detected by comparing the new normal sample data with the actual samples, and abnormal data samples for enhancing model training are generated according to the abnormal points, and the abnormal data samples are aggregated into a comprehensive data set, which includes:

[0024] The generator part of the generative adversarial network is trained based on the effective features, which includes the generator receiving random noise as input and generating new samples similar to the actual normal data distribution. During the training process, the difference between the generated samples and the actual samples is minimized;

[0025] Based on the trained generator, new normal samples are sampled from the noise distribution, the generated new samples are compared with the actual normal samples, and the anomaly detection algorithm is used to detect abnormal points that deviate from the normal distribution;

[0026] According to the detected anomalies, a generator of a generative adversarial network is used to generate abnormal data samples, where the abnormal data samples are abnormal noise inputs, and the abnormal data samples are simultaneously aggregated into a comprehensive data set.

[0027] Preferably, a multi-task learning framework is constructed based on the comprehensive data set, and multi-task learning of anomaly detection, load forecasting and fault diagnosis is performed simultaneously, and a multi-task optimization model is obtained through shared features between tasks, including:

[0028] Based on a comprehensive dataset, a multi-task learning framework is designed to simultaneously train anomaly detection, load forecasting, and fault diagnosis tasks, where the framework is used to improve the overall performance by sharing the underlying feature extractors and exploiting the correlation between different tasks;

[0029] Using a network structure to extract features from an input comprehensive data set, in a multi-task learning framework of a neural network, extract shared features by sharing hidden layers or shared convolution kernels, wherein the network structure includes a convolutional neural network, a recurrent neural network, or a transformer;

[0030] Multi-task learning for anomaly detection, load forecasting, and fault diagnosis is trained based on loss functions;

[0031] After multi-task learning training, a comprehensive multi-task optimization model is obtained based on shared features, where the multi-task optimization model is used to process anomaly detection, load forecasting and fault diagnosis tasks at the same time.

[0032] Preferably, the multi-task optimization model is based on the design of the optimal abnormal response strategy by using reinforcement learning, and the abnormal detection model is optimized in real time through online learning and adaptive adjustment to obtain the final optimized abnormal response system that adapts to dynamic changes, and then the task of distributed data abnormality diagnosis of the virtual power plant is completed through the abnormal response system, which includes:

[0033] The multi-task optimization model is learned according to the abnormal response strategy designed by reinforcement learning. In the reinforcement learning environment, the state s is defined as the current system state, the action a is the response strategy, and the reward r is the effect after the system responds. The Q-learning algorithm is used to maximize the cumulative reward through the optimization strategy to obtain the optimal abnormal response strategy.

[0034] Based on the optimal abnormal response strategy, the reinforcement learning model is continuously updated through real-time online learning to adapt to the dynamic changes of the virtual power plant system, and the deep deterministic policy gradient algorithm is used to optimize the policy parameters in the real-time data stream to obtain the final updated reinforcement learning model;

[0035] According to the updated reinforcement learning model, adaptive adjustment and optimization are performed through adaptive dynamic programming method, and model parameters are adjusted through real-time feedback to obtain the optimized anomaly detection model;

[0036] The optimized anomaly detection model is used to monitor and respond to abnormal situations in the distributed data of the virtual power plant in real time and complete the anomaly diagnosis task.

[0037] In a second aspect, the present application also provides a virtual power plant distributed data anomaly diagnosis device based on a deep neural network, comprising:

[0038] Memory for storing computer programs;

[0039] A processor is used to implement the steps of the virtual power plant distributed data anomaly diagnosis method based on deep neural network when executing the computer program.

[0040] In a third aspect, the present application also provides a readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned distributed data anomaly diagnosis method for a virtual power plant based on a deep neural network are implemented.

[0041] The beneficial effects of the present invention are:

[0042] Based on the preprocessing and fusion of multimodal data, the present invention solves the problem that traditional methods are difficult to process complex data, ensures the high quality of data input, and models the nodes of the virtual power plant and their relationships as a graph structure through multi-time scale spatiotemporal data modeling. The graph neural network (GNN) is used to capture spatiotemporal dependencies and realize multi-time scale spatiotemporal data modeling. This process effectively extracts short-term, medium-term and long-term spatiotemporal features, enhances the ability to capture complex abnormal patterns, and improves the accuracy of anomaly detection.

[0043] Based on self-supervised learning and feature extraction, the present invention constructs self-supervised learning tasks (such as future data point prediction, data reconstruction), performs deep feature extraction, and uses contrastive learning methods to extract effective features between different time scales and modal data. This process significantly improves the model's feature expression ability and adaptability to different data modes; by generating a normal data distribution model and high-quality abnormal data samples through a generative adversarial network (GAN), the model can generate and learn more abnormal samples, thereby enhancing the model's robustness and generalization ability.

[0044] The present invention constructs a multi-task learning framework based on multi-task learning and transfer learning, and simultaneously performs related tasks such as anomaly detection, load forecasting and fault diagnosis, and improves the detection effect by sharing features between tasks. The multi-task learning framework improves data utilization efficiency and improves the comprehensive performance of the model; the present invention designs the optimal anomaly response strategy based on reinforcement learning, optimizes the anomaly detection model in real time through online learning and adaptive adjustment, adapts to the dynamically changing system environment, and finally obtains an optimized anomaly response system, which can quickly respond to and handle abnormal situations to ensure the stable operation of the virtual power plant.

[0045] The present invention solves the limitations of traditional methods in processing complex data of virtual power plants by constructing a distributed data anomaly diagnosis method for virtual power plants based on deep neural networks, and realizes intelligent management of distributed energy resources. This method not only improves the accuracy and response speed of anomaly detection, but also realizes the rapid deployment and application of models through multi-task learning and transfer learning, significantly improving the operating efficiency and safety of virtual power plants.

[0046] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or be understood by implementing the embodiments of the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0048] Figure 1 It is a schematic flow chart of a method for diagnosing anomalies in distributed data of a virtual power plant based on a deep neural network according to an embodiment of the present invention;

[0049] Figure 2 Schematic diagram of the structure of the distributed data anomaly diagnosis device of a virtual power plant based on a deep neural network described in an embodiment of the present invention.

[0050] In the figure: 800, tag receiving device; 801, processor; 802, memory; 803, multimedia component; 804, I / O interface; 805, communication component. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0052] It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in the subsequent drawings. At the same time, in the description of the present invention, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0053] Embodiment 1:

[0054] This embodiment provides a virtual power plant distributed data anomaly diagnosis method based on deep neural network.

[0055] See also Figure 1 , the figure shows that the method includes step S100, step S200, step S300, step S400 and step S500.

[0056] S100. Acquire multimodal data of each distributed energy source of the virtual power plant, perform data cleaning and normalization on the multimodal data, and obtain processed distributed data, wherein the multimodal data includes power output, meteorological data, and equipment operating status.

[0057] It is understandable that in this step, in the virtual power plant system, the operating data of distributed energy resources such as wind power, photovoltaics and energy storage equipment are highly heterogeneous and multimodal, and these data types include but are not limited to power output data, meteorological data (such as wind speed, temperature, irradiance, etc.) and equipment operating status data (such as voltage, current, frequency, temperature, fault status, etc.), and obtaining these multimodal data is the basis of the abnormal diagnosis method. Among them, the data collection includes power output data, meteorological data and equipment operating status data, wherein the power output data includes data from the energy management system or the monitoring system of distributed energy, and the meteorological data includes wind speed, wind direction, temperature, humidity, irradiance, etc. These data can be obtained by means of meteorological stations, satellite remote sensing, Internet of Things sensors, etc.; the equipment operating status data involves the working parameters of the equipment, such as voltage, current, frequency, temperature and the operating status of the equipment (such as whether a fault occurs), and these data come from the monitoring system and sensor network of the equipment.

[0058] Specifically, the data cleaning step aims to remove noise and outliers in the data. For missing values, interpolation, mean filling and other methods can be used to process them. By detecting and correcting sensor failures, data entry errors and abnormal fluctuations, the integrity and accuracy of the data can be ensured. Common data cleaning techniques include mean filling, median filling, interpolation and more machine learning methods such as anomaly detection models. Secondly, the purpose of normalization is to convert data of different dimensions and ranges to the same scale for subsequent modeling and analysis. Common normalization methods include minimum-maximum normalization (scaling the data to the [0,1] interval) and standardization (converting the data to a distribution with a mean of 0 and a standard deviation of 1).

[0059] Therefore, by cleaning and normalizing the multimodal data of distributed energy in virtual power plants, the quality and consistency of the data can be significantly improved, and the impact of noise and outliers on subsequent analysis can be reduced. The cleaned data can better reflect the actual operating status and environmental conditions of the equipment, while the normalization process ensures the compatibility between different data types, facilitating model processing and feature extraction.

[0060] S200. Decompose the processed distributed data into three time scales: short-term, medium-term and long-term. Model the nodes of the virtual power plant and their relationships as a graph structure, and use GNN to capture the spatiotemporal dependencies to obtain multi-time-scale spatiotemporal feature vectors, where the nodes include wind farms, photovoltaic power stations and energy storage equipment. The short-term, medium-term and long-term time scales are data from minutes to hours, days to weeks, and months to years, respectively.

[0061] It can be understood that step S200 includes S201, S202 and S203:

[0062] S201. Short-term data is used to capture rapidly changing dynamic information in the power system, including instantaneous power fluctuations and instantaneous changes in equipment status; medium-term data is used to capture daily and periodic changes in the power system, including daily and periodic changes in daytime load fluctuations and weekend power consumption patterns; long-term data is used to capture seasonal and annual trends in the power system, including seasonal load changes and annual energy consumption patterns.

[0063] It should be noted that short-term data generally refers to minute-level or hour-level data, which is used to capture fast-changing dynamic information in the power system, such as instantaneous power fluctuations and instantaneous changes in equipment status. For example, due to the influence of weather conditions, the output power of wind power and photovoltaic power generation will fluctuate significantly in a short period of time, and instantaneous power fluctuation data can help identify these changes. Using short-term data, sliding window technology and fast Fourier transform (FFT) and other methods can be used for analysis to identify fast-changing abnormal patterns. Medium-term data is usually daily or weekly data, which is used to capture daily and periodic changes in the power system. Typical daily changes include daytime load fluctuations (such as morning peak and evening peak) and periodic changes (such as weekend power consumption patterns). For medium-term data, time series analysis methods such as autoregressive moving average model (ARIMA) and periodic pattern recognition algorithms can be applied to predict and detect daily and periodic anomalies. Long-term data refers to monthly, quarterly or annual data, which is used to capture seasonal and annual trends in the power system. For example, electricity consumption in winter and summer may increase significantly due to heating and cooling demand, and annual energy consumption patterns can reflect changes in overall economic activities and energy policies. Long-term data analysis can use trend analysis, seasonal decomposition methods (such as STL decomposition), and long short-term memory networks (LSTM) to identify anomalies in long-term trends and seasonal changes.

[0064] S202. Model each energy node, load node, and substation node in the virtual power plant as a node in a graph.

[0065] S203, modeling the physical connection, power flow and geographical location relationship between nodes as edges in a graph, wherein the weight of the edge is set according to the power transmission capacity, distance and transmission efficiency, thereby forming a network graph containing all nodes and edges, wherein the network graph includes the topological structure and operating status of the virtual power plant, and its calculation formula is as follows:

[0066] G=(V,E)

[0067] Where V is the node set and E is the edge set.

[0068] It should be noted that in step S202, the distributed energy nodes include wind turbines, photovoltaic panels, and energy storage devices, etc. The attributes of these nodes include the power output, device status, and geographic location of the equipment, etc. The load nodes represent power-consuming equipment or power-consuming areas, and the attributes of these nodes include load demand, power consumption mode, and geographic location, etc. The substation nodes represent key nodes for power transmission and distribution, and the attributes of these nodes include transmission capacity, substation status, and geographic location, etc.

[0069] In step S203, in the graph model, it is necessary not only to represent each node, but also to represent the connection relationship between the nodes. The physical connection, power flow and geographical location relationship between the nodes are represented by the edges in the graph, and the weight of the edge reflects the strength and nature of the connection. Among them, the physical connection represents the direct connection between power equipment and facilities, such as cables, wires and transformer connections, and each edge represents the physical connection between two nodes. And according to the actual physical layout of the virtual power plant, determine which nodes have direct connections, and represent these connections as edges in the graph. The power flow represents the path and direction of power transmission in the system. The weight of the edge can reflect the power size, transmission loss and transmission efficiency of the transmission. Among them, the geographical location relationship reflects the geographical distribution and spatial relationship of the nodes. There may be strong connections between nodes that are close to each other. Using geographic information system (GIS) data, the geographical distance between nodes is calculated, and the distance information is used as part of the weight of the edge to ensure that the model can reflect spatial dependencies.

[0070] That is, by modeling the physical connection, power flow and geographical location relationship between nodes as edges in the graph, and setting the weight of the edge according to the power transmission capacity, distance and transmission efficiency, a network graph containing all nodes and edges is formed. This network graph not only reflects the topological structure of the virtual power plant, but also contains information on the operating status. This structured representation lays the foundation for the application of graph neural networks, enabling the model to capture complex spatiotemporal dependencies and improve the accuracy and effectiveness of anomaly detection. Through this detailed and dynamic network graph modeling, the operating status of the virtual power plant is fully monitored and analyzed, enhancing the stability and responsiveness of the system.

[0071] It can be understood that step S200 also includes S204, S205 and S206:

[0072] S204. Based on the three time scales and graph structure, the graph neural network is used to perform spatiotemporal convolution operations, and graph convolution and time convolution processing are performed on the data of each node at different time scales. The calculation formula of the spatiotemporal convolution is as follows:

[0073]

[0074] In the formula, is the feature vector of node i at the kth layer, N(i) is the set of neighbor nodes of node i, c ij is the normalized coefficient between nodes i and j, W (k) and is the weight matrix of the kth layer, σ is the activation function;

[0075] S205, combining with time convolution, capturing the dependency of nodes in the time dimension, and obtaining a graph feature vector containing spatiotemporal features;

[0076] S206, performing feature extraction and fusion processing on the graph feature vector, including performing feature fusion on the feature vector of each node at different time scales to obtain a first fusion feature, integrating the feature vectors of each node at different time scales based on the first fusion feature to obtain a final multi-time-scale spatiotemporal feature vector.

[0077] It should be noted that, specifically, graph convolution is used to capture the spatial dependencies between nodes, and temporal convolution is used to capture the dependencies between nodes in the temporal dimension. Commonly used temporal convolution methods include one-dimensional convolutional neural networks (1D-CNN) and gated recurrent units (GRU). Therefore, by using graph neural networks to perform spatiotemporal convolution operations, the dependencies between nodes in the virtual power plant in the spatial and temporal dimensions can be captured at the same time. This method not only improves the ability to model complex relationships between nodes, but also enhances the ability to capture dynamic changes in time series. Graph convolution captures the spatial associations between nodes, and temporal convolution captures the changing trends of nodes at different time scales, enabling the model to more accurately identify abnormal patterns.

[0078] Specifically, in step S205, in the spatiotemporal convolution operation, the dependency of the time dimension is crucial. Data of different time scales are processed by the temporal convolution layer to extract patterns and features in the time series. By applying temporal convolution (such as 1D-CNN or GRU), data of different time scales are convolved, where the temporal convolution layer can identify patterns and changing trends of data in the time series. Combined with the previous graph convolution output, the results of the temporal convolution further enhance the spatiotemporal features of the nodes, and these feature vectors contain the spatial relationships of the nodes in the graph structure and the dynamic information in the time series. The generation of spatiotemporal feature vectors provides high-quality data input for subsequent feature extraction and fusion processing, improving the accuracy and robustness of anomaly detection.

[0079] Specifically, in step S206, the feature vectors of each node at different time scales are fused to obtain a first fused feature. The feature fusion method can adopt weighted average, splicing or self-attention mechanism (such as Transformer). Based on the first fused feature, the feature vectors of each node at different time scales are integrated to form the final multi-time scale spatiotemporal feature vector. The integration method can be implemented by weighted aggregation, ensemble learning and other means.

[0080] In summary, this step improves the representation and generalization capabilities of the model, making anomaly detection more accurate and reliable. The feature fusion vector provides high-quality input data for subsequent anomaly detection, load forecasting, and fault diagnosis, further enhancing the detection effect and adaptability of the system.

[0081] S300, based on the feature vector, pre-training is performed by constructing a self-supervised learning task to extract deep features, and at the same time, the contrastive learning method is used to extract effective features between different time scales and modal data.

[0082] It can be understood that step S300 includes S301, S302 and S303, wherein:

[0083] S301. Based on the spatiotemporal feature vectors of multiple time scales, a self-supervised learning task is constructed, which includes pre-training an autoencoder of time scale and modal data, training an encoder and a decoder for the feature vector of each node, reconstructing the input data and minimizing the reconstruction error to obtain deep features;

[0084] It should be noted that the input data of the multi-time-scale spatiotemporal feature vectors includes spatiotemporal feature vectors extracted from different time scales, which have been preliminarily processed and fused through the previous steps. The process of training the autoencoder is to map the input spatiotemporal feature vector X to a low-dimensional latent space Z, and train the autoencoder by minimizing the reconstruction error. After training, the latent space representation Z extracted by the encoder is the deep feature, which contains the high-order information of the input data. This method can better characterize the intrinsic structure and pattern of the data, thereby providing higher quality input for subsequent anomaly detection and prediction tasks. The application of self-supervised learning greatly reduces the dependence on manually labeled data and improves the training efficiency and generalization ability of the model.

[0085] S302, based on the deep features, using the contrastive learning method, learning to compare feature vectors from different time scales and modal data, building a twin network for each node, calculating the similarity loss of positive and negative samples, and obtaining optimized deep features;

[0086] S303. According to the optimized deep features, the feature vectors of each node at the short-term, medium-term and long-term time scales are integrated. Through the feature fusion technology, the effective feature vectors of different time scales and modal data are integrated, and finally a multi-time scale feature representation for anomaly detection and prediction tasks is obtained.

[0087] It should be noted that the deep features obtained in step S301 are used as the input of contrastive learning, and a twin network, that is, a neural network composed of two shared parameters, is used to process paired samples respectively, wherein the twin network is used to learn the similarity between feature vectors, pair data of different time scales and modes, and construct positive sample pairs (similar data pairs, such as different time scale data of the same node) and negative sample pairs (dissimilar data pairs, such as time scale data of different nodes). In step S302, the similarity loss calculation is the loss function of contrastive learning, which can be calculated using contrastive loss. By minimizing the contrastive loss, the twin network is trained to optimize the deep features, so that the distance between positive sample pairs is smaller and the distance between negative sample pairs is larger. Using the contrastive learning method, the deep features can be further optimized to have better distinguishing ability between different time scales and modal data. In step S303, for each node, its optimized deep features on short-term, medium-term and long-term time scales are collected, and these feature vectors are spliced ​​together to form a comprehensive feature vector containing multi-time scale information. Among them, feature fusion technology can adopt a variety of feature fusion technologies, including weighted averaging, splicing, attention mechanism, etc. The attention mechanism is particularly suitable for capturing the correlation and importance between different time scales. Through feature fusion, the final multi-time scale feature representation is obtained, which integrates the effective information of different time scales and modal data.

[0088] Therefore, it can be understood that through feature integration and fusion processing, the final feature representation containing multi-time scale information is obtained, and the correlation between different time scales and modal data can be captured. The fused feature representation provides richer and more accurate input data for anomaly detection and prediction tasks, improving the prediction accuracy and detection effect of the model. The application of feature fusion technology greatly enhances the comprehensive characterization ability of data, making the abnormal diagnosis of virtual power plants more efficient and reliable.

[0089] S400. Based on effective features, a normal data distribution model is generated using a generative adversarial network. New normal sample data is analyzed from the normal data distribution model. By comparing the new normal sample data with the actual sample, abnormal points that deviate from the normal distribution are detected, and abnormal data samples for enhancing model training are generated based on the abnormal points, and the abnormal data samples are aggregated into a comprehensive data set.

[0090] It can be understood that the step S400 includes S401, S402 and S403, wherein:

[0091] S401, training a generator part of a generative adversarial network according to effective features, wherein the generator receives random noise as input and generates new samples similar to the actual normal data distribution, and during the training process, minimizes the difference between the generated samples and the actual samples;

[0092] S402, based on the trained generator, sampling from the noise distribution to obtain new normal samples, comparing the generated new samples with the actual normal samples, and using anomaly detection algorithms to detect abnormal points that deviate from the normal distribution;

[0093] S403. Generate abnormal data samples using a generator of a generative adversarial network according to the detected abnormal points, wherein the abnormal data samples are abnormal noise inputs, and at the same time aggregate the abnormal data samples into a comprehensive data set.

[0094] It should be noted that the valid feature vectors obtained in the previous steps are used as basic data. These feature vectors are processed in multiple layers, contain data features of different time scales and modes, and train the generator. The adversarial training mechanism of the generative adversarial network enables the generator to continuously improve the quality of the generated samples, and the samples finally generated can effectively simulate the distribution characteristics of normal data. In step S402, a noise vector is sampled from a standard normal distribution, and the noise vector is mapped to a new sample through the generator G. The normal sample obtained from the actual data is used, and an anomaly detection algorithm (such as a density estimation-based method or a machine learning-based method) is used to compare the generated sample and the actual sample to detect abnormal points that deviate from the normal distribution. When detecting abnormal points, the difference metric between the generated sample and the actual sample can be calculated. Commonly used methods include Euclidean distance, Mahalanobis distance, etc., and by setting a threshold ∈, the sample whose difference exceeds the threshold is identified as an abnormal point. By comparing the generated new sample with the actual normal sample, the abnormal point that deviates from the normal distribution can be effectively detected. The new sample generated by the generative adversarial network has a distribution characteristic similar to that of the actual sample, so it can provide a reliable benchmark for abnormal detection.

[0095] It is understandable that, in step S403, based on the anomaly detection result in step S402, anomalies that deviate from the normal distribution are identified, and the anomalies are used as noise input to generate abnormal data samples through the generator. The generated abnormal data samples are collected and aggregated to form a comprehensive data set for model training, including normal data and abnormal data, which can effectively enrich the training data set, so that the model can be exposed to more abnormal situations during the training process, which helps to improve the model's ability to identify abnormal situations, thereby enhancing the robustness and accuracy of the system.

[0096] S500. Construct a multi-task learning framework based on the comprehensive data set, and perform multi-task learning of anomaly detection, load forecasting and fault diagnosis simultaneously. Through the shared features between tasks, a multi-task optimization model is obtained. Based on the multi-task optimization model, the optimal abnormal response strategy is designed using reinforcement learning. Through online learning and adaptive adjustment, the anomaly detection model is optimized in real time to obtain the final optimized abnormal response system that adapts to dynamic changes. Then, the task of distributed data anomaly diagnosis of the virtual power plant is completed through the abnormal response system.

[0097] It can be understood that step S500 includes S501, S502, S503 and S504, wherein:

[0098] S501. Based on a comprehensive dataset, a multi-task learning framework is designed to simultaneously train anomaly detection, load forecasting, and fault diagnosis tasks, where the framework is used to improve the overall performance by sharing the underlying feature extractors and exploiting the correlation between different tasks.

[0099] It should be noted that the comprehensive dataset contains normal and abnormal samples generated in the previous steps, covering features of multiple data modalities and time scales, and constructs a neural network architecture with shared layers and task-specific branches. The shared layers are used to extract underlying features, while each task-specific branch is optimized for a specific task. Among them, the initial layers of the network are designed as shared layers, which extract common features by processing all samples in the comprehensive dataset, including spatiotemporal features of different time scales, equipment operating status, etc. Above the shared layers, the network is divided into three branches, each of which focuses on a specific task (anomaly detection, load forecasting, fault diagnosis), and these tasks have certain correlations, and using these correlations can help improve the learning effect of each task. Therefore, by sharing the underlying feature extractor, each task can use the relevant information of each other to improve the efficiency and accuracy of feature extraction; the use of correlations between tasks can help the model of each task better learn complex features and improve the overall performance and generalization ability.

[0100] S502, extracting features from the input comprehensive data set using a network structure, and extracting shared features by sharing hidden layers or shared convolution kernels in a multi-task learning framework of a neural network, wherein the network structure includes a convolutional neural network, a recurrent neural network, or a transformer;

[0101] It should be noted that convolutional neural networks are suitable for extracting spatial features, especially for processing spatiotemporal data and image data. Shared features in space can be extracted by sharing convolution kernels, while recurrent neural networks are suitable for processing sequence data, especially time series. Long-term dependencies of time series can be captured by sharing hidden layers. In the multi-task learning framework, the initial layers of the network are designed as shared layers, which include convolutional layers, hidden layers or self-attention layers. These layers extract shared features that can be used for multiple tasks by processing comprehensive data sets. Feature extraction is the process of processing the comprehensive data set by the network to extract feature representations, which include spatiotemporal features of nodes, power output features, device status features, etc. In this step, by using an appropriate network structure, multimodal features and multi-time scale features in the comprehensive data set can be effectively extracted. The design of the shared layer enables different tasks to share the underlying features, thereby improving the efficiency and accuracy of feature extraction.

[0102] S503, training the multi-task learning of anomaly detection, load forecasting and fault diagnosis according to the loss function, the calculation formula is as follows:

[0103]

[0104] In the formula, λ anomaly ,λ load ,λ fault is the task weight coefficient, is the overall loss function;

[0105] S504. After multi-task learning training, a comprehensive multi-task optimization model is obtained based on shared features, wherein the multi-task optimization model is used to process anomaly detection, load forecasting and fault diagnosis tasks at the same time.

[0106] It should be noted that in the multi-task learning framework, the model simultaneously receives training data for anomaly detection, load forecasting, and fault diagnosis. Through the back propagation algorithm, the parameters of the model are optimized under the joint action of multiple tasks. During the training process, the model uses the shared feature layer to extract common features. These shared features provide a unified basis for multiple tasks, so that each task can benefit from them. The trained model contains the optimization parameters of multiple tasks. The structural design and training process of the model enable it to efficiently handle multiple tasks at the same time. Through the multi-task learning framework, the model can balance the requirements between different tasks when optimizing multiple tasks at the same time, thereby improving the overall performance. The optimization goal is not only the accuracy of a single task, but also the comprehensive performance of multiple tasks. Therefore, the comprehensive multi-task optimization model can handle multiple tasks at the same time, avoiding the redundant calculation of multiple individual models and improving the overall efficiency of the system. Multi-task learning can utilize the correlation between tasks to improve the performance of each task. The shared feature layer ensures that different tasks promote each other during the training process, thereby improving the comprehensive performance of the model.

[0107] It can be understood that step S500 also includes S505, S506, S507 and S508, wherein:

[0108] S505, learning the multi-task optimization model according to the abnormal response strategy designed by reinforcement learning, wherein in the reinforcement learning environment, the state s is defined as the current system state, the action a is the response strategy, and the reward r is the effect after the system responds. The Q-learning algorithm is used to maximize the cumulative reward by optimizing the strategy to obtain the optimal abnormal response strategy;

[0109] S506. Based on the optimal abnormal response strategy, the reinforcement learning model is continuously updated through real-time online learning to adapt to the dynamic changes of the virtual power plant system, and the policy parameters in the real-time data stream are optimized using a deep deterministic policy gradient algorithm to obtain the final updated reinforcement learning model;

[0110] S507, according to the updated reinforcement learning model, adaptive adjustment and optimization are performed through an adaptive dynamic programming method, and model parameters are adjusted through real-time feedback to obtain an optimized anomaly detection model;

[0111] It should be noted that in step S505, the abnormal response strategy is optimized through the Q-learning algorithm, which can select the optimal action under different states, maximize the system's cumulative rewards, and improve the system's fault response and recovery capabilities. Specifically, in step S506, the system continuously receives new data and updates the model parameters in real time to ensure that the model can quickly adapt to changes. Through the combination of real-time online learning and the DDPG algorithm, the reinforcement learning model can adaptively adjust the strategy parameters, optimize the system's response capabilities in real time, and ensure that the best performance is maintained in a dynamically changing environment. And through the adaptive dynamic programming method, the system can adjust and optimize the model parameters in real time, further improve the accuracy and response speed of abnormal detection, and ensure that the virtual power plant can quickly resume normal operation under abnormal conditions.

[0112] S508. Use the optimized anomaly detection model to monitor and respond to abnormal situations in the distributed data of the virtual power plant in real time and complete the abnormal diagnosis task. The calculation formula is as follows:

[0113] y diagnosis =M final (x input )

[0114] In the formula, x input is the input data, y diagnosis For the diagnosis result.

[0115] S500. Construct a multi-task learning framework based on the comprehensive data set, and perform multi-task learning of anomaly detection, load forecasting and fault diagnosis simultaneously. Through the shared features between tasks, a multi-task optimization model is obtained. Based on the multi-task optimization model, the optimal abnormal response strategy is designed using reinforcement learning. Through online learning and adaptive adjustment, the anomaly detection model is optimized in real time to obtain the final optimized abnormal response system that adapts to dynamic changes. Then, the task of distributed data anomaly diagnosis of the virtual power plant is completed through the abnormal response system.

[0116] Embodiment 2:

[0117] Corresponding to the above method embodiments, this embodiment also provides a virtual power plant distributed data anomaly diagnosis device based on a deep neural network. The virtual power plant distributed data anomaly diagnosis device based on a deep neural network described below and the virtual power plant distributed data anomaly diagnosis method based on a deep neural network described above can refer to each other.

[0118] Figure 2 8 is a block diagram of a virtual power plant distributed data anomaly diagnosis device 800 based on a deep neural network according to an exemplary embodiment. Figure 2As shown, the virtual power plant distributed data anomaly diagnosis device 800 based on deep neural network includes: a processor 801 and a memory 802. The virtual power plant distributed data anomaly diagnosis device 800 based on deep neural network also includes one or more of a multimedia component 803, an I / O interface 804, and a communication component 805.

[0119] The processor 801 is used to control the overall operation of the virtual power plant distributed data anomaly diagnosis device 800 based on a deep neural network to complete all or part of the steps in the above-mentioned virtual power plant distributed data anomaly diagnosis method based on a deep neural network. The memory 802 is used to store various types of data to support the operation of the virtual power plant distributed data anomaly diagnosis device 800 based on a deep neural network. These data may include, for example, instructions for any application or method operating on the virtual power plant distributed data anomaly diagnosis device 800 based on a deep neural network, as well as application-related data, such as contact data, sent and received messages, pictures, audio, video, etc. The memory 802 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, disk or optical disk. The multimedia component 803 may include a screen and an audio component. The screen may be, for example, a touch screen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signal may be further stored in the memory 802 or sent via the communication component 805. The audio component also includes at least one speaker for outputting audio signals. The I / O interface 804 provides an interface between the processor 801 and other interface modules, and the above-mentioned other interface modules may be a keyboard, a mouse or buttons, etc. These buttons may be virtual buttons or physical buttons. The communication component 805 is used for wired or wireless communication between the virtual power plant distributed data anomaly diagnosis device 800 based on deep neural network and other devices. Wireless communication, such as Wi-Fi, Bluetooth, near field communication (NFC), 2G, 3G or 4G, or a combination of one or more of them, so the corresponding communication component 805 may include: a Wi-Fi module, a Bluetooth module or an NFC module.

[0120] In an exemplary embodiment, the distributed data anomaly diagnosis device 800 of a virtual power plant based on a deep neural network can be implemented by one or more application-specific integrated circuits (ASIC), digital signal processors (DSP), digital signal processing devices (DSPD), programmable logic devices (PLD), field programmable gate arrays (FPGA), controllers, microcontrollers, microprocessors or other electronic components to execute the above-mentioned distributed data anomaly diagnosis method of a virtual power plant based on a deep neural network.

[0121] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided, and when the program instructions are executed by a processor, the steps of the above-mentioned virtual power plant distributed data anomaly diagnosis method based on a deep neural network are implemented. For example, the computer-readable storage medium can be the above-mentioned memory 802 including program instructions, and the above-mentioned program instructions can be executed by the processor 801 of the virtual power plant distributed data anomaly diagnosis device 800 based on a deep neural network to complete the above-mentioned virtual power plant distributed data anomaly diagnosis method based on a deep neural network.

[0122] Corresponding to the above method embodiment, a readable storage medium is also provided in this embodiment. The readable storage medium described below and the distributed data anomaly diagnosis method of a virtual power plant based on a deep neural network described above can refer to each other.

[0123] A computer program is stored on the readable storage medium, and when the computer program is executed by the processor, the steps of the distributed data anomaly diagnosis method of a virtual power plant based on a deep neural network in the above-mentioned method embodiment are implemented.

[0124] The readable storage medium may specifically be a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or other readable storage medium that can store program codes.

[0125] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0126] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A virtual power plant distributed data anomaly diagnosis method based on deep neural network, characterized in that: include: Obtain multimodal data of each distributed energy source of the virtual power plant, perform data cleaning and normalization on the multimodal data, and obtain processed distributed data, wherein the multimodal data includes power output, meteorological data, and equipment operating status; The processed distributed data is decomposed into three time scales: short-term, medium-term and long-term. The nodes of the virtual power plant and their relationships are modeled as a graph structure, and GNN is used to capture the spatiotemporal dependencies to obtain multi-time-scale spatiotemporal feature vectors. The nodes include wind farms, photovoltaic power stations and energy storage equipment. The short-term, medium-term and long-term time scales are data from minutes to hours, days to weeks, and months to years, respectively. Based on the feature vector, pre-training is performed by constructing a self-supervised learning task to extract deep features. At the same time, the contrastive learning method is used to extract effective features between data of different time scales and modalities. Based on effective features, a normal data distribution model is generated using a generative adversarial network. New normal sample data is analyzed from the normal data distribution model. By comparing the new normal sample data with the actual sample, anomalies that deviate from the normal distribution are detected. Based on the anomalies, abnormal data samples for enhancing model training are generated, and the abnormal data samples are aggregated into a comprehensive data set. A multi-task learning framework is built based on the comprehensive data set. Multi-task learning of anomaly detection, load forecasting, and fault diagnosis is performed simultaneously. The multi-task optimization model is obtained through the shared features between tasks. Based on the multi-task optimization model, reinforcement learning is used to design the optimal abnormal response strategy. Through online learning and adaptive adjustment, the anomaly detection model is optimized in real time to obtain the final optimized abnormal response system that adapts to dynamic changes. Then, the task of distributed data anomaly diagnosis of the virtual power plant is completed through the abnormal response system.

2. The method for distributed data anomaly diagnosis of virtual power plants based on deep neural networks according to claim 1 is characterized in that: The processed distributed data is decomposed into three time scales: short-term, medium-term and long-term, including: short-term data is used to capture rapidly changing dynamic information in the power system, where the dynamic information includes instantaneous power fluctuations and instantaneous state changes of equipment; medium-term data is used to capture daily and periodic changes in the power system, where daily and periodic changes include daytime load fluctuations and weekend power consumption patterns; long-term data is used to capture seasonal and annual trends in the power system, where seasonal and annual trends include seasonal load changes and annual energy consumption patterns.

3. The method for distributed data anomaly diagnosis of virtual power plants based on deep neural networks according to claim 1 is characterized in that: The nodes of the virtual power plant and their relationships are modeled as a graph structure, which includes: Model each energy node, load node, and substation node in the virtual power plant as a node in the graph; The physical connections, power flows, and geographic location relationships between nodes are modeled as edges in the graph, where the weights of the edges are set according to the power transmission capacity, distance, and transmission efficiency, thereby forming a network graph containing all nodes and edges. The network graph includes the topology and operating status of the virtual power plant, and its calculation formula is as follows: G=(V,E) Where V is the node set and E is the edge set.

4. The method for distributed data anomaly diagnosis of virtual power plants based on deep neural networks according to claim 1 is characterized in that: The GNN is used to capture the spatiotemporal dependencies and obtain spatiotemporal feature vectors at multiple time scales, including: Based on three time scales and graph structures, the graph neural network is used to perform spatiotemporal convolution operations, and graph convolution and time convolution are performed on the data of each node at different time scales. The calculation formula for the spatiotemporal convolution is as follows: In the formula, is the feature vector of node i in the kth layer, N(i) is the set of neighbor nodes of node i, c ij is the normalized coefficient between nodes i and j, W (k) and is the weight matrix of the kth layer, σ is the activation function; Combined with time convolution, it captures the dependency of nodes in the time dimension and obtains the graph feature vector containing spatiotemporal features; The feature vector of the graph is subjected to feature extraction and fusion processing, which includes feature fusion of the feature vector of each node at different time scales to obtain the first fusion feature, and the feature vectors of each node at different time scales are integrated based on the first fusion feature to obtain the final multi-time-scale spatiotemporal feature vector.

5. The method for distributed data anomaly diagnosis of virtual power plants based on deep neural networks according to claim 1 is characterized in that: According to the feature vector, the self-supervised learning task is constructed for pre-training to extract deep features, and the contrastive learning method is used to extract effective features between different time scales and modal data, including: Based on the spatiotemporal feature vectors of multiple time scales, a self-supervised learning task is constructed, which includes pre-training the autoencoder with time scale and modality data, training the encoder and decoder for the feature vector of each node, reconstructing the input data and minimizing the reconstruction error to obtain deep features; Based on deep features, the contrastive learning method is used to learn to compare feature vectors from different time scales and modal data. For each node, a twin network is constructed to calculate the similarity loss of positive and negative samples to obtain optimized deep features. According to the optimized deep features, the feature vectors of each node in the short-term, medium-term and long-term time scales are integrated. Through the feature fusion technology, the effective feature vectors of different time scales and modal data are integrated, and finally a multi-time scale feature representation for anomaly detection and prediction tasks is obtained.

6. The method for distributed data anomaly diagnosis of virtual power plants based on deep neural networks according to claim 1 is characterized in that: Based on the effective features, a normal data distribution model is generated by using a generative adversarial network, and new normal sample data is obtained from the normal data distribution model. By comparing the new normal sample data with the actual sample, abnormal points that deviate from the normal distribution are detected, and abnormal data samples for enhancing model training are generated according to the abnormal points, and the abnormal data samples are summarized into a comprehensive data set, which includes: The generator part of the generative adversarial network is trained based on the effective features, including the generator receiving random noise as input and generating new samples similar to the actual normal data distribution. During the training process, the difference between the generated samples and the actual samples is minimized; Based on the trained generator, new normal samples are sampled from the noise distribution, the generated new samples are compared with the actual normal samples, and the anomaly detection algorithm is used to detect abnormal points that deviate from the normal distribution; According to the detected anomalies, a generator of a generative adversarial network is used to generate abnormal data samples, where the abnormal data samples are abnormal noise inputs, and the abnormal data samples are summarized into a comprehensive data set at the same time.

7. The method for distributed data anomaly diagnosis of virtual power plants based on deep neural networks according to claim 1 is characterized in that: The multi-task learning framework is constructed based on the comprehensive data set, and multi-task learning of anomaly detection, load forecasting and fault diagnosis is performed simultaneously. Through the shared features between tasks, a multi-task optimization model is obtained, which includes: Based on a comprehensive dataset, a multi-task learning framework is designed to simultaneously train anomaly detection, load forecasting, and fault diagnosis tasks, where the framework is used to improve the overall performance by sharing the underlying feature extractors and exploiting the correlation between different tasks; Using a network structure to extract features from an input comprehensive data set, in a multi-task learning framework of a neural network, extract shared features by sharing hidden layers or shared convolution kernels, wherein the network structure includes a convolutional neural network, a recurrent neural network, or a transformer; The multi-task learning of anomaly detection, load forecasting and fault diagnosis is trained according to the loss function, and its calculation formula is as follows: In the formula, λ anomaly ,λ load ,λ fault is the task weight coefficient, is the overall loss function; After multi-task learning training, a comprehensive multi-task optimization model is obtained based on shared features, where the multi-task optimization model is used to process anomaly detection, load forecasting and fault diagnosis tasks at the same time.

8. The method for distributed data anomaly diagnosis of virtual power plants based on deep neural networks according to claim 1 is characterized in that: Based on the multi-task optimization model, the optimal abnormal response strategy is designed by using reinforcement learning. Through online learning and adaptive adjustment, the abnormal detection model is optimized in real time to obtain the final optimized abnormal response system that adapts to dynamic changes. Then, the task of distributed data abnormality diagnosis of the virtual power plant is completed through the abnormal response system, which includes: The multi-task optimization model is learned according to the abnormal response strategy designed by reinforcement learning. In the reinforcement learning environment, the state s is defined as the current system state, the action a is the response strategy, and the reward r is the effect after the system responds. The Q-learning algorithm is used to maximize the cumulative reward through the optimization strategy to obtain the optimal abnormal response strategy. Based on the optimal abnormal response strategy, the reinforcement learning model is continuously updated through real-time online learning to adapt to the dynamic changes of the virtual power plant system, and the deep deterministic policy gradient algorithm is used to optimize the policy parameters in the real-time data stream to obtain the final updated reinforcement learning model; According to the updated reinforcement learning model, adaptive adjustment and optimization are performed through adaptive dynamic programming method, and model parameters are adjusted through real-time feedback to obtain the optimized anomaly detection model; The optimized anomaly detection model is used to monitor and respond to anomalies in the distributed data of the virtual power plant in real time and complete the anomaly diagnosis task. The calculation formula is as follows: y diagnosis =M final (x input ) In the formula, x input is the input data, y diagnosis For the diagnosis result.

Images