Intelligent monitoring method and system for operation state of electric power system

Abstract

The invention provides a power system operation state intelligent monitoring method and system, and relates to the technical field of power data monitoring, and the method comprises the steps: obtaining multi-source heterogeneous operation data, obtaining a standard operation data set, adaptively extracting high-order multi-modal feature representation through a confidence weighted fusion method, and obtaining an operation state multi-modal feature set. Constructing an intelligent monitoring cognitive map by learning topological association characteristics and semantic association rules of the running state; mapping the real-time operation data into an intelligent monitoring cognitive map, dynamically sensing an operation state spatio-temporal evolution rule through a spatio-temporal dependence modeling unit and a graph feature propagation unit in an operation state graph learning model, and generating a prediction and diagnosis result; and learning and generating an optimal control strategy of the power system through a depth deterministic strategy gradient algorithm, decoupling the optimal control strategy of the power system into a plurality of local sub-strategies, issuing the local sub-strategies to corresponding power control equipment, and obtaining a distributed cooperative control scheme through a federated learning algorithm.

Description

Technical Field

[0001] The present invention relates to the technical field of power data monitoring, and in particular to a method and system for intelligently monitoring the operating status of a power system. Background Art

[0002] As the scale of power systems continues to expand and their complexity increases, ensuring the safe and stable operation of power systems faces new challenges. Traditional power system monitoring methods mainly rely on manual inspections and regular maintenance, which have problems such as slow response, low efficiency, and low accuracy. In order to meet the development needs of power systems, intelligent monitoring methods for system operation status are urgently needed.

[0003] In the existing technology, monitoring and diagnosis are carried out on a single device or a single type of fault, and there is a lack of comprehensive monitoring and analysis of the overall operating status of the power system; a large amount of multi-source heterogeneous data is generated during the operation of the power system, but the degree of data mining and utilization is limited, and the value of the data is not fully explored; the status judgment relies on pre-defined rules and thresholds, and the degree of intelligence is not high, making it difficult to adapt to the complex and changeable operating conditions of the power system; the monitoring and diagnosis results lack transparency and explainability, which is not conducive to the monitoring personnel to make decisions and disposal.

[0004] In summary, the intelligent monitoring method for the operation status of the power system of the present invention realizes comprehensive perception, real-time analysis, intelligent diagnosis and explainable decision support of the operation status of the power system by integrating technologies such as the Internet of Things, big data, and artificial intelligence. Compared with the prior art, the present invention can solve the problems in the prior art. Summary of the invention

[0005] The embodiments of the present invention provide a method and system for intelligently monitoring the operating status of an electric power system, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] Provided is a method for intelligently monitoring the operation status of a power system, comprising:

[0008] Acquire multi-source heterogeneous operation data of the power system, perform standardized preprocessing on the multi-source heterogeneous operation data to obtain a standard operation data set, input the standard operation data set into a heterogeneous data feature extraction model, adaptively extract high-order multimodal feature representations through a confidence weighted fusion method, obtain a multimodal feature set of the operation state, and construct an intelligent monitoring cognitive map by learning the topological association features and semantic association rules of the operation state;

[0009] Based on the intelligent monitoring cognitive graph, an operation state graph learning model is constructed to map the real-time operation data to the intelligent monitoring cognitive graph. Through the spatiotemporal dependency modeling unit and the graph feature propagation unit in the operation state graph learning model, the spatiotemporal evolution law of the operation state is dynamically perceived to generate prediction and diagnosis results.

[0010] Based on the prediction and diagnosis results, the optimal control strategy of the power system is generated through deep deterministic policy gradient algorithm learning, and the optimal control strategy of the power system is decoupled into multiple local sub-strategies. The strategies are sent to the corresponding power control equipment, and the distributed collaborative control scheme is obtained through the federated learning algorithm. The control process corresponding to the distributed collaborative control scheme is displayed in real time through the visualization equipment.

[0011] In an optional embodiment,

[0012] Acquire multi-source heterogeneous operation data of the power system, perform standardized preprocessing on the multi-source heterogeneous operation data to obtain a standard operation data set, input the standard operation data set into a heterogeneous data feature extraction model, and adaptively extract high-order multimodal feature representations through a confidence weighted fusion method to obtain an operation state multimodal feature set including:

[0013] Acquire multi-source heterogeneous operation data of the power system, wherein the multi-source heterogeneous operation data includes real-time operation parameters, historical operation data, weather data, geographic information data, and equipment status data;

[0014] Correct the multi-source heterogeneous operation data through outlier detection, missing value filling and duplicate value deletion, unify the scale and distribution of the multi-source heterogeneous operation data through maximum and minimum value normalization and Z-score normalization, and obtain a standard operation data set;

[0015] Inputting the standard operation data set into a heterogeneous data feature extraction model, wherein the heterogeneous data feature extraction model includes a multi-layer deep feature extraction network and an adversarial network;

[0016] Based on the dimension corresponding to the standard running data set, a first neural network node group is set to construct an input observation layer, according to a preset feature quantity threshold, a second neural network node group is set to construct a potential capture layer, an interaction is constructed between the input observation layer and the potential capture layer, according to a preset shielding probability value, some nodes in the first neural network node group and the second neural network node group are randomly shielded, a hierarchical feature extraction neural network is determined, the hierarchical feature extraction neural network is stacked in multiple layers to construct a multi-layer deep feature extraction network, based on a contrast divergence algorithm, the corresponding weight parameters of each layer of the hierarchical feature extraction neural network are iteratively adjusted, and a high-order multimodal feature representation is output;

[0017] Inputting the high-order multimodal feature representation into an adversarial network, and capturing the multimodal feature distribution through game learning between a generator and a discriminator in the adversarial network;

[0018] Based on the multimodal feature distribution, combined with the confidence weighted fusion method, the confidence of different modal features is determined, and the weights of different modal features are calculated through the softmax function. Adaptive fusion is performed through weighted averaging to obtain a multimodal feature set of the operating state.

[0019] In an optional embodiment,

[0020] The training of the heterogeneous data feature extraction model includes:

[0021] Based on the heterogeneous data feature extraction model, the number of nodes in the second neural network node group, the number of layers in the multi-layer stack, and the corresponding model hyperparameters are used as model configuration to determine the solution space;

[0022] In the solution space, individuals are randomly initialized to determine an initial population, a fitness value corresponding to each individual in the initial population is calculated, and based on the fitness value, a first individual corresponding to an optimal solution, a second individual corresponding to a suboptimal solution, a third individual corresponding to a third optimal solution, and a remaining common individual group are determined;

[0023] Based on the first individual, the second individual and the third individual, randomly determine basic offspring individuals according to a preset neighborhood range, perform crossover and mutation operations on the basic offspring individuals, and generate a preferred offspring individual group;

[0024] Based on the common individual group, a plurality of individuals are randomly selected, and crossover and mutation operations are performed to generate common offspring and add them to the common individual group, and based on the positions of the first individual, the second individual, and the third individual, the position of each individual in the common individual group is updated according to the individual position update formula;

[0025] The first individual, the second individual, the third individual, the preferred offspring individual group and the updated common individual group constitute an iterative population;

[0026] Calculating the fitness value corresponding to each individual in the iterative population, and re-determining the first individual, the second individual, the third individual and the common individual group;

[0027] Repeat the iterative update until the preset number of iterations is reached, and the first individual finally obtained becomes the optimal individual, determining the optimal structural data feature extraction model.

[0028] In an optional embodiment,

[0029] The individual position update formula is as follows:

[0030]

[0031] Among them, X1 represents the position update of the ordinary individual under the guidance of the first individual, X2 represents the position update of the ordinary individual under the guidance of the second individual, X3 represents the position update of the ordinary individual under the guidance of the third individual, Xα represents the position of the first individual, X β represents the second individual position, X δ represents the position of the third individual, r1 represents the first random vector, r2 represents the second random vector, r3 represents the third random vector, k represents the iteration control parameter, C1 represents the scaling factor vector with the first individual position, C2 represents the scaling factor vector with the second individual position, C3 represents the scaling factor vector with the third individual position, t represents the current number of iterations, X(t) represents the position of the ordinary individual at the tth iteration, X(t+1) represents the position of the ordinary individual at the t+1th iteration, T represents the maximum number of iterations, ω1 represents the influence weight of the first individual, ω2 represents the influence weight of the second individual, and ω3 represents the influence weight of the third individual.

[0032] In an optional embodiment,

[0033] Based on the intelligent monitoring cognitive graph, an operation state graph learning model is constructed, and real-time operation data is mapped to the intelligent monitoring cognitive graph. Through the spatiotemporal dependency modeling unit and graph feature propagation unit in the operation state graph learning model, the spatiotemporal evolution law of the operation state is dynamically perceived, and prediction and diagnosis results are generated, including:

[0034] Mapping the real-time operation data to the intelligent monitoring cognitive graph, representing the nodes in the intelligent monitoring cognitive graph as low-dimensional dense vectors through node embedding, and generating an operation status feature matrix;

[0035] Input the running state feature matrix into the gated recurrent unit, learn the state time evolution pattern of the node by calculating the update gate, reset gate and candidate state, and obtain the node hidden state sequence; based on each node, dot product the hidden state corresponding to the node with the hidden state of the neighbor node, calculate the similarity between the node and the neighbor node, determine the attention weight through normalization operation, and sum the hidden state corresponding to the neighbor node according to the attention weight to obtain the node spatial context feature; splice the node hidden state sequence with the node spatial context feature, integrate the spatiotemporal dependency, and generate the node state representation;

[0036] Based on the node state representation, combined with the causal semantics between nodes, the causal attention between nodes is calculated to construct a causal attention matrix; the causal attention matrix is ​​used as an adjacency matrix, the information of neighbor nodes is aggregated through the adjacency matrix, and feature transformation is performed in combination with a parameter matrix and a nonlinear function. After multiple layers of graph convolution are stacked, the interaction between nodes is captured, and the node state representation is spatially updated;

[0037] The updated node state representation is input into the prediction and diagnosis module, and the node state distribution at the future moment is predicted through the sequence-to-sequence model. The abnormality measurement of each node state representation in the node state distribution is performed to determine the abnormal nodes and associated paths.

[0038] A loss function including state prediction loss and abnormal diagnosis loss is constructed. The parameters of the operation state graph learning model are trained through the back propagation algorithm and the gradient descent optimization method. The optimal operation state graph learning model is generated by minimizing the loss function to obtain the prediction and diagnosis results.

[0039] In an optional embodiment,

[0040] Based on the prediction and diagnosis results, the optimal control strategy of the power system is generated by deep deterministic policy gradient algorithm learning, and the optimal control strategy of the power system is decoupled into multiple local sub-strategies including:

[0041] Based on the prediction and diagnosis results, determining the optimal control problem, constructing a state space and an action space, wherein the state space includes a current state, a future state in the prediction and diagnosis results, and an abnormal state, and the action space includes a power system control variable corresponding to the control strategy;

[0042] Based on the deep deterministic policy gradient algorithm, the optimal control problem is solved to generate a global optimal control strategy;

[0043] Based on the preset topological structure and control requirements of the power system, the global control domain is divided into control sub-areas, and the global optimal control strategy is decoupled into multiple local sub-strategies, each of which corresponds to a control sub-area;

[0044] Based on the global optimal control strategy, the local sub-strategy corresponding to the control sub-region is optimized and adjusted through the deep deterministic policy gradient algorithm to generate the local optimal sub-strategy;

[0045] The deep deterministic policy gradient algorithm includes:

[0046] Based on the state input of the state space, output the action distribution of the action space, construct a policy network, and based on the evaluation value corresponding to the action distribution and the state input, construct a value function network;

[0047] Initializing the policy network and the value function network, and setting initial exploration noise;

[0048] Based on the current state, an action distribution is generated through the policy network, a deterministic action is sampled from the action distribution, an exploratory action is generated in combination with the exploratory noise, and a next state and reward obtained after executing the exploratory action are calculated;

[0049] The current state, exploratory action, reward and next state are taken as a set of state transition samples and stored in an experience replay buffer;

[0050] Sampling a set of state transition samples from the experience replay buffer, and calculating the loss function of the value function network and the policy network;

[0051] Optimizing the loss function, updating the value function network parameters and the strategy network parameters, and adjusting the intensity of the exploration noise;

[0052] Repeat the training steps until the preset number of iterations is reached, and finally the optimal control strategy is obtained.

[0053] In an optional embodiment,

[0054] Through the federated learning algorithm, the distributed collaborative control solutions include:

[0055] Each control sub-area is used as a blockchain node. The blockchain node locally trains and generates a local sub-strategy through the deep deterministic policy gradient algorithm to obtain local policy parameters, and uploads the local policy parameters to the blockchain network to generate a parameter upload event.

[0056] In the blockchain network, a smart aggregation contract is set to monitor the parameter upload event, record the local policy parameters, trigger aggregation calculation of multiple local policy parameters, generate global policy parameters, and broadcast them to the blockchain network according to a preset broadcast cycle;

[0057] Each of the control sub-areas synchronizes the global policy parameters from the blockchain network, updates the local sub-policies, and iterates the training to generate local sub-policies;

[0058] The iteration is repeated until the preset upper limit of the number of learning times is reached, and the local sub-strategies corresponding to all the control sub-areas constitute a distributed collaborative control solution.

[0059] According to a second aspect of the embodiments of the present invention,

[0060] Provided is an intelligent monitoring system for power system operation status, comprising:

[0061] The first unit is used to obtain multi-source heterogeneous operation data of the power system, perform standardized preprocessing on the multi-source heterogeneous operation data to obtain a standard operation data set, input the standard operation data set into a heterogeneous data feature extraction model, adaptively extract high-order multimodal feature representations through a confidence weighted fusion method, obtain an operation state multimodal feature set, and construct an intelligent monitoring cognitive map by learning topological association features and semantic association rules of the operation state;

[0062] The second unit is used to build an operation state graph learning model based on the intelligent monitoring cognitive graph, map the real-time operation data to the intelligent monitoring cognitive graph, dynamically perceive the spatiotemporal evolution law of the operation state through the spatiotemporal dependency modeling unit and the graph feature propagation unit in the operation state graph learning model, and generate prediction and diagnosis results;

[0063] The third unit is used to generate the optimal control strategy of the power system based on the prediction and diagnosis results through the deep deterministic policy gradient algorithm learning, decouple the optimal control strategy of the power system into multiple local sub-strategies, and send them to the corresponding power control equipment, obtain the distributed collaborative control scheme through the federated learning algorithm, and display the control process corresponding to the distributed collaborative control scheme in real time through the visualization equipment.

[0064] According to a third aspect of the embodiments of the present invention,

[0065] An electronic device is provided, comprising:

[0066] processor;

[0067] a memory for storing processor-executable instructions;

[0068] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0069] A fourth aspect of the embodiments of the present invention is:

[0070] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0071] In the embodiment of the present invention, by preprocessing the standardized multi-source heterogeneous data, the format and scale differences between different data sources are solved, and the consistency and comparability of the data are improved; by using the confidence weighted fusion method, the high-order feature representation of the multimodal data can be adaptively extracted, and the ability to capture complex operating states is enhanced; by learning the topological association features and semantic association rules of the operating state, the intelligent monitoring cognitive map constructed provides a comprehensive representation of the operating state of the power system, which is conducive to more accurate monitoring and diagnosis; by using the spatiotemporal dependency modeling unit, the temporal and spatial dependencies of the power system are accurately captured, and the model's ability to understand complex operating states is improved; By propagating and updating the features of the system status through the graph feature propagation unit, accurate prediction and diagnosis results can be generated, which improves the prediction accuracy of the power system operating status and the timeliness of fault diagnosis; the optimal control strategy generated by deep deterministic policy gradient algorithm learning can improve the operating efficiency and stability of the power system and avoid the suboptimal solutions that may be brought about by traditional methods; the overall control strategy is decoupled into multiple local sub-strategies to adapt to the characteristics and requirements of different equipment, and improve the accuracy and response speed of control; the federated learning algorithm is used to realize distributed collaborative control to ensure that all devices can work together in a decentralized manner, further enhancing the stability and robustness of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 A schematic diagram of a flow chart of a method for intelligently monitoring the operating status of a power system according to an embodiment of the present invention;

[0073] Figure 2 It is a structural diagram of an intelligent monitoring system for the operation status of a power system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0074] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution 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 only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0075] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0076] Figure 1 FIG. 1 is a flow chart of an intelligent monitoring method for power system operation status according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0077] S101. Acquire multi-source heterogeneous operation data of the power system, perform standardization preprocessing on the multi-source heterogeneous operation data to obtain a standard operation data set, input the standard operation data set into a heterogeneous data feature extraction model, adaptively extract high-order multimodal feature representations through a confidence weighted fusion method, obtain a multimodal feature set of the operation state, and construct an intelligent monitoring cognitive map by learning the topological association features and semantic association rules of the operation state;

[0078] In this embodiment, by standardizing and preprocessing multi-source heterogeneous data, the format and scale differences between different data sources are resolved, and the consistency and comparability of the data are improved; by using the confidence weighted fusion method, the high-order feature representation of multimodal data can be adaptively extracted, and the ability to capture complex operating conditions is enhanced; by learning the topological association features and semantic association rules of the operating conditions, the constructed intelligent monitoring cognitive graph provides a comprehensive representation of the operating conditions of the power system, which is conducive to more accurate monitoring and diagnosis; the constructed intelligent monitoring cognitive graph integrates the high-order features of multi-source heterogeneous data, making the intelligent monitoring of the power system more accurate and reliable.

[0079] In an optional embodiment, multi-source heterogeneous operation data of the power system is obtained, and the multi-source heterogeneous operation data is standardized and preprocessed to obtain a standard operation data set, and the standard operation data set is input into a heterogeneous data feature extraction model, and a high-order multimodal feature representation is adaptively extracted through a confidence weighted fusion method to obtain an operation state multimodal feature set including:

[0080] Acquire multi-source heterogeneous operation data of the power system, wherein the multi-source heterogeneous operation data includes real-time operation parameters, historical operation data, weather data, geographic information data, and equipment status data;

[0081] Correct the multi-source heterogeneous operation data through outlier detection, missing value filling and duplicate value deletion, unify the scale and distribution of the multi-source heterogeneous operation data through maximum and minimum value normalization and Z-score normalization, and obtain a standard operation data set;

[0082] Inputting the standard operation data set into a heterogeneous data feature extraction model, wherein the heterogeneous data feature extraction model includes a multi-layer deep feature extraction network and an adversarial network;

[0083] Based on the dimension corresponding to the standard running data set, a first neural network node group is set to construct an input observation layer, according to a preset feature quantity threshold, a second neural network node group is set to construct a potential capture layer, an interaction is constructed between the input observation layer and the potential capture layer, according to a preset shielding probability value, some nodes in the first neural network node group and the second neural network node group are randomly shielded, a hierarchical feature extraction neural network is determined, the hierarchical feature extraction neural network is stacked in multiple layers to construct a multi-layer deep feature extraction network, based on a contrast divergence algorithm, the corresponding weight parameters of each layer of the hierarchical feature extraction neural network are iteratively adjusted, and a high-order multimodal feature representation is output;

[0084] Inputting the high-order multimodal feature representation into an adversarial network, and capturing the multimodal feature distribution through game learning between a generator and a discriminator in the adversarial network;

[0085] Based on the multimodal feature distribution, combined with the confidence weighted fusion method, the confidence of different modal features is determined, and the weights of different modal features are calculated through the softmax function. Adaptive fusion is performed through weighted averaging to obtain a multimodal feature set of the operating state.

[0086] Acquire multi-source heterogeneous operation data of the power system, including real-time operation parameters, historical operation data, weather data, geographic information data, equipment status data and other data from multiple sources and types.

[0087] Preprocess and clean the acquired multi-source heterogeneous operation data. Use the outlier detection algorithm to identify outliers in the data and process them, such as deleting or replacing them; fill missing values ​​in the data by using methods such as mean filling, median filling or interpolation; delete duplicate values ​​in the data to ensure the uniqueness of the data; and correct multi-source heterogeneous operation data.

[0088] In order to unify the scale and distribution of multi-source heterogeneous operation data, the data normalization method is used. For numerical features, the maximum and minimum value normalization is used to scale them to the interval [0, 1]. For features with normal distribution, Z-score normalization is used to convert them to a standard normal distribution with a mean of 0 and a variance of 1. After normalization, the standard operation data set is obtained.

[0089] The standard running data set is input into the heterogeneous data feature extraction model, which is composed of a multi-layer deep feature extraction network and an adversarial network.

[0090] Construct a multi-layer deep feature extraction network. According to the dimension of the standard running data set, set the same number of neural network nodes as the data dimension to form the input observation layer. According to the preset feature number threshold, set fewer neural network nodes than the input observation layer to form the potential capture layer. Construct a fully connected interaction between the input observation layer and the potential capture layer. In order to improve the robustness and generalization ability of feature extraction, according to the preset masking probability value, randomly mask some nodes in the input observation layer and the potential capture layer to form a random masking mechanism. Through the above steps, determine the single-layer hierarchical feature extraction neural network structure. Stack multiple hierarchical feature extraction neural networks to construct a multi-layer deep feature extraction network. Using the contrastive divergence algorithm, through unsupervised learning, iteratively adjust the weight parameters of each layer of the neural network layer by layer to extract the high-order multimodal feature representation of the input data.

[0091] The high-order multimodal feature representation output by the multi-layer deep feature extraction network is input into the adversarial network. The adversarial network consists of two parts: the generator and the discriminator. Through the game learning between the two, the true distribution of multimodal features is captured. The generator receives random noise and generates synthetic features similar to the distribution of real features. The discriminator receives real features and synthetic features and tries to distinguish them. Through repeated games between the generator and the discriminator, the parameters of the two are continuously optimized, so that the synthetic features generated by the generator are closer and closer to the distribution of real features, while improving the ability of the discriminator to distinguish between real and synthetic features. After the game converges, the adversarial network learns the true distribution of multimodal features.

[0092] Based on the multimodal feature distribution learned by the adversarial network, the confidence weighted fusion method is used to adaptively fuse the features of different modalities. First, the confidence score of each modal feature is calculated based on the error of reconstruction in the adversarial network. The smaller the error, the higher the confidence. Then, the confidence score is converted into a weight coefficient through the softmax function, and the sum of the weights is 1. Finally, the features of different modalities are weighted averaged according to the weight coefficients to obtain the fused multimodal feature set of the operating state. The modal features with higher weights dominate the fusion, while the modal features with lower weights contribute less, thus achieving adaptive feature fusion.

[0093] In this embodiment, through preprocessing steps such as outlier detection, missing value filling, and duplicate value deletion, the quality and reliability of multi-source heterogeneous data are improved, and the accuracy and consistency of the data are ensured; a multi-layer deep feature extraction network and a random shielding mechanism are adopted, so that the model can effectively capture high-order multimodal features, and enhance the robustness and generalization ability of feature extraction; through the game learning between the generator and the discriminator of the adversarial network, the model can capture and simulate the true distribution of multimodal features, and improve the authenticity and effectiveness of feature representation; based on the confidence weighted fusion method, different modal features are adaptively fused, ensuring the dominant role of key features with higher weights in the fusion process, thereby improving the representativeness and accuracy of the final multimodal feature set; the fused multimodal feature set finally output provides a comprehensive and efficient feature representation for subsequent power system monitoring and diagnosis tasks, and improves the prediction and anomaly detection capabilities of the system.

[0094] In an optional embodiment, the training of the heterogeneous data feature extraction model includes:

[0095] Based on the heterogeneous data feature extraction model, the number of nodes in the second neural network node group, the number of layers in the multi-layer stack, and the corresponding model hyperparameters are used as model configuration to determine the solution space;

[0096] In the solution space, individuals are randomly initialized to determine an initial population, a fitness value corresponding to each individual in the initial population is calculated, and based on the fitness value, a first individual corresponding to an optimal solution, a second individual corresponding to a suboptimal solution, a third individual corresponding to a third optimal solution, and a remaining common individual group are determined;

[0097] Based on the first individual, the second individual and the third individual, randomly determine basic offspring individuals according to a preset neighborhood range, perform crossover and mutation operations on the basic offspring individuals, and generate a preferred offspring individual group;

[0098] Based on the common individual group, a plurality of individuals are randomly selected, and crossover and mutation operations are performed to generate common offspring and add them to the common individual group, and based on the positions of the first individual, the second individual, and the third individual, the position of each individual in the common individual group is updated according to the individual position update formula;

[0099] The first individual, the second individual, the third individual, the preferred offspring individual group and the updated common individual group constitute an iterative population;

[0100] Calculating the fitness value corresponding to each individual in the iterative population, and re-determining the first individual, the second individual, the third individual and the common individual group;

[0101] Repeat the iterative update until the preset number of iterations is reached, and the first individual finally obtained becomes the optimal individual, determining the optimal structural data feature extraction model.

[0102] Based on the heterogeneous data feature extraction model, its optimized solution space is determined. The number of nodes in the second neural network node group (i.e., the potential capture layer), the number of layers in the multi-layer stack, and related model hyperparameters are used as variables to be optimized to form a solution space. The solution space defines all possible configuration combinations of the heterogeneous data feature extraction model.

[0103] In the solution space, a group of initial individuals are randomly generated to form an initial population. Each individual represents a possible configuration of the heterogeneous data feature extraction model. For each individual in the initial population, its corresponding model configuration is applied to the heterogeneous data feature extraction model, and the performance indicators of the model, such as reconstruction error or classification accuracy, are calculated on the validation data set as the fitness value of the individual. The fitness value measures the quality of the corresponding model configuration of the individual.

[0104] The initial population is sorted according to the fitness value, and the individual with the best performance is determined as the first individual, the individual with the second best performance is determined as the second individual, and the individual with the third best performance is determined as the third individual. The remaining individuals are divided into the common individual group.

[0105] Based on the first individual, the second individual, and the third individual, a group of basic offspring individuals are randomly generated in the preset neighborhood range near them. These basic offspring individuals are crossover and mutation operations are performed to generate a group of preferred offspring individuals. The crossover operation generates new individuals by exchanging some genes of two individuals; the mutation operation introduces new search directions by randomly changing some genes of individuals. The group of preferred offspring individuals inherits the genes of excellent individuals and introduces new diversity.

[0106] Multiple individuals are randomly selected from the common individual group, and crossover and mutation operations are performed on them to generate common offspring individuals, which are then added to the common individual group. At the same time, the position of each individual in the common individual group is updated according to the position of the first individual, the second individual, and the third individual in the solution space. The update process is implemented through the individual position update formula to guide the common individuals to move in a better direction.

[0107] The first individual, the second individual, the third individual, the preferred offspring individual group and the updated common individual group are merged to form a new iterative population. The fitness value of each individual in the iterative population is calculated, and the first individual, the second individual, the third individual and the common individual group are re-determined.

[0108] Repeat the above iterative update process until the preset number of iterations is reached or the convergence condition is met. During the iteration process, the population continues to evolve, the individuals continue to optimize, and gradually approach the optimal solution. When the iteration ends, the final first individual is the optimal individual searched, and its corresponding model configuration determines the optimal heterogeneous data feature extraction model.

[0109] In this embodiment, by constructing and optimizing the solution space, various configuration combinations of the heterogeneous data feature extraction model are systematically explored to ensure that the model configuration with the best performance is found; using the idea of ​​evolutionary algorithm, the fitness value of the individual is gradually improved through the initial population, crossover and mutation operations, to achieve efficient solution space search and quickly approach the optimal solution; the offspring individuals are selected to inherit excellent genes, and new diversity is introduced through mutation, so as to maintain the balance between exploration and development in the optimization process and prevent falling into the local optimum; the position and genes of the ordinary individual group are gradually approached by excellent individuals through continuous iterative updates, and the optimization process has strong adaptability and dynamic adjustment capabilities; through multiple rounds of iterations, the population continues to evolve and gradually approaches the global optimal solution, ensuring that the configuration of the heterogeneous data feature extraction model reaches the optimal state, thereby improving the performance and adaptability of the model.

[0110] In an optional embodiment, the individual position update formula is as follows:

[0111]

[0112] Among them, X1 represents the position update of the ordinary individual under the guidance of the first individual, X2 represents the position update of the ordinary individual under the guidance of the second individual, X3 represents the position update of the ordinary individual under the guidance of the third individual, Xα represents the position of the first individual, X β represents the second individual position, X δ represents the position of the third individual, r1 represents the first random vector, r2 represents the second random vector, r3 represents the third random vector, k represents the iteration control parameter, C1 represents the scaling factor vector with the first individual position, C2 represents the scaling factor vector with the second individual position, C3 represents the scaling factor vector with the third individual position, t represents the current number of iterations, X(t) represents the position of the ordinary individual at the tth iteration, X(t+1) represents the position of the ordinary individual at the t+1th iteration, T represents the maximum number of iterations, ω1 represents the influence weight of the first individual, ω2 represents the influence weight of the second individual, and ω3 represents the influence weight of the third individual.

[0113] In this algorithm, the position update process of an ordinary individual consists of three steps, and each step considers the influence of three individuals with better performance, namely the first individual, the second individual, and the third individual.

[0114] The position of the normal individual is adjusted based on the position of the first individual. The adjustment is controlled by a random factor, which is determined by calculating the scale of the current relative position of the first individual and the normal individual. Specifically, the position of the normal individual will move closer to the first individual, and the amplitude is determined by a random vector, which is regenerated at each iteration, increasing randomness and search diversity.

[0115] Similar to the first step, the ordinary individual will also make adjustments based on the positions of the second and third individuals.

[0116] The three new positions calculated separately are combined by weighted average to form the new position of the common individual in the next iteration. The weight coefficient is determined by the weights of the three influencing individuals, and the weight reflects the contribution ratio of each influencing individual to the final position of the common individual.

[0117] An iterative control parameter is introduced in this algorithm. As the number of iterations increases, the parameter gradually decreases, so that the adjustment range during individual updates gradually decreases, controlling the convergence speed of the algorithm.

[0118] In this embodiment, the positions of ordinary individuals are updated by guiding the first individual, the second individual and the third individual respectively, so that the information of the optimal individual is fully utilized, making it more likely that the ordinary individuals will move towards the direction of the global optimal solution; an iterative control parameter is introduced, which gradually decreases with the increase of the number of iterations, helping the algorithm to conduct extensive exploration in the early stage and gradually converge to a better solution in the later stage; random vectors are used to introduce randomness, increase the ability to explore the solution space, and effectively avoid the algorithm from falling into the local optimum; combined with the positions of multiple individuals and the role of the scaling factor vector, ordinary individuals can approach the global optimal solution more quickly, thereby accelerating the convergence of the optimization process.

[0119] S102. Based on the intelligent monitoring cognitive map, an operation state graph learning model is constructed, and real-time operation data is mapped to the intelligent monitoring cognitive map. Through the spatiotemporal dependency modeling unit and the graph feature propagation unit in the operation state graph learning model, the spatiotemporal evolution law of the operation state is dynamically perceived to generate prediction and diagnosis results;

[0120] In this embodiment, real-time operation data is mapped to the intelligent monitoring cognitive graph, which can dynamically perceive the spatiotemporal evolution of the power system and provide real-time status monitoring and analysis. The spatiotemporal dependency modeling unit is used to accurately capture the temporal and spatial dependencies of the power system, thereby improving the model's ability to understand complex operating states. The feature propagation unit is used to propagate and update the system state, thereby generating accurate prediction and diagnosis results, improving the accuracy of prediction of the power system operation state and the timeliness of fault diagnosis. The prediction and diagnosis results generated by the model provide intelligent decision support for the operation and maintenance of the power system, which helps to prevent faults in advance and optimize system operation.

[0121] In an optional embodiment, based on the intelligent monitoring cognitive graph, an operation state graph learning model is constructed, and real-time operation data is mapped to the intelligent monitoring cognitive graph. Through the spatiotemporal dependency modeling unit and the graph feature propagation unit in the operation state graph learning model, the spatiotemporal evolution law of the operation state is dynamically perceived, and the prediction and diagnosis results are generated, including:

[0122] Mapping the real-time operation data to the intelligent monitoring cognitive graph, representing the nodes in the intelligent monitoring cognitive graph as low-dimensional dense vectors through node embedding, and generating an operation status feature matrix;

[0123] Input the running state feature matrix into the gated recurrent unit, learn the state time evolution pattern of the node by calculating the update gate, reset gate and candidate state, and obtain the node hidden state sequence; based on each node, dot product the hidden state corresponding to the node with the hidden state of the neighbor node, calculate the similarity between the node and the neighbor node, determine the attention weight through normalization operation, and sum the hidden state corresponding to the neighbor node according to the attention weight to obtain the node spatial context feature; splice the node hidden state sequence with the node spatial context feature, integrate the spatiotemporal dependency, and generate the node state representation;

[0124] Based on the node state representation, combined with the causal semantics between nodes, the causal attention between nodes is calculated to construct a causal attention matrix; the causal attention matrix is ​​used as an adjacency matrix, the information of neighbor nodes is aggregated through the adjacency matrix, and feature transformation is performed in combination with a parameter matrix and a nonlinear function. After multiple layers of graph convolution are stacked, the interaction between nodes is captured, and the node state representation is spatially updated;

[0125] The updated node state representation is input into the prediction and diagnosis module, and the node state distribution at the future moment is predicted through the sequence-to-sequence model. The abnormality measurement of each node state representation in the node state distribution is performed to determine the abnormal nodes and associated paths.

[0126] A loss function including state prediction loss and abnormal diagnosis loss is constructed. The parameters of the operation state graph learning model are trained through the back propagation algorithm and the gradient descent optimization method. The optimal operation state graph learning model is generated by minimizing the loss function to obtain the prediction and diagnosis results.

[0127] The real-time operation data of the power system is mapped to the intelligent monitoring cognitive graph, so that the operation data is associated with the nodes in the graph. Through node embedding technology, each node in the graph is represented as a low-dimensional dense vector, which captures the semantic information and contextual relationship of the node. Based on the node embedding vector, a feature matrix representing the current operation status is generated, with each row corresponding to the feature vector of a node.

[0128] The running state feature matrix is ​​input into the gated recurrent unit for time series modeling. The gated recurrent unit learns the time evolution pattern of the node state by calculating the update gate, reset gate and candidate state. The update gate controls the degree to which the state information of the previous moment is retained to the current moment, the reset gate controls the degree to which the state information of the previous moment is ignored, and the candidate state represents the new state information at the current moment. Through the recursive calculation of the gated recurrent unit, the hidden state sequence of each node is obtained to capture the time dependency of the node state.

[0129] For each node, consider its spatial dependency. Calculate the dot product between the hidden state of the node and the hidden state of its neighboring nodes to obtain the similarity between the node and the neighboring nodes. By normalizing the similarity, the attention weight of the node to the neighboring nodes is obtained. The hidden states of the neighboring nodes are weighted and summed according to the attention weights to obtain the spatial context features of the node, which indicates the degree of influence of the node by the neighboring nodes.

[0130] The hidden state sequence of the node is concatenated with its spatial context features, and the temporal and spatial dependencies of the node state are integrated to generate a comprehensive node state representation that takes into account both the state evolution of the node itself and the interaction between the node and its neighboring nodes.

[0131] Based on the node state representation and combined with the causal semantic information between nodes in the intelligent monitoring cognitive graph, the causal attention between nodes is calculated. Causal attention reflects the causal dependency between nodes, that is, the degree of influence of a node state change on another node state. Through causal attention calculation, a causal attention matrix is ​​constructed to represent the causal dependency strength between nodes.

[0132] Using the causal attention matrix as the adjacency matrix, the graph convolutional network is used to spatially update the node state representation. The graph convolutional network aggregates the information of neighbor nodes through the adjacency matrix, and combines the parameter matrix and nonlinear activation function to transform the features. By stacking multiple layers of graph convolutional networks, high-order interactions and spatial propagation patterns between nodes are captured. The output of the graph convolutional network is the updated node state representation, which integrates the graph topology and the spatial dependency of the node state.

[0133] The updated node state representation is input into the prediction and diagnosis module for state prediction and anomaly diagnosis. The sequence-to-sequence model is used to take the node state representation at the current moment as input to predict the node state distribution at multiple moments in the future. The abnormal nodes are identified by performing anomaly measurement on the predicted node state distribution. The anomaly measurement can be based on the degree of deviation of the node state representation from the normal state, or based on indicators such as reconstruction error. At the same time, by analyzing the association paths between abnormal nodes, the path of anomaly propagation and the scope of influence are located.

[0134] Construct a loss function including state prediction loss and abnormal diagnosis loss, and perform end-to-end training on the running state graph learning model. The state prediction loss measures the difference between the predicted node state distribution and the true state distribution, and the abnormal diagnosis loss measures the model's ability to identify and locate abnormal nodes. The gradient of the loss function to the model parameters is calculated through the back propagation algorithm, and the model parameters are updated using the gradient descent optimization method to minimize the loss function. After multiple rounds of iterative training, the optimal running state graph learning model is obtained.

[0135] The trained optimal operating state graph learning model is used to predict and diagnose the real-time operating state of the power system. The model can predict the node state distribution at future moments, identify abnormal nodes and abnormal propagation paths, and provide decision support for the safe and stable operation of the power system. Through the application of the graph learning model, real-time monitoring, early warning and diagnosis of the operating state are realized, which improves the intelligence level of the power system and its ability to deal with complex abnormal situations.

[0136] In this embodiment, through node embedding and time series modeling, the time evolution and spatial dependency of the operating state of the power system are accurately captured to achieve accurate monitoring of the operating state of the power system; the time dependency and spatial context features of the node are fused to generate a comprehensive node state representation, so that the model can not only capture the changes of the node itself, but also consider the interaction between the node and the neighboring nodes; based on the node state representation and causal semantic information, the causal attention between nodes is calculated, and the causal dependency between nodes is identified, which helps to understand the causal logic of node state changes; through the graph convolutional network and sequence-to-sequence model, the model can identify abnormal nodes and their propagation paths, provide efficient abnormal diagnosis capabilities, and enhance the ability of the power system to cope with abnormal situations; by constructing a loss function including state prediction loss and abnormal diagnosis loss, the entire model is trained end-to-end to ensure the prediction accuracy and diagnostic capability of the model; the trained model can be applied to real-time monitoring and diagnosis of the power system, provide intelligent decision support, and ensure the safe and stable operation of the power system.

[0137] S103. Based on the prediction and diagnosis results, the optimal control strategy of the power system is generated through deep deterministic policy gradient algorithm learning, the optimal control strategy of the power system is decoupled into multiple local sub-strategies, and sent to the corresponding power control equipment, and a distributed collaborative control scheme is obtained through a federated learning algorithm. The control process corresponding to the distributed collaborative control scheme is displayed in real time through a visualization device.

[0138] Based on the prediction and diagnosis results of the power system, a deep deterministic policy gradient algorithm is used to learn and generate the optimal control strategy, aiming to optimize the operating efficiency and stability of the power system; the generated optimal control strategy is decoupled into multiple local sub-strategies, which correspond to different power control devices, making the control more refined and decentralized; through the federated learning algorithm, distributed collaborative control is achieved among multiple power control devices to ensure that each device can work together to jointly optimize the operation of the entire power system; using visualization equipment, the execution process of the distributed collaborative control scheme is displayed in real time, allowing operators to intuitively monitor and adjust the system control strategy.

[0139] In this embodiment, the optimal control strategy generated by deep deterministic policy gradient algorithm learning can improve the operating efficiency and stability of the power system and avoid suboptimal solutions that may be caused by traditional methods; decouple the overall control strategy into multiple local sub-strategies to adapt to the characteristics and needs of different devices, and improve the accuracy and response speed of control; use federated learning algorithms to achieve distributed collaborative control, ensure that each device can work together while operating in a decentralized manner, and further enhance the stability and robustness of the system; use visualization equipment to display the execution process of the control scheme in real time, help operators to promptly discover and adjust potential problems, and improve the safety and reliability of the power system.

[0140] In an optional embodiment, based on the prediction and diagnosis results, the optimal control strategy of the power system is generated by learning through a deep deterministic policy gradient algorithm, and the optimal control strategy of the power system is decoupled into multiple local sub-strategies, including:

[0141] Based on the prediction and diagnosis results, determining the optimal control problem, constructing a state space and an action space, wherein the state space includes a current state, a future state in the prediction and diagnosis results, and an abnormal state, and the action space includes a power system control variable corresponding to the control strategy;

[0142] Based on the deep deterministic policy gradient algorithm, the optimal control problem is solved to generate a global optimal control strategy;

[0143] Based on the preset topological structure and control requirements of the power system, the global control domain is divided into control sub-areas, and the global optimal control strategy is decoupled into multiple local sub-strategies, each of which corresponds to a control sub-area;

[0144] Based on the global optimal control strategy, the local sub-strategy corresponding to the control sub-region is optimized and adjusted through the deep deterministic policy gradient algorithm to generate the local optimal sub-strategy;

[0145] The deep deterministic policy gradient algorithm includes:

[0146] Based on the state input of the state space, output the action distribution of the action space, construct a policy network, and based on the evaluation value corresponding to the action distribution and the state input, construct a value function network;

[0147] Initializing the policy network and the value function network, and setting initial exploration noise;

[0148] Based on the current state, an action distribution is generated through the policy network, a deterministic action is sampled from the action distribution, an exploratory action is generated in combination with the exploratory noise, and a next state and reward obtained after executing the exploratory action are calculated;

[0149] The current state, exploratory action, reward and next state are taken as a set of state transition samples and stored in an experience replay buffer;

[0150] Sampling a set of state transition samples from the experience replay buffer, and calculating the loss function of the value function network and the policy network;

[0151] Optimizing the loss function, updating the value function network parameters and the strategy network parameters, and adjusting the intensity of the exploration noise;

[0152] Repeat the training steps until the preset number of iterations is reached, and finally the optimal control strategy is obtained.

[0153] Based on the prediction and diagnosis results, the optimal control problem of the power system is determined. The state space and action space are constructed, where the state space includes the current state, the future state in the prediction and diagnosis results, and the identified abnormal state, and the action space includes the power system control variables corresponding to the optional control strategies, such as generator output, load control, etc.

[0154] The deep deterministic policy gradient algorithm is used to solve the optimal control problem and generate the global optimal control policy. The deep deterministic policy gradient algorithm maximizes the long-term cumulative reward by learning the optimal policy.

[0155] Considering the topological structure and control requirements of the power system, the global control domain is divided into multiple control sub-regions. Each control sub-region corresponds to a local control problem. The global optimal control strategy is decoupled into multiple local sub-strategies, and each local sub-strategy performs optimal control for a control sub-region.

[0156] Based on the global optimal control strategy, the deep deterministic policy gradient algorithm is used to optimize and adjust the local sub-strategy of each control sub-region. Through the optimization of the local strategy, the local optimal sub-strategy that adapts to the characteristics of the control sub-region is generated. The local optimal sub-strategy works in conjunction with the global optimal control strategy to achieve hierarchical distributed optimization control of the power system.

[0157] The specific steps of the deep deterministic policy gradient algorithm are as follows:

[0158] Construct a policy network and a value function network. The policy network takes the state in the state space as input and outputs the action distribution in the action space. The value function network takes the state and action as input and evaluates the value of the state-action pair.

[0159] Initialize the parameters of the policy network and the value function network, and set the initial exploration noise. The exploration noise is used to introduce exploration and help the algorithm escape from the local optimum.

[0160] Based on the current state, an action distribution is generated through the policy network. Deterministic actions are sampled from the action distribution, and exploratory noise is added to generate exploratory actions. The exploratory actions are applied to the power system to obtain the next state and reward value. The reward value measures the degree of improvement in system performance after executing the action.

[0161] The current state, exploratory action, reward, and next state are stored as a set of state transition samples in the experience replay buffer. The experience replay buffer is used to store multiple sets of state transition samples for algorithm training.

[0162] A set of state transition samples are randomly sampled from the experience replay buffer, and the loss functions of the value function network and the policy network are calculated. The loss function of the value function network measures the difference between the predicted state-action value and the actual value, and the loss function of the policy network measures the quality of the policy.

[0163] Optimize the loss function and update the parameters of the value function network and the policy network. Use optimization algorithms such as gradient descent to adjust network parameters to minimize the loss function. At the same time, adjust the intensity of exploration noise to balance exploration and utilization.

[0164] Repeat the above training steps and continuously iterate and optimize until the preset number of iterations or convergence conditions are reached. Finally, the trained optimal control strategy is obtained, including the global optimal control strategy and the local optimal sub-strategy.

[0165] The optimal control strategy obtained through training is applied to the actual control of the power system. According to the current state of the system and the prediction and diagnosis results, the corresponding control strategy is selected, the optimal control instructions are generated, the operating state of the power system is adjusted, and safe, stable and economical optimal control is achieved.

[0166] In this embodiment, by using a deep deterministic policy gradient algorithm, the global optimal control strategy of the power system can be effectively generated, and further decoupled into local sub-strategies to ensure the accuracy and global coordination of the control strategy; the global control strategy is decoupled into multiple local sub-strategies, and optimized in combination with the characteristics of each sub-region, thereby realizing hierarchical distributed control of the power system and enhancing the flexibility and adaptability of the system; through continuous iteration and optimization, the deep deterministic policy gradient algorithm can continuously improve the control strategy, enhance the response speed and stability of the power system under different states, and reduce possible fluctuations and anomalies; in the optimization process, exploration noise is introduced to help the algorithm jump out of the local optimal solution, and at the same time, by gradually adjusting the noise intensity, the exploration and utilization are balanced to ensure that it finally converges to the optimal control strategy.

[0167] In an optional embodiment, obtaining a distributed collaborative control solution through a federated learning algorithm includes:

[0168] Each control sub-area is used as a blockchain node. The blockchain node locally trains and generates a local sub-strategy through the deep deterministic policy gradient algorithm to obtain local policy parameters, and uploads the local policy parameters to the blockchain network to generate a parameter upload event.

[0169] In the blockchain network, a smart aggregation contract is set to monitor the parameter upload event, record the local policy parameters, trigger aggregation calculation of multiple local policy parameters, generate global policy parameters, and broadcast them to the blockchain network according to a preset broadcast cycle;

[0170] Each of the control sub-areas synchronizes the global policy parameters from the blockchain network, updates the local sub-policies, and iterates the training to generate local sub-policies;

[0171] The iteration is repeated until the preset upper limit of the number of learning times is reached, and the local sub-strategies corresponding to all the control sub-areas constitute a distributed collaborative control solution.

[0172] Each control sub-area of ​​the power system is used as a blockchain node to form a blockchain network. Each control sub-area corresponds to a participant.

[0173] Participants use the deep deterministic policy gradient algorithm to train and generate local sub-policies locally. Through iterative training of the deep deterministic policy gradient algorithm, the optimized local policy parameters such as the weights and biases of the policy network and value function network are obtained.

[0174] Participants upload the local strategy parameters obtained through training to the blockchain network. The upload process triggers a parameter upload event, with the local strategy parameters as the data content of the event.

[0175] A smart aggregation contract is deployed in the blockchain network to aggregate local policy parameters uploaded by multiple participants. The smart aggregation contract listens to parameter upload events, obtains the uploaded local policy parameters, and records them in the state of the blockchain.

[0176] When the smart aggregation contract collects enough local policy parameters, the aggregation calculation process is triggered. Aggregation calculation combines multiple local policy parameters to generate global policy parameters. Aggregation calculation can adopt weighted average, voting mechanism and other methods to comprehensively consider the contribution of each local policy.

[0177] The smart aggregation contract broadcasts the generated global policy parameters to the blockchain network according to the preset broadcast cycle. The broadcast process packages the global policy parameters into blocks and reaches consensus in the blockchain network through the consensus mechanism to ensure the consistency and immutability of the global policy parameters.

[0178] Each control sub-region acts as a blockchain node and synchronously obtains the broadcasted global policy parameters from the blockchain network. Each control sub-region merges the global policy parameters with the local sub-policy and updates the parameters of the local sub-policy.

[0179] The updated local sub-policy continues to be iteratively trained locally, and the local policy parameters are optimized using the deep deterministic policy gradient algorithm. Through the guidance and constraints of the global policy parameters, the training process of the local sub-policy can take into account global information and the policies of other control sub-regions to achieve collaborative optimization of the policy.

[0180] Repeat the above process of parameter uploading, aggregation calculation, global parameter broadcasting, local strategy updating and iterative training, and continuously iterate and optimize until the preset upper limit of learning times is reached or the convergence conditions are met.

[0181] After multiple rounds of iterative optimization, the local sub-strategies of all control sub-areas formed a distributed collaborative control scheme, which fully utilized the decentralized, transparent, trustworthy and tamper-proof characteristics of blockchain to achieve secure distributed learning and execution of control strategies.

[0182] In the actual control process, each control sub-region generates corresponding control instructions based on the optimized local sub-strategy, combined with local state information and control requirements, and performs distributed collaborative control of the power system. Through the collaborative execution of local sub-strategies, the global optimization control goal is achieved.

[0183] In this embodiment, the local strategy parameters of each control sub-area are aggregated and integrated through the blockchain network, and the strategies of each sub-area are collaboratively optimized by using smart contracts and consensus mechanisms to ensure the consistency and effectiveness of the global control strategy. The decentralized, tamper-proof, transparent and reliable characteristics of the blockchain are used to ensure the security and consistency of the uploaded, aggregated and broadcasted global strategy parameters to prevent tampering or loss of strategy data. Each control sub-area locally updates and optimizes its local strategy using the global strategy parameters, and autonomously generates control instructions during the actual control process, thereby enhancing the flexibility and adaptability of distributed control of the power system. Through continuous iterative updates and training in the blockchain network, the local strategies of each control sub-area can continuously adapt to new control requirements and environmental changes, thereby achieving continuous optimization of the power system and long-term stability of the global control objectives.

[0184] Figure 2 FIG. 1 is a schematic diagram of the structure of the intelligent monitoring system for the operation status of a power system according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0185] The first unit is used to obtain multi-source heterogeneous operation data of the power system, perform standardized preprocessing on the multi-source heterogeneous operation data to obtain a standard operation data set, input the standard operation data set into a heterogeneous data feature extraction model, adaptively extract high-order multimodal feature representations through a confidence weighted fusion method, obtain an operation state multimodal feature set, and construct an intelligent monitoring cognitive map by learning topological association features and semantic association rules of the operation state;

[0186] The second unit is used to build an operation state graph learning model based on the intelligent monitoring cognitive graph, map the real-time operation data to the intelligent monitoring cognitive graph, dynamically perceive the spatiotemporal evolution law of the operation state through the spatiotemporal dependency modeling unit and the graph feature propagation unit in the operation state graph learning model, and generate prediction and diagnosis results;

[0187] The third unit is used to generate the optimal control strategy of the power system based on the prediction and diagnosis results through the deep deterministic policy gradient algorithm learning, decouple the optimal control strategy of the power system into multiple local sub-strategies, and send them to the corresponding power control equipment, obtain the distributed collaborative control scheme through the federated learning algorithm, and display the control process corresponding to the distributed collaborative control scheme in real time through the visualization equipment.

[0188] According to a third aspect of the embodiments of the present invention,

[0189] An electronic device is provided, comprising:

[0190] processor;

[0191] a memory for storing processor-executable instructions;

[0192] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0193] A fourth aspect of the embodiments of the present invention is:

[0194] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0195] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0196] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent monitoring method for the operation status of a power system, characterized in that: include: Acquire multi-source heterogeneous operation data of the power system, perform standardized preprocessing on the multi-source heterogeneous operation data to obtain a standard operation data set, input the standard operation data set into a heterogeneous data feature extraction model, adaptively extract high-order multimodal feature representations through a confidence weighted fusion method, obtain a multimodal feature set of the operation state, and construct an intelligent monitoring cognitive map by learning the topological association features and semantic association rules of the operation state; Based on the intelligent monitoring cognitive graph, an operation state graph learning model is constructed to map the real-time operation data to the intelligent monitoring cognitive graph. Through the spatiotemporal dependency modeling unit and the graph feature propagation unit in the operation state graph learning model, the spatiotemporal evolution law of the operation state is dynamically perceived to generate prediction and diagnosis results. Based on the prediction and diagnosis results, the optimal control strategy of the power system is generated through deep deterministic policy gradient algorithm learning, and the optimal control strategy of the power system is decoupled into multiple local sub-strategies. The strategies are sent to the corresponding power control equipment, and the distributed collaborative control scheme is obtained through the federated learning algorithm. The control process corresponding to the distributed collaborative control scheme is displayed in real time through the visualization equipment.

2. The method according to claim 1, characterized in that Acquire multi-source heterogeneous operation data of the power system, perform standardized preprocessing on the multi-source heterogeneous operation data to obtain a standard operation data set, input the standard operation data set into a heterogeneous data feature extraction model, and adaptively extract high-order multimodal feature representations through a confidence weighted fusion method to obtain an operation state multimodal feature set including: Acquire multi-source heterogeneous operation data of the power system, wherein the multi-source heterogeneous operation data includes real-time operation parameters, historical operation data, weather data, geographic information data, and equipment status data; Correct the multi-source heterogeneous operation data through outlier detection, missing value filling and duplicate value deletion, unify the scale and distribution of the multi-source heterogeneous operation data through maximum and minimum value normalization and Z-score normalization, and obtain a standard operation data set; Inputting the standard operation data set into a heterogeneous data feature extraction model, wherein the heterogeneous data feature extraction model includes a multi-layer deep feature extraction network and an adversarial network; Based on the dimension corresponding to the standard running data set, a first neural network node group is set to construct an input observation layer, according to a preset feature quantity threshold, a second neural network node group is set to construct a potential capture layer, an interaction is constructed between the input observation layer and the potential capture layer, according to a preset shielding probability value, some nodes in the first neural network node group and the second neural network node group are randomly shielded, a hierarchical feature extraction neural network is determined, the hierarchical feature extraction neural network is stacked in multiple layers to construct a multi-layer deep feature extraction network, based on a contrast divergence algorithm, the corresponding weight parameters of each layer of the hierarchical feature extraction neural network are iteratively adjusted, and a high-order multimodal feature representation is output; Inputting the high-order multimodal feature representation into an adversarial network, and capturing the multimodal feature distribution through game learning between a generator and a discriminator in the adversarial network; Based on the multimodal feature distribution, combined with the confidence weighted fusion method, the confidence of different modal features is determined, and the weights of different modal features are calculated through the softmax function. Adaptive fusion is performed through weighted averaging to obtain a multimodal feature set of the operating state.

3. The method according to claim 2, characterized in that The training of the heterogeneous data feature extraction model includes: Based on the heterogeneous data feature extraction model, the number of nodes in the second neural network node group, the number of layers in the multi-layer stack, and the corresponding model hyperparameters are used as model configuration to determine the solution space; In the solution space, individuals are randomly initialized to determine an initial population, a fitness value corresponding to each individual in the initial population is calculated, and based on the fitness value, a first individual corresponding to an optimal solution, a second individual corresponding to a suboptimal solution, a third individual corresponding to a third optimal solution, and a remaining common individual group are determined; Based on the first individual, the second individual and the third individual, randomly determine basic offspring individuals according to a preset neighborhood range, perform crossover and mutation operations on the basic offspring individuals, and generate a preferred offspring individual group; Based on the common individual group, a plurality of individuals are randomly selected, and crossover and mutation operations are performed to generate common offspring and add them to the common individual group, and based on the positions of the first individual, the second individual, and the third individual, the position of each individual in the common individual group is updated according to the individual position update formula; The first individual, the second individual, the third individual, the preferred offspring individual group and the updated common individual group constitute an iterative population; Calculating the fitness value corresponding to each individual in the iterative population, and re-determining the first individual, the second individual, the third individual and the common individual group; Repeat the iterative update until the preset number of iterations is reached, and the first individual finally obtained becomes the optimal individual, determining the optimal structural data feature extraction model.

4. The method according to claim 3, characterized in that The individual position update formula is as follows: Among them, X1 represents the position update of the ordinary individual under the guidance of the first individual, X2 represents the position update of the ordinary individual under the guidance of the second individual, X3 represents the position update of the ordinary individual under the guidance of the third individual, Xα represents the position of the first individual, X β represents the second individual position, X δ represents the position of the third individual, r1 represents the first random vector, r2 represents the second random vector, r3 represents the third random vector, k represents the iteration control parameter, C1 represents the scaling factor vector with the first individual position, C2 represents the scaling factor vector with the second individual position, C3 represents the scaling factor vector with the third individual position, t represents the current number of iterations, X(t) represents the position of the ordinary individual at the tth iteration, X(t+1) represents the position of the ordinary individual at the t+1th iteration, T represents the maximum number of iterations, ω1 represents the influence weight of the first individual, ω2 represents the influence weight of the second individual, and ω3 represents the influence weight of the third individual.

5. The method according to claim 1, characterized in that Based on the intelligent monitoring cognitive graph, an operation state graph learning model is constructed, and real-time operation data is mapped to the intelligent monitoring cognitive graph. Through the spatiotemporal dependency modeling unit and graph feature propagation unit in the operation state graph learning model, the spatiotemporal evolution law of the operation state is dynamically perceived, and prediction and diagnosis results are generated, including: Mapping the real-time operation data to the intelligent monitoring cognitive graph, representing the nodes in the intelligent monitoring cognitive graph as low-dimensional dense vectors through node embedding, and generating an operation status feature matrix; Input the running state feature matrix into the gated recurrent unit, learn the state time evolution pattern of the node by calculating the update gate, reset gate and candidate state, and obtain the node hidden state sequence; based on each node, dot product the hidden state corresponding to the node with the hidden state of the neighbor node, calculate the similarity between the node and the neighbor node, determine the attention weight through normalization operation, and sum the hidden state corresponding to the neighbor node according to the attention weight to obtain the node spatial context feature; splice the node hidden state sequence with the node spatial context feature, integrate the spatiotemporal dependency, and generate the node state representation; Based on the node state representation, combined with the causal semantics between nodes, the causal attention between nodes is calculated to construct a causal attention matrix; the causal attention matrix is ​​used as an adjacency matrix, the information of neighbor nodes is aggregated through the adjacency matrix, and feature transformation is performed in combination with a parameter matrix and a nonlinear function. After multiple layers of graph convolution are stacked, the interaction between nodes is captured, and the node state representation is spatially updated; The updated node state representation is input into the prediction and diagnosis module, and the node state distribution at the future moment is predicted through the sequence-to-sequence model. The abnormality measurement of each node state representation in the node state distribution is performed to determine the abnormal nodes and associated paths. A loss function including state prediction loss and abnormal diagnosis loss is constructed. The parameters of the operation state graph learning model are trained through the back propagation algorithm and the gradient descent optimization method. The optimal operation state graph learning model is generated by minimizing the loss function to obtain the prediction and diagnosis results.

6. The method according to claim 1, characterized in that Based on the prediction and diagnosis results, the optimal control strategy of the power system is generated by deep deterministic policy gradient algorithm learning, and the optimal control strategy of the power system is decoupled into multiple local sub-strategies including: Based on the prediction and diagnosis results, determining the optimal control problem, constructing a state space and an action space, wherein the state space includes a current state, a future state in the prediction and diagnosis results, and an abnormal state, and the action space includes a power system control variable corresponding to the control strategy; Based on the deep deterministic policy gradient algorithm, the optimal control problem is solved to generate a global optimal control strategy; Based on the preset topological structure and control requirements of the power system, the global control domain is divided into control sub-areas, and the global optimal control strategy is decoupled into multiple local sub-strategies, each of which corresponds to a control sub-area; Based on the global optimal control strategy, the local sub-strategy corresponding to the control sub-region is optimized and adjusted through the deep deterministic policy gradient algorithm to generate the local optimal sub-strategy; The deep deterministic policy gradient algorithm includes: Based on the state input of the state space, output the action distribution of the action space, construct a policy network, and based on the evaluation value corresponding to the action distribution and the state input, construct a value function network; Initializing the policy network and the value function network, and setting initial exploration noise; Based on the current state, an action distribution is generated through the policy network, a deterministic action is sampled from the action distribution, an exploratory action is generated in combination with the exploratory noise, and a next state and reward obtained after executing the exploratory action are calculated; The current state, exploratory action, reward and next state are taken as a set of state transition samples and stored in an experience replay buffer; Sampling a set of state transition samples from the experience replay buffer, and calculating the loss function of the value function network and the policy network; Optimizing the loss function, updating the value function network parameters and the strategy network parameters, and adjusting the intensity of the exploration noise; Repeat the training steps until the preset number of iterations is reached, and finally the optimal control strategy is obtained.

7. The method according to claim 6, characterized in that Through the federated learning algorithm, the distributed collaborative control solutions include: Each control sub-area is used as a blockchain node. The blockchain node locally trains and generates a local sub-strategy through the deep deterministic policy gradient algorithm to obtain local policy parameters, and uploads the local policy parameters to the blockchain network to generate a parameter upload event. In the blockchain network, a smart aggregation contract is set to monitor the parameter upload event, record the local policy parameters, trigger aggregation calculation of multiple local policy parameters, generate global policy parameters, and broadcast them to the blockchain network according to a preset broadcast cycle; Each of the control sub-areas synchronizes the global policy parameters from the blockchain network, updates the local sub-policies, and iterates the training to generate local sub-policies; The iteration is repeated until the preset upper limit of the number of learning times is reached, and the local sub-strategies corresponding to all the control sub-areas constitute a distributed collaborative control solution.

8. An intelligent monitoring system for the operation status of a power system, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain multi-source heterogeneous operation data of the power system, perform standardized preprocessing on the multi-source heterogeneous operation data to obtain a standard operation data set, input the standard operation data set into a heterogeneous data feature extraction model, adaptively extract high-order multimodal feature representations through a confidence weighted fusion method, obtain an operation state multimodal feature set, and construct an intelligent monitoring cognitive map by learning topological association features and semantic association rules of the operation state; The second unit is used to build an operation state graph learning model based on the intelligent monitoring cognitive graph, map the real-time operation data to the intelligent monitoring cognitive graph, dynamically perceive the spatiotemporal evolution law of the operation state through the spatiotemporal dependency modeling unit and the graph feature propagation unit in the operation state graph learning model, and generate prediction and diagnosis results; The third unit is used to generate the optimal control strategy of the power system based on the prediction and diagnosis results through the deep deterministic policy gradient algorithm learning, decouple the optimal control strategy of the power system into multiple local sub-strategies, and send them to the corresponding power control equipment, obtain the distributed collaborative control scheme through the federated learning algorithm, and display the control process corresponding to the distributed collaborative control scheme in real time through the visualization equipment.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images