Power system energy supply prediction method and system based on transfer learning

Abstract

The present invention provides a method and system for predicting energy supply of an electric power system based on transfer learning, the method comprising the following steps: caching regional real-time power data of a risky power supply area; obtaining regional power supply topology data and regional historical power operation data of a target area; extracting power operation timing features by combining a multi-scale convolutional neural network and transfer learning technology; generating a regional power map network of a target area based on regional power supply topology data, and intercepting a regional power map network of a risky power supply area according to regional power supply topology data; extracting power topology structure features by combining a regional power map network and a regional power map network and using a graph neural network; fusing power operation timing features and power topology structure features into comprehensive power features, and outputting a power supply gap prediction result of a risky power supply area through a GCN‑LSTM prediction model. The present invention has the effect of accurately predicting the power supply gap.

Description

Technical Field

[0001] The present invention belongs to the technical field of power management, and specifically relates to a method and system for predicting energy supply of a power system based on transfer learning. Background Art

[0002] With the development of social economy and the continuous growth of electricity demand, the scale and complexity of power systems are also increasing, and power dispatching systems are facing more and more challenges. At the same time, with the widespread application of smart grids and renewable energy, the types and quantities of data in power systems have increased dramatically, which provides rich data resources for power dispatching systems, but also puts forward higher data processing and analysis requirements.

[0003] The power system itself is highly complex and uncertain. Emergencies such as natural disasters and equipment failures will have a significant impact on the power supply system. Emergencies in the fault area will cause a power supply gap in the fault area. At this time, it is necessary to dispatch the power of the entire power system in a timely manner to make up for the power supply gap in the fault area while minimizing the impact on the power supply of all other areas. This requires that the size of the power supply gap can be accurately predicted when an emergency occurs. However, most traditional prediction methods for power systems rely on a single data source, such as historical electricity consumption data or simple real-time monitoring data, which cannot fully reflect the dynamic changes and complexity of the power system. Therefore, they lack sufficient flexibility and adaptability, resulting in the inability to adjust and respond to these emergencies in a timely manner. On the other hand, existing statistical models and simple prediction algorithms are difficult to capture the complex time series characteristics and mutual influences between regions in the power system, resulting in insufficient accuracy in the prediction results of the power supply gap. Summary of the invention

[0004] The present invention provides a method and system for predicting energy supply of an electric power system based on transfer learning, so as to solve the problem of insufficient accuracy of prediction results of power supply gap.

[0005] In a first aspect, the present invention provides a method for predicting energy supply of a power system based on transfer learning, which is applied to a power dispatching system in a target area, wherein the target area includes a plurality of different power supply areas, and comprises the following steps:

[0006] In response to the regional power supply risk information received by the power dispatching system, a risk power supply area in the target area where the emergency power supply risk occurs is determined, and starting from the determination of the risk power supply area, regional real-time power data of the risk power supply area within a preset time period is cached, wherein the regional real-time power data includes regional power supply topology data and regional real-time power operation data;

[0007] Acquiring regional power supply topology data and regional historical power operation data of the target area through the power dispatching system;

[0008] Preprocessing the real-time power operation data of the region and the historical power operation data of the region, and extracting power operation time series features from the preprocessed real-time power operation data of the region and the historical power operation data of the region by combining a multi-scale convolutional neural network and transfer learning technology;

[0009] generating a regional power map network of the target area based on the regional power supply topology data, and extracting a regional power map network of the risk power supply area from the regional power map network according to the regional power supply topology data;

[0010] Combining the regional power graph network with the area power graph network and using a graph neural network to extract power topology features;

[0011] The power operation timing characteristics and the power topology structure characteristics are integrated into a comprehensive power characteristic, and the comprehensive power characteristic is input into a preset GCN-LSTM prediction model, and the power supply gap prediction result of the risk power supply area is output through the GCN-LSTM prediction model.

[0012] Optionally, the preprocessing of the regional real-time power operation data and the regional historical power operation data, and extracting power operation time series features from the preprocessed regional real-time power operation data and the regional historical power operation data by combining a multi-scale convolutional neural network and transfer learning technology comprises the following steps:

[0013] preprocessing the regional real-time power operation data based on the grey prediction model theory to obtain preprocessed regional real-time power time series data;

[0014] Preprocessing the historical power operation data of the region by using a fuzzy entropy algorithm to obtain preprocessed historical power time series data of the region;

[0015] Based on the real-time power time series data of the region and the historical power time series data of the region, the power operation time series characteristics are extracted by using multi-scale convolutional neural network and transfer learning technology.

[0016] Optionally, the preprocessing of the regional real-time power operation data based on the grey prediction model theory to obtain the preprocessed regional real-time power time series data comprises the following steps:

[0017] constructing a regional real-time power operation matrix based on the regional real-time power operation data;

[0018] Performing a cumulative calculation on the real-time power operation data of the region to obtain a cumulative sequence data;

[0019] Based on the grey prediction model theory and according to the once accumulated sequence data, obtaining the once accumulated adjacent sequence data;

[0020] Constructing a primary accumulated neighbor sequence matrix according to the primary accumulated neighbor sequence data;

[0021] The grey coefficient of the grey prediction model is calculated by combining the regional real-time power operation matrix and the primary accumulated neighboring sequence matrix;

[0022] Using the grey coefficient to construct the differential equation of the grey prediction model and the corresponding system response function;

[0023] Performing cumulative subtraction calculation on the once accumulated sequence data by using the system response function to obtain restored power operation data;

[0024] The restored power operation data is subjected to residual inspection and residual correction processing to obtain regional real-time power time series data.

[0025] Optionally, the extracting of power operation time series features based on the regional real-time power time series data and the regional historical power time series data by using a multi-scale convolutional neural network and transfer learning technology comprises the following steps:

[0026] Inputting the regional real-time power time series data into a preset local feature extractor, performing local feature extraction through the local feature extractor, and obtaining regional real-time local time series features;

[0027] Inputting the regional historical power time series data into a preset global feature extractor, performing global feature extraction through the global feature extractor, obtaining initial regional historical global time series features and updating all extractor parameters in the global feature extractor;

[0028] Initialize the fully connected layer parameters of the last fully connected layer in the global feature extractor, and freeze all other extractor parameters in the global feature extractor;

[0029] Inputting the regional real-time power time series data into the global feature extractor, using the regional real-time power time series data to implement transfer learning of the global feature extractor, and updating the fully connected layer parameters of the last fully connected layer to obtain the global feature extractor after transfer learning;

[0030] Extracting regional historical global time series features from the regional historical power time series data by the global feature extractor after transfer learning;

[0031] The real-time local time series characteristics of the region and the historical global time series characteristics of the region are fused into the multi-scale time series characteristics of electric power by using preset fusion parameters;

[0032] The convolution layers of the multi-scale convolutional neural network are configured according to the characteristic scale of the multi-scale time series features of the electric power, the multi-scale time series features of the electric power are input into the configured multi-scale convolutional neural network, and the electric power operation time series features are output after hierarchical transfer learning processing of the multi-scale convolutional neural network.

[0033] Optionally, configuring the convolutional layers of the multi-scale convolutional neural network according to the characteristic scale of the multi-scale time series features of the electric power, inputting the multi-scale time series features of the electric power into the configured multi-scale convolutional neural network, and outputting the electric power operation time series features after hierarchical transfer learning processing of the multi-scale convolutional neural network comprises the following steps:

[0034] Configure a common convolution layer of a multi-scale convolutional neural network based on the average characteristic scale of all time series features of the multi-scale time series features of the power;

[0035] Sorting all the time series features in the power multi-scale time series features in descending order of feature scale;

[0036] Based on the number of scale types of the characteristic scale and according to the sorting result of the temporal features, sequentially configuring a plurality of standard convolutional layers of different scales in the multi-scale convolutional neural network, wherein the number of the standard convolutional layers is the same as the number of scale types;

[0037] According to the sorting result, the first-ranked time series feature is used as the target time series feature, the target time series feature is input into the common convolution layer and the target standard convolution layer of the corresponding scale, and the common parameters of the common convolution layer and the target standard parameters of the target standard convolution layer are updated and retained;

[0038] According to the sorting result, the next time series feature is used as the target time series feature in turn, the target time series feature is input into the multi-scale convolutional neural network, and a parameter updating step is performed based on a layered transfer learning method until the standard parameters of each standard convolutional layer are updated;

[0039] The optimized timing features output by each of the standard convolutional layers are integrated into the power operation timing features.

[0040] Optionally, the step of inputting the target time series features into the multi-scale convolutional neural network and performing parameter updating based on a layered transfer learning method comprises the following steps:

[0041] Inputting the target time series features into the common convolution layer, outputting common extraction features based on the current common parameters, and updating and retaining the common parameters;

[0042] Inputting the target time series features into a target standard convolution layer of a corresponding scale, loading an upper layer standard parameter of the standard convolution layer of a previous scale into the target standard convolution layer, and outputting a standard extraction feature based on the upper layer standard parameter;

[0043] Calculating feature similarity between the common extracted features and the standard extracted features;

[0044] If the feature similarity is less than a preset similarity threshold, the standard parameters in the target standard convolution layer are reset, the target time series feature is re-input into the target standard convolution layer, and the target time series feature is used to update and retain the standard parameters of the target standard convolution layer;

[0045] If the feature similarity is greater than or equal to the similarity threshold, the upper layer standard parameters are retained as the standard parameters of the target standard convolution layer.

[0046] Optionally, the method further comprises the following steps:

[0047] Inputting the regional real-time local time series features into a preset ARMA model, and outputting a first preliminary prediction result through the ARMA model;

[0048] Inputting the power operation time series characteristics into a preset LSTM model, and outputting a second initial prediction result through the LSTM model;

[0049] Calculating the difference of the initial prediction results by combining the first initial prediction result and the second initial prediction result;

[0050] Based on the initial prediction result difference, the fusion parameters and the model parameters of the ARMA model and the LSTM model are adjusted through a parameter adjustment algorithm.

[0051] Optionally, generating a regional power map network of the target area based on the regional power supply topology data, and extracting a regional power map network of the risk power supply area from the regional power map network according to the regional power supply topology data comprises the following steps:

[0052] Traversing the regional power supply topology data, extracting all power supply nodes and power consumption nodes in the regional power supply topology data, and identifying the supply relationship between all the power supply nodes and all the power consumption nodes;

[0053] Combining all the power supply nodes, all the power consumption nodes and the supply relationship to construct a regional power map network of the target area;

[0054] Traversing the regional power supply topology data, and extracting all regional edge power supply nodes and regional edge power consumption nodes according to node attributes of topological nodes in the regional power supply topology data;

[0055] The regional power map network of the risk power supply area is determined and intercepted in the regional power map network by using all the regional edge power supply nodes and all the regional edge power consumption nodes.

[0056] Optionally, combining the regional power graph network and the area power graph network and extracting power topology structure features using a graph neural network comprises the following steps:

[0057] Inputting the regional power graph network into a preset graph neural network, and using the graph convolution layer and the graph pooling layer in the graph neural network to sequentially perform convolution processing and pooling processing on the regional power graph network to obtain initial topological features;

[0058] Inputting the initial topological features into a graph fully connected layer in the graph neural network, and updating graph fully connected parameters in the graph fully connected layer;

[0059] Inputting the regional power graph network into the graph fully connected layer and fine-tuning the graph fully connected parameters;

[0060] The initial topological features are re-input into the graph fully connected layer after parameter fine-tuning, and the power topological structure features are output through the graph fully connected layer after parameter fine-tuning.

[0061] In the second aspect, the present invention also provides an electric power system energy supply prediction system based on transfer learning, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the electric power system energy supply prediction method based on transfer learning as described in the first aspect is implemented.

[0062] The beneficial effects of the present invention are:

[0063] By responding to the regional power supply risk information received by the power dispatching system, the risk power supply area with emergency power supply risk in the target area can be determined in time, and the real-time power data of the area can be cached within the preset time period, providing a detailed data basis for subsequent analysis. Then the power supply topology data and historical power operation data of the target area are obtained, and these data are preprocessed. The power operation time series features extracted by combining multi-scale convolutional neural network and transfer learning technology can more accurately reflect the dynamic changes of the power system. Combining the regional power graph network and the regional power graph network, the power topology structure features extracted by graph neural network are helpful to fully understand the structural characteristics of the power system. Finally, the power operation time series features and power topology structure features are fused into power comprehensive features and input into the preset GCN-LSTM prediction model. The power supply gap prediction results of the risk power supply area are output through this model, which can significantly improve the accuracy and response speed of the prediction, so as to effectively evaluate and respond to power supply risks and ensure the stable operation of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 This is a flowchart of the steps of a method for predicting energy supply of a power system based on transfer learning in one embodiment of the present application.

[0065] Figure 2 This is a framework flow chart of a method for predicting energy supply of a power system based on transfer learning in one embodiment of the present application.

[0066] Figure 3 A framework flow chart for extracting power operation timing characteristics in one embodiment of the present application.

[0067] Figure 4 This is a framework flow chart for adjusting fusion parameters in one implementation of the present application.

[0068] Figure 5 A framework flow chart for extracting power topology features in one embodiment of the present application. DETAILED DESCRIPTION

[0069] The following will be combined with the drawings in the embodiments of the present application to clearly describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments in the present application belong to the scope of protection of this application.

[0070] The terms "first", "second", etc. in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described here, and the objects distinguished by "first", "second", etc. are generally of one type, and the number of objects is not limited. For example, the first object can be one or more. In addition, "and / or" in the specification and claims represents at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0071] Figure 1 FIG. 1 is a flow chart of a method for predicting power system energy supply based on transfer learning in one embodiment. It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps. The power system energy supply prediction method based on transfer learning disclosed in the present invention is applied to the power dispatching system of the target area, and the target area includes multiple different power supply areas. Figure 1 As shown, the power system energy supply prediction method based on transfer learning disclosed in the present invention specifically includes the following steps:

[0072] S101. In response to regional power supply risk information received by the power dispatching system, determine the risk power supply area in the target area where emergency power supply risk occurs, and cache regional real-time power data of the risk power supply area within a preset time period starting from the determination of the risk power supply area.

[0073] Among them, regional real-time power data includes regional power supply topology data and regional real-time power operation data. Once the power dispatching system detects that there is power supply risk information in a certain area, the system will immediately analyze the source and content of the information to determine which areas may be affected. This process may involve a comprehensive analysis of multiple factors such as power load and equipment failure. After determining the risk power supply area, the system will begin to cache the real-time power data of the area within a preset time period. Real-time power data includes regional power supply topology data and regional real-time power operation data. Power supply topology data describes the structure of the power network, such as the connection relationship between substations, transmission lines and loads, while real-time power operation data includes dynamic operating parameters such as voltage, current, and power. By caching this data, detailed input information can be provided for subsequent analysis and prediction, ensuring that a quick and accurate response can be made in emergency situations.

[0074] S102. Obtain regional power supply topology data and regional historical power operation data of the target area through the power dispatching system.

[0075] Among them, the power supply topology data describes the structure of the entire power network, including the specific locations of each substation, transmission line and load and their interconnection relationship. The historical power operation data records the power operation status of the region over a period of time in the past, including changes in parameters such as voltage, current, and power. These data can be obtained through the database of the power dispatching system, which usually involves a lot of data processing and cleaning work to ensure the quality and consistency of the data. By obtaining these data, the necessary basic information can be provided for subsequent time series feature extraction and topology structure analysis.

[0076] S103. Pre-process the regional real-time power operation data and the regional historical power operation data, and extract the power operation time series characteristics from the pre-processed regional real-time power operation data and the regional historical power operation data by combining the multi-scale convolutional neural network and transfer learning technology.

[0077] Among them, refer to Figure 2The basic preprocessing process includes data cleaning, normalization, noise reduction and other links. Data cleaning refers to removing outliers and missing values ​​in the data to ensure the integrity and accuracy of the data. Normalization is to convert the data to the same dimension to facilitate subsequent analysis and processing. Noise reduction is to remove random noise in the data through filtering and other technologies to improve the signal-to-noise ratio of the data. The preprocessed data will be input into the multi-scale convolutional neural network, which can extract multi-level time series features from the data, while the transfer learning technology can use existing models and knowledge to improve the efficiency and accuracy of feature extraction. By combining these two technologies, the power operation time series features can be extracted from the preprocessed regional real-time power operation data and the regional historical power operation data, providing important input information for subsequent predictions.

[0078] S104. Generate a regional power map network of the target area based on the regional power supply topology data, and extract the regional power map network of the risky power supply area from the regional power map network according to the regional power supply topology data.

[0079] Among them, the regional power graph network is a graphical representation of the entire power system structure, with nodes representing substations (power supply equipment) or loads (power consumption equipment), and edges representing transmission lines. The process of generating a power graph network includes reading the power supply topology data and building a graph structure according to the relationship between nodes and edges. Subsequently, based on the determined risk power supply area, the corresponding regional power graph network is extracted from the regional power graph network. This process requires precise matching of nodes and edges within the region to ensure that the extracted subgraph can accurately reflect the power supply structure of the risk area.

[0080] S105. Combine the regional power graph network and the regional power graph network and use graph neural network to extract power topology structure features.

[0081] Among them, the graph neural network is a deep learning model that can process graph structure data and can effectively extract node features and edge features in the graph. In the specific implementation, the regional power graph network and the regional power graph network are first input into the graph neural network, and the structural features in the graph are extracted layer by layer through multi-layer graph convolution operations. These features include the connection relationship between nodes, the degree distribution of nodes, the structural pattern of subgraphs, etc. In this way, the topological structural characteristics of the power system can be fully captured, and potential structural risks and weak links can be identified.

[0082] S106. The power operation timing characteristics and the power topology structure characteristics are integrated into the power comprehensive characteristics, and the power comprehensive characteristics are input into the preset GCN-LSTM prediction model, and the power supply gap prediction results of the risk power supply area are output through the GCN-LSTM prediction model.

[0083] Among them, the GCN-LSTM prediction model combines the advantages of graph convolutional networks (GCN) and long short-term memory networks (LSTM), and can process graph structure data and time series data at the same time. Figure 2 As shown in the figure, the extracted power operation time series features and power topology structure features are first fused to generate comprehensive power features. Then, the comprehensive power features are input into the GCN-LSTM model. The GCN part is responsible for processing the graph structure information and extracting the features of nodes and edges, while the LSTM part is responsible for processing the time series information and capturing the time dependency in the data. In this way, the model can comprehensively consider the structural characteristics and dynamic operation characteristics of the power system and output the prediction results of the power supply gap in the risk power supply area. This result can provide important decision-making support for power dispatching, help take timely countermeasures, and ensure the stable operation of the power system.

[0084] In one implementation manner, step S103 specifically includes the following steps:

[0085] Preprocess the regional real-time power operation data based on the grey prediction model theory to obtain the preprocessed regional real-time power time series data;

[0086] The fuzzy entropy algorithm is used to preprocess the regional historical power operation data to obtain the preprocessed regional historical power time series data;

[0087] Based on regional real-time power time series data and regional historical power time series data, the power operation time series characteristics are extracted by using multi-scale convolutional neural network and transfer learning technology.

[0088] In this embodiment, the grey system theory is a method for dealing with uncertainty and incomplete information, which is particularly suitable for small samples and poor information problems. In the specific implementation, the GM(1,1) model in the grey prediction model theory is applied to preprocess these data. The GM(1,1) model is a first-order univariate grey prediction model, which generates a sequence by accumulating the original data, then establishes a differential equation model, and finally obtains the predicted value by inverse accumulation and restoration. In this way, random fluctuations and noise in the data can be effectively removed, and the smoothness and stability of the data can be improved, thereby obtaining the preprocessed regional real-time power time series data.

[0089] The fuzzy entropy algorithm is a nonlinear analysis method for measuring the complexity of time series, which can effectively handle the ambiguity and uncertainty in the data. In the specific implementation, the fuzzy entropy algorithm needs to be applied to preprocess the regional historical power operation data. The specific steps of the fuzzy entropy algorithm include: first, the time series is divided into multiple subsequences, each subsequence has a length of m; then the distance between each subsequence is calculated, and the distance can be calculated using Euclidean distance or other distance measurement methods; then, the fuzzy membership is calculated according to the distance, and the fuzzy membership indicates the similarity between the two subsequences; the fuzzy entropy value is calculated by the fuzzy membership, and the higher the fuzzy entropy value, the greater the complexity of the time series; finally, based on the fuzzy entropy value and through the fuzzy entropy algorithm, the components of the fuzzy value approximation are superimposed and merged to obtain the regional historical power time series data, thereby reducing the computational complexity. In this way, random noise and outliers in the data can be effectively removed, the smoothness and stability of the data can be improved, and the complexity of the data can be reduced.

[0090] The pre-processed regional real-time power time series data and regional historical power time series data are integrated to form a comprehensive time series data set. Next, the multi-scale convolutional neural network (MSCNN) is used to extract features from the comprehensive time series data set. MSCNN extracts multi-level features in the data through convolution kernels of different scales, and can capture local patterns and global trends in the data. In the specific implementation, the comprehensive time series data is first input into the MSCNN, and after multiple layers of convolution operations, feature maps of different scales are extracted; then, the feature maps are reduced in dimension through the pooling layer, the main features are retained, and redundant information is removed. Next, the model parameters are adjusted using transfer learning technology to improve the efficiency and accuracy of feature extraction. Ultimately, the extracted power operation time series features will provide important input information for subsequent prediction models, significantly improving the accuracy and reliability of predictions.

[0091] In one embodiment, the step of preprocessing the regional real-time power operation data based on the grey prediction model theory to obtain the preprocessed regional real-time power time series data specifically includes the following steps:

[0092] Construct a regional real-time power operation matrix based on regional real-time power operation data;

[0093] Perform a cumulative calculation on the regional real-time power operation data to obtain a cumulative sequence data;

[0094] Based on the grey prediction model theory and according to the once accumulated sequence data, the once accumulated adjacent sequence data is obtained;

[0095] Constructing a cumulative neighbor sequence matrix according to the cumulative neighbor sequence data;

[0096] The grey coefficient of the grey prediction model is calculated by combining the regional real-time power operation matrix and the one-time cumulative neighbor sequence matrix.

[0097] Use grey coefficients to construct the differential equation of grey prediction model and the corresponding system response function;

[0098] The system response function is used to perform cumulative subtraction calculation on the primary accumulated sequence data to obtain the restored power operation data;

[0099] The restored power operation data is subjected to residual inspection and residual correction processing to obtain regional real-time power time series data.

[0100] In this embodiment, constructing a regional real-time power operation matrix based on the regional real-time power operation data is the first step in the application of the entire gray prediction model. Assuming that there are N time points and M parameters, the regional real-time power operation data can be represented as a matrix of N rows and M columns. Each row represents the power operation status at a time point, and each column represents a specific power parameter. The regional real-time power operation data is accumulated and calculated once. The purpose of the accumulation calculation is to smooth out the random fluctuations in the data and make the data sequence more regular. In the specific implementation, each column of data in the regional real-time power operation matrix is ​​accumulated and calculated. Assume that the original data sequence is , the cumulative sequence is ,in Through one cumulative calculation, the noise and random components in the data can be effectively reduced, making the data sequence smoother and more stable, thus laying the foundation for the subsequent grey prediction model.

[0101] Based on the theory of grey prediction model and obtaining the data of the first cumulative adjacent sequence according to the first cumulative sequence data is a key step in grey system theory. Grey prediction model, especially GM (1,1) model, realizes the prediction of original data by modeling the cumulative sequence. In the specific implementation, it is necessary to calculate the adjacent sequence of the first cumulative sequence first. Assume that the first cumulative sequence is , the cumulative adjacent sequence is ,in .

[0102] Organize the accumulated adjacent sequence data into a matrix form. Assume that the accumulated adjacent sequence is , the constructed matrix is ​​recorded as , where each line is an element of the accumulated adjacent sequence. The gray coefficient is estimated using the least squares method. Suppose the regional real-time power operation matrix is , the cumulative adjacent sequence matrix is , the gray coefficient vector is recorded as . Solve the gray coefficient vector by the least squares method , so that the sum of squared errors is minimized. The specific calculation formula is: .

[0103] Using the gray coefficient to construct the differential equation of the gray prediction model and the corresponding system response function is an important step in realizing gray prediction. In the specific implementation, the gray coefficient is used to construct the differential equation of the GM (1,1) model. Assume that the gray coefficient is , a cumulative sequence is , the differential equation is ,in is a constant term. By solving the differential equation, we get the system response function. The system response function is in the form of ,in is the initial value.

[0104] Then use the system response function to perform cumulative subtraction calculation on the primary cumulative sequence. Suppose the system response function is , restore the power operation data to The cumulative calculation formula is: . Then it is necessary to perform residual inspection and residual correction on the restored power operation data. First, perform residual inspection to calculate the residual between the restored data and the original data. Then perform residual correction to reduce the residual by adjusting model parameters or introducing compensation items. Specific correction methods include adjusting the gray coefficient or introducing adaptive correction items. Finally, more accurate regional real-time power time series data can be obtained through residual inspection and correction processing. These data will provide a reliable basis for real-time monitoring and prediction of power operation.

[0105] In one embodiment, the step of extracting power operation time series features based on regional real-time power time series data and regional historical power time series data by using multi-scale convolutional neural network and transfer learning technology specifically includes the following steps:

[0106] Inputting the regional real-time power time series data into a preset local feature extractor, extracting local features through the local feature extractor, and obtaining regional real-time local time series features;

[0107] Input the regional historical power time series data into a preset global feature extractor, perform global feature extraction through the global feature extractor, obtain the initial regional historical global time series features and update all extractor parameters in the global feature extractor;

[0108] Initialize the fully connected layer parameters of the last fully connected layer in the global feature extractor and freeze all other extractor parameters in the global feature extractor;

[0109] The regional real-time power time series data is input into the global feature extractor, and the regional real-time power time series data is used to realize the transfer learning of the global feature extractor, and the fully connected layer parameters of the last fully connected layer are updated to obtain the global feature extractor after transfer learning;

[0110] The regional historical global time series features are extracted from the regional historical power time series data through the global feature extractor after transfer learning;

[0111] The preset fusion parameters are used to fuse the regional real-time local time series characteristics and the regional historical global time series characteristics into the multi-scale time series characteristics of power;

[0112] The convolution layers of the multi-scale convolutional neural network are configured according to the characteristic scale of the multi-scale time series characteristics of electric power, the multi-scale time series characteristics of electric power are input into the configured multi-scale convolutional neural network, and the electric power operation time series characteristics are output after hierarchical transfer learning processing of the multi-scale convolutional neural network.

[0113] In this embodiment, refer to Figure 3 , the regional real-time power time series data is input into the preset local feature extractor, and the local feature extractor is used to extract local features to obtain the process of regional real-time local time series features. First, it is necessary to clarify the structure and content of the regional real-time power time series data. The preset local feature extractor can use a convolutional neural network (CNN), which is designed to use convolution layers and pooling layers to capture local patterns in the data. In specific implementation, the time series data is organized into a two-dimensional matrix form, each row represents a time point, and each column represents a power parameter. When the input data passes through the first convolution layer, the convolution kernel slides on the data to extract the features of the local area, such as the voltage change trend within a certain period of time. Then, the pooling layer will perform dimensionality reduction processing on the features output by the convolution layer, retain the main features and reduce the amount of data. After multiple layers of convolution and pooling operations, the final feature matrix contains local features in the regional real-time power time series data, such as short-term fluctuations and mutation points of power parameters. These local features provide detailed information for understanding the immediate state of the regional power system.

[0114] The process of inputting the regional historical power time series data into the preset global feature extractor, extracting global features through the global feature extractor, obtaining the initial regional historical global time series features and updating all extractor parameters in the global feature extractor usually adopts the long short-term memory network (LSTM) or other recurrent neural networks (RNN). These networks are good at capturing long-term dependencies in time series data. In specific implementation, the historical power time series data is first organized into a format suitable for input into the LSTM, such as a three-dimensional tensor, and each time step contains multiple power parameters. Each unit of the LSTM can memorize and process long-term dependencies in the time series, and selectively remember or forget information through the mechanism of input, forgetting and output gates. After the regional historical power time series data is input into the LSTM, it is processed by multiple layers of LSTM units to extract global features in the data, such as long-term power demand trends and seasonal changes. During the training process, the back propagation algorithm is used to update all parameters in the global feature extractor so that it can accurately capture the global patterns in the historical data. The initial regional historical global time series features finally obtained provide comprehensive background information for subsequent analysis.

[0115] The process of initializing the fully connected layer parameters of the last fully connected layer in the global feature extractor and freezing the parameters of all other extractors in the global feature extractor is to ensure that the model remains stable during the transfer learning process. First, the parameters of the fully connected layer are randomly initialized so that it can adapt to the new data distribution. The function of the fully connected layer is to map the features output by the previous layer to the target space. This process requires flexible parameter settings to adapt to different data distributions. Then, the parameters of other extractors in the global feature extractor are frozen, that is, these parameters are no longer updated in the subsequent training process. The purpose of freezing parameters is to maintain the good performance of the model on historical data and avoid excessive impact on the overall structure of the model during the transfer learning process. In this way, it can be ensured that the global feature extractor still has stable feature extraction capabilities in the new data environment, while providing flexibility for learning new data.

[0116] The regional real-time power time series data is input into the global feature extractor for forward propagation to calculate the output of the model. Then, the error back propagation algorithm is used to update only the parameters of the last fully connected layer. During the forward propagation process, the regional real-time power time series data passes through the extractor layer with frozen parameters to extract preliminary features, and then is mapped to the target space through the fully connected layer. During the back propagation process, only the parameters of the fully connected layer are updated to enable it to better adapt to the regional real-time power time series data. In this way, the generalization ability of the model to new data can be improved while maintaining the original global feature extraction ability. The global feature extractor obtained after transfer learning can effectively extract global features in the new data environment. The global feature extractor after transfer learning should be able to more accurately reflect the global pattern in the historical data. By comparing the global features before and after transfer learning, the effect of transfer learning can be verified to ensure that the model still has good feature extraction capabilities under the new data distribution. Through the above steps, the regional historical global time series features with global background information can be finally obtained.

[0117] According to the preset fusion parameters Fusion weights are assigned to the regional real-time local time series features and the regional historical global time series features, where the fusion weight assigned to the real-time local time series features is , the fusion weight of the regional historical global temporal feature allocation is , combined with the assigned fusion weights, the regional real-time local time series features and the regional historical global time series features are weighted fused and spliced ​​to obtain the multi-scale time series features of power. The splicing method splices the local features and the global features in the feature dimension to form a new feature vector. The multi-scale time series features of power can simultaneously reflect the local changes and global trends of power operation. According to the different scales of the multi-scale time series features of power, the convolution layers of the multi-scale convolutional neural network are designed. For example, for local features in a short period of time, a smaller convolution kernel and a shallower convolution layer can be used; for long-term global features, a larger convolution kernel and a deeper convolution layer can be used. Then, the multi-scale time series features of power are input into the configured multi-scale convolutional neural network, and more advanced power operation time series features are extracted through multi-layer convolution and pooling operations. During the training process, the parameters of the convolution layer are updated layer by layer using the transfer learning method to improve the feature extraction ability of the model. Finally, the output power operation time series features can fully reflect the multi-scale information of power operation and provide reliable feature support for power prediction and scheduling.

[0118] In one embodiment, the convolutional layers of the multi-scale convolutional neural network are configured according to the characteristic scale of the multi-scale time series features of the electric power, the multi-scale time series features of the electric power are input into the configured multi-scale convolutional neural network, and the step of outputting the electric power operation time series features after the hierarchical transfer learning processing of the multi-scale convolutional neural network specifically includes the following steps:

[0119] The common convolution layer of the multi-scale convolutional neural network is configured based on the average characteristic scale of all time series features of the multi-scale time series features of power;

[0120] Sort all the time series features in the multi-scale time series features of electric power in descending order of feature scale;

[0121] Based on the number of scale types of feature scales and according to the sorting results of temporal features, multiple standard convolutional layers of different scales in the multi-scale convolutional neural network are sequentially configured, and the number of standard convolutional layers is the same as the number of scale types;

[0122] According to the sorting result, the first time series feature is used as the target time series feature, and the target time series feature is input into the common convolution layer and the target standard convolution layer of the corresponding scale, and the common parameters of the common convolution layer and the target standard parameters of the target standard convolution layer are updated and retained;

[0123] According to the sorting result, the next time series feature is used as the target time series feature in turn, the target time series feature is input into the multi-scale convolutional neural network, and the parameter updating step is performed based on the layered transfer learning method until the standard parameters of each standard convolutional layer are updated;

[0124] The optimized timing features output by each standard convolutional layer are integrated into the power operation timing features.

[0125] In this embodiment, in the design of the multi-scale convolutional neural network, it is first necessary to configure the common convolution layer according to the average feature scale of all time series features in the electric power multi-scale time series features. In order to determine the average feature scale, the scale value of each time series feature can be calculated first, for example, by counting the time window length or frequency component of each feature. The scale values ​​of all features are added together and then divided by the total number of features to obtain the average feature scale. Based on this average feature scale, the convolution kernel size and number of layers of the common convolution layer are designed so that it can effectively capture the feature pattern at the average scale. Next, all the time series features in the electric power multi-scale time series features need to be sorted in order from large to small according to the feature scale. Sorting these features from large to small according to the scale value means that features with long time spans or low frequencies are ranked first, and features with short time spans or high frequencies are ranked later.

[0126] According to the number of types of feature scales and the sorting results of temporal features, multiple standard convolutional layers of different scales in the multi-scale convolutional neural network are sequentially configured. Assuming that the number of types of feature scales is N, N standard convolutional layers need to be configured, and each convolutional layer corresponds to a feature scale. For example, if the types of feature scales include short-term, medium-term, and long-term, three standard convolutional layers are configured, corresponding to short-term features, medium-term features, and long-term features, respectively. The convolution kernel size and number of layers of each standard convolutional layer are designed according to the corresponding feature scale to ensure that the feature pattern at the scale can be effectively extracted. According to the sorting results, the first temporal feature is used as the target temporal feature and input into the common convolutional layer and the target standard convolutional layer of the corresponding scale. The target temporal feature passes through the common convolutional layer to extract preliminary feature patterns. Then, these preliminary features are input into the standard convolutional layer of the corresponding scale for further feature extraction. The common parameters of the common convolutional layer and the target standard parameters of the target standard convolutional layer are updated through the back propagation algorithm. The updated common parameters are retained in the common convolutional layer, and the target standard parameters are retained in the target standard convolutional layer. Ensure that the common convolutional layer can adapt to inputs of different feature scales, and optimize the parameters of the target standard convolutional layer so that it can better extract features of the corresponding scale.

[0127] Each time the next time series feature in the sorted list is selected, it is input into the multi-scale convolutional neural network, and the parameter update step is performed based on the layered transfer learning method. The above process is repeated and input into the common convolution layer and the standard convolution layer of the corresponding scale. The feature output is calculated by forward propagation, and then the parameters of the common convolution layer and the standard convolution layer are updated by back propagation. The layered transfer learning method ensures that only the standard convolution layer parameters corresponding to the current target feature are updated each time, and the parameters of the common convolution layer are adjusted in each update. This layer-by-layer update method can effectively utilize the learning results of the previous features and improve the accuracy and efficiency of subsequent feature extraction. This process continues until the standard parameters of each standard convolution layer are updated, ensuring that features of all scales are fully trained and optimized.

[0128] Finally, the optimized time series features output by each standard convolution layer are integrated into the power operation time series features. After processing the time series features of the corresponding scale, each standard convolution layer outputs features that represent the changes in power parameters at different time scales. These optimized time series features can be integrated by methods such as splicing and weighted averaging. The splicing method is to connect features of different scales in the feature dimension to form a comprehensive feature vector; the weighted average method is to weighted sum the features according to their importance to obtain comprehensive features. The integrated power operation time series features can fully reflect the operating status of the power system at different time scales.

[0129] In one embodiment, the step of inputting the target time series feature into the multi-scale convolutional neural network and performing the parameter updating step based on the hierarchical transfer learning method specifically includes the following steps:

[0130] Input the target time series features into the common convolution layer, output the common extracted features based on the current common parameters, and update and retain the common parameters at the same time;

[0131] Input the target time series features into the target standard convolution layer of the corresponding scale, load the upper standard parameters of the standard convolution layer of the previous scale into the target standard convolution layer, and output the standard extraction features based on the upper standard parameters;

[0132] Calculate the feature similarity between the common extracted features and the standard extracted features;

[0133] If the feature similarity is less than the preset similarity threshold, the standard parameters in the target standard convolution layer are reset, the target temporal features are re-input into the target standard convolution layer, and the target temporal features are used to update and retain the standard parameters of the target standard convolution layer;

[0134] If the feature similarity is greater than or equal to the similarity threshold, the upper layer standard parameters are retained as the standard parameters of the target standard convolution layer.

[0135] In this embodiment, in the common convolution layer, the target time series feature is convolved with the convolution kernel to obtain the convolution feature; then the nonlinear feature is enhanced through the activation function, such as ReLU; finally, the feature size is reduced and the main feature pattern is extracted through the pooling layer, such as maximum pooling or average pooling. The parameters of the common convolution layer will be updated through the back propagation algorithm in this process, and the updated parameters are retained for subsequent feature extraction. In the target standard convolution layer, the upper standard parameters of the previous scale standard convolution layer are first loaded. The parameters of the previous scale standard convolution layer contain the experience of extracting features at a larger scale, and these parameters can provide good initial values ​​for feature extraction at the current scale. After the target time series feature passes through the target standard convolution layer loaded with the upper standard parameters, the standard extraction feature generated contains the feature representation of the time series feature at the current scale. This process is also implemented by the forward propagation algorithm. The purpose of loading the upper standard parameters is to use the learning results of the previous layer to improve the efficiency and accuracy of feature extraction at the current layer.

[0136] Then the feature similarity between the common extracted features and the standard extracted features is calculated. Feature similarity is an indicator to measure the similarity between two feature representations. It can be calculated by a variety of methods, such as cosine similarity, Euclidean distance or Pearson correlation coefficient. Taking cosine similarity as an example, the specific calculation method is as follows: First, the common extracted features and standard extracted features are represented as vectors respectively; then, the dot product of the two vectors is calculated; then, the modulus of the two vectors is calculated; finally, the dot product is divided by the product of the modulus to obtain the similarity value. The similarity value is between 0 and 1. The closer the value is to 1, the more similar the two features are. The purpose of calculating feature similarity is to evaluate the consistency of the features extracted by the common convolution layer and the standard convolution layer to ensure the accuracy and effectiveness of feature extraction.

[0137] If the feature similarity is less than the preset similarity threshold, the standard parameters in the target standard convolution layer are reset, and the target time series features are re-input into the target standard convolution layer. Specifically, the parameters of the target standard convolution layer are initialized to random values ​​or pre-trained values; then, the target time series features are re-input into the target standard convolution layer, and new standard features are extracted through the forward propagation algorithm; then, the standard parameters are updated through the back propagation algorithm to ensure that the new feature extraction meets the requirements of the current feature scale. This process may require multiple iterations until the standard parameters are stable and the feature extraction effect is as expected. If the feature similarity is greater than or equal to the similarity threshold, it means that the upper-layer standard parameters can effectively extract the features of the current scale and there is no need to reset the parameters. The purpose of retaining the upper-layer standard parameters is to use the learning results of the previous layer, reduce the amount of calculation for parameter updates, and improve the efficiency of feature extraction. Finally, after the standard parameters of the target standard convolution layer are optimized, they can effectively extract the features of the current scale, providing a reliable data basis for subsequent feature integration and analysis.

[0138] In one embodiment, the power system energy supply prediction method based on transfer learning further includes the following steps:

[0139] Input the regional real-time local time series features into the preset ARMA model, and output the first preliminary prediction result through the ARMA model;

[0140] Input the power operation time series characteristics into the preset LSTM model, and output the second preliminary prediction result through the LSTM model;

[0141] Combine the first preliminary prediction result and the second preliminary prediction result to calculate the difference between the preliminary prediction results;

[0142] Based on the difference in the initial prediction results, the fusion parameters and the model parameters of the ARMA model and LSTM model are adjusted through the parameter adjustment algorithm.

[0143] In this embodiment, refer to Figure 4, the ARMA model, or the autoregressive moving average model, is a statistical model used for time series analysis and prediction. It mainly consists of two parts: the autoregressive (AR) part and the moving average (MA) part. The autoregressive part uses the past values ​​of the time series itself for prediction, while the moving average part uses the past prediction errors for correction. Since the regional real-time local time series characteristics have been stabilized through residual correction and the residual values ​​have been eliminated, the regional real-time local time series characteristics are time series feature data that conform to the ARMA model. In the specific implementation process, the regional real-time local time series characteristics are first input as input data into the ARMA model. Then, the model performs parameter estimation and model fitting through the set autoregressive order and moving average order to generate prediction results. Assume that the input data is , the mathematical expression of the ARMA model is:

[0144]

[0145] in, is a constant term, is the autoregression coefficient, is the sliding average coefficient, is the white noise error term. Through model calculation, the first initial prediction result is obtained ,This result reflects the predicted value of the regional real-time local time series characteristics.

[0146] The power operation time series characteristics are input into the preset LSTM model, and the second initial prediction result is output through the LSTM model. The LSTM model is a long short-term memory network. The specific process is to first input the power operation time series characteristics as input data into the LSTM model. The model processes the input data step by step through the series connection of multiple LSTM units and updates its internal state at each moment. Through model calculation, the second initial prediction result is obtained, which reflects the predicted value of the power operation time series characteristics.

[0147] The difference of the initial prediction results is calculated by combining the first initial prediction result and the second initial prediction result. The difference of the initial prediction results is an important indicator to measure the difference between the prediction results of the two models. In the specific implementation process, the first initial prediction result and the second initial prediction result are first compared, and their difference is calculated, so as to evaluate the degree of deviation of the two models when predicting the same time series characteristics. The smaller the difference, the closer the prediction results of the two models are, and the higher the prediction accuracy; the larger the difference, the larger the prediction results of the two models have a large deviation, and the model parameters need to be further adjusted. Based on the difference of the initial prediction results and the parameter adjustment algorithm, the fusion parameters and the model parameters of the ARMA model and the LSTM model are adjusted. The parameter adjustment algorithm is an optimization strategy used to minimize the difference of the initial prediction results and thus improve the prediction accuracy. Through the optimization algorithm (such as gradient descent method, genetic algorithm, etc.), the fusion parameters and the internal parameters of the ARMA model and the LSTM model (such as autoregressive coefficient, sliding average coefficient, LSTM unit weight, etc.) are adjusted to minimize the difference of the initial prediction results.

[0148] In one implementation manner, step S104 specifically includes the following steps:

[0149] Traversing the regional power supply topology data, extracting all power supply nodes and power consumption nodes in the regional power supply topology data, and identifying the supply relationship between all power supply nodes and all power consumption nodes;

[0150] Construct a regional power map network of the target area by combining all power supply nodes, all power consumption nodes and supply relationships;

[0151] Traversing the regional power supply topology data, extracting all regional edge power supply nodes and regional edge power consumption nodes according to the node attributes of the topology nodes in the regional power supply topology data;

[0152] The regional power map network of risky power supply areas is determined and intercepted by utilizing all regional edge power supply nodes and all regional edge power consumption nodes.

[0153] In this embodiment, in the regional power supply topology data, nodes represent power supply facilities (such as substations, power plants) and power consumption facilities (such as residential areas, industrial areas), and edges represent power supply lines. First, read the power supply topology data, and traverse each node to distinguish between power supply nodes and power consumption nodes. Power supply nodes usually have specific attributes, such as power generation capacity, transformer capacity, etc., while power consumption nodes have power demand attributes. By analyzing the node attributes, the type of node can be determined. Next, the supply relationship between the power supply node and the power consumption node is identified by analyzing the edges connecting these nodes. Each edge represents a power supply line, and its attributes include line capacity, voltage level, etc., through which the relationship between the power supply node and the power consumption node can be determined. Finally, all power supply nodes, power consumption nodes and their supply relationships are extracted to form a complete power supply topology structure.

[0154] In the regional power supply topology data, edge nodes refer to nodes located at the edge of the power supply network. These nodes are usually connected to other areas or directly facing end users. Traverse each node in the topology data and analyze its attributes to determine whether it is an edge node. The attributes of a power supply node may include the number of lines it is connected to, its geographical location, etc., while the attributes of a power consumption node may include its power consumption, location, etc. By comparing these attributes, it is possible to identify which nodes are edge power supply nodes and edge power consumption nodes. For example, if a power supply node is only connected to a few power consumption nodes and is located at the periphery of the network, it can be considered an edge power supply node. Similarly, a power consumption node can also be considered an edge power consumption node if it only obtains power from one or a few power supply nodes and is located at the edge of the network.

[0155] All identified regional edge power supply nodes and edge power consumption nodes are marked in the regional power graph network. Next, the connection relationship between these edge nodes is analyzed to determine which parts of the network may form risky power supply areas. Risky power supply areas usually refer to those areas with poor power supply stability and are easily affected by external factors. By analyzing the connection relationship of edge nodes, it is possible to identify which areas rely on a few nodes or lines for power supply, thereby determining the risk level of these areas. Then, the subgraph network of these risky areas is intercepted to form a subset of the regional power graph network.

[0156] In one implementation manner, step S105 specifically includes the following steps:

[0157] The regional power graph network is input into the preset graph neural network, and the graph convolution layer and graph pooling layer in the graph neural network are used to perform convolution and pooling processing on the regional power graph network in turn to obtain the initial topological features;

[0158] Input the initial topological features into the graph fully connected layer in the graph neural network, and update the graph fully connected parameters in the graph fully connected layer;

[0159] Input the regional power graph network into the graph fully connected layer and fine-tune the graph fully connected parameters;

[0160] The initial topological features are re-input into the fully connected graph layer after parameter fine-tuning, and the power topological structure features are output through the fully connected graph layer after parameter fine-tuning.

[0161] In this embodiment, refer to Figure 5 , input the regional power graph network into the preset graph neural network, and use the graph convolution layer and graph pooling layer in the graph neural network to perform convolution and pooling on the regional power graph network in turn to obtain the initial topological features. Graph neural network (GNN) is a deep learning model that can process graph structured data. Its core components include graph convolution layer and graph pooling layer. The graph convolution layer updates the node features by aggregating the neighbor information of the node. The formula is expressed as:

[0162]

[0163] in, It is Node in layer Features, Is a node The neighbor set of and The nodes are and The degree, It is The weight matrix of the layer, is the activation function. The graph pooling layer reduces the number of nodes in the graph through downsampling operations and retains important structural information. Through multi-layer graph convolution and pooling, high-order features of the graph can be extracted to form initial topological features.

[0164] The initial topological features are input into the graph fully connected layer in the graph neural network, and the graph fully connected parameters in the graph fully connected layer are updated. The graph fully connected layer is similar to the fully connected layer in the traditional neural network, but its input is the feature matrix of the graph. Through forward propagation, the output features of each node are calculated, and the formula is ,in Is a node The output features of is the weight matrix, is the bias vector. During the back propagation process, the gradient is calculated according to the loss function, and the weights and biases are updated by the gradient descent method. The loss function can be mean square error (MSE), cross entropy, etc. The updated graph fully connected layer parameters can better adapt to the initial topological features of the input, thereby improving the effect of feature extraction.

[0165] The regional power graph network is a subset of the regional power graph network and contains more detailed local structural information. It is input into the graph fully connected layer and the output is calculated through forward propagation. The formula is the same as before. During the fine-tuning process, the focus is on the characteristics of the regional power graph network so that the graph fully connected layer can more accurately capture the power topology structure in the region. The fine-tuned parameters can better reflect the characteristics of the regional power network and improve the performance of the model in local areas. After parameter fine-tuning, the initial topological features are input again, and forward propagation is performed through the fine-tuned graph fully connected layer to calculate the final feature output of each node. Since the parameters of the graph fully connected layer have been optimized according to the regional power graph network, the output features can more accurately reflect the topological structure of the power network.

[0166] The present invention also discloses a power system energy supply prediction system based on transfer learning, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the power system energy supply prediction method based on transfer learning as described in any one of the above embodiments.

[0167] Among them, the processor can adopt a central processing unit (CPU). Of course, according to actual usage, other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. can also be adopted. The general-purpose processor can adopt a microprocessor or any conventional processor, etc., and this application does not impose any restrictions on this.

[0168] Among them, the memory can be an internal storage unit of a computer device, such as a hard disk or memory of a computer device, or an external storage device of a computer device, such as a plug-in hard disk, a smart memory card (SMC), a secure digital card (SD) or a flash memory card (FC) equipped on the computer device, and the memory can also be a combination of an internal storage unit and an external storage device of a computer device. The memory is used to store computer programs and other programs and data required by the computer device. The memory can also be used to temporarily store data that has been output or is to be output, and this application does not impose any restrictions on this.

[0169] A person skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of protection of the present application is limited to these examples. In line with the concept of the present application, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of different aspects of one or more embodiments of the present application as above, which are not provided in detail for the sake of simplicity.

[0170] One or more embodiments of the present application are intended to cover all such substitutions, modifications and variations that fall within the broad scope of the present application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of one or more embodiments of the present application should be included in the protection scope of the present application.

Claims

1. A method for predicting energy supply of a power system based on transfer learning, characterized in that: A power dispatching system applied to a target area, wherein the target area includes a plurality of different power supply areas, comprises the following steps: In response to the regional power supply risk information received by the power dispatching system, a risk power supply area in the target area where the emergency power supply risk occurs is determined, and starting from the determination of the risk power supply area, regional real-time power data of the risk power supply area within a preset time period is cached, wherein the regional real-time power data includes regional power supply topology data and regional real-time power operation data; Acquiring regional power supply topology data and regional historical power operation data of the target area through the power dispatching system; preprocessing the regional real-time power operation data based on the grey prediction model theory to obtain preprocessed regional real-time power time series data; Preprocessing the historical power operation data of the region by using a fuzzy entropy algorithm to obtain preprocessed historical power time series data of the region; Inputting the regional real-time power time series data into a preset local feature extractor, performing local feature extraction through the local feature extractor, and obtaining regional real-time local time series features; Inputting the regional historical power time series data into a preset global feature extractor, performing global feature extraction through the global feature extractor, obtaining initial regional historical global time series features and updating all extractor parameters in the global feature extractor; Initialize the fully connected layer parameters of the last fully connected layer in the global feature extractor, and freeze all other extractor parameters in the global feature extractor; Inputting the regional real-time power time series data into the global feature extractor, using the regional real-time power time series data to implement transfer learning of the global feature extractor, and updating the fully connected layer parameters of the last fully connected layer to obtain the global feature extractor after transfer learning; Extracting regional historical global time series features from the regional historical power time series data by the global feature extractor after transfer learning; The real-time local time series characteristics of the region and the historical global time series characteristics of the region are fused into the multi-scale time series characteristics of electric power by using preset fusion parameters; Configuring the convolutional layers of the multi-scale convolutional neural network according to the characteristic scale of the multi-scale time series features of the power, inputting the multi-scale time series features of the power into the configured multi-scale convolutional neural network, and outputting the power operation time series features after hierarchical transfer learning processing of the multi-scale convolutional neural network; generating a regional power map network of the target area based on the regional power supply topology data, and extracting a regional power map network of the risk power supply area from the regional power map network according to the regional power supply topology data; Combining the regional power graph network with the area power graph network and using a graph neural network to extract power topology features; The power operation timing characteristics and the power topology structure characteristics are integrated into a comprehensive power characteristic, and the comprehensive power characteristic is input into a preset GCN-LSTM prediction model, and the power supply gap prediction result of the risk power supply area is output through the GCN-LSTM prediction model.

2. The method for predicting power system energy supply based on transfer learning according to claim 1 is characterized in that: The preprocessing of the regional real-time power operation data based on the grey prediction model theory to obtain the preprocessed regional real-time power time series data comprises the following steps: constructing a regional real-time power operation matrix based on the regional real-time power operation data; Performing a cumulative calculation on the real-time power operation data of the region to obtain a cumulative sequence data; Based on the grey prediction model theory and according to the once accumulated sequence data, obtaining the once accumulated adjacent sequence data; Constructing a primary accumulation neighbor sequence matrix according to the primary accumulation neighbor sequence data; The grey coefficient of the grey prediction model is calculated by combining the regional real-time power operation matrix and the primary accumulated neighboring sequence matrix; Using the grey coefficient to construct the differential equation of the grey prediction model and the corresponding system response function; Performing cumulative subtraction calculation on the once accumulated sequence data by using the system response function to obtain restored power operation data; The restored power operation data is subjected to residual inspection and residual correction processing to obtain regional real-time power time series data.

3. The method for predicting power system energy supply based on transfer learning according to claim 1, characterized in that: The step of configuring the convolutional layers of the multi-scale convolutional neural network according to the characteristic scale of the multi-scale time series feature of the electric power, inputting the multi-scale time series feature of the electric power into the configured multi-scale convolutional neural network, and outputting the electric power operation time series feature after the hierarchical transfer learning processing of the multi-scale convolutional neural network comprises the following steps: Configure a common convolution layer of a multi-scale convolutional neural network based on the average characteristic scale of all time series features of the multi-scale time series features of the power; Sorting all the time series features in the power multi-scale time series features in descending order of feature scale; Based on the number of scale types of the characteristic scale and according to the sorting result of the temporal features, sequentially configuring a plurality of standard convolutional layers of different scales in the multi-scale convolutional neural network, wherein the number of the standard convolutional layers is the same as the number of scale types; According to the sorting result, the first-ranked time series feature is used as the target time series feature, the target time series feature is input into the common convolution layer and the target standard convolution layer of the corresponding scale, and the common parameters of the common convolution layer and the target standard parameters of the target standard convolution layer are updated and retained; According to the sorting result, the next time series feature is used as the target time series feature in turn, the target time series feature is input into the multi-scale convolutional neural network, and a parameter updating step is performed based on a layered transfer learning method until the standard parameters of each standard convolutional layer are updated; The optimized timing features output by each of the standard convolutional layers are integrated into the power operation timing features.

4. The method for predicting power system energy supply based on transfer learning according to claim 3 is characterized in that: The step of inputting the target time series features into the multi-scale convolutional neural network and performing parameter updating based on the layered transfer learning method comprises the following steps: Inputting the target time series features into the common convolution layer, outputting common extraction features based on the current common parameters, and updating and retaining the common parameters; Inputting the target time series features into a target standard convolution layer of a corresponding scale, loading an upper layer standard parameter of the standard convolution layer of a previous scale into the target standard convolution layer, and outputting a standard extraction feature based on the upper layer standard parameter; Calculating feature similarity between the common extracted features and the standard extracted features; If the feature similarity is less than a preset similarity threshold, the standard parameters in the target standard convolution layer are reset, the target time series feature is re-input into the target standard convolution layer, and the target time series feature is used to update and retain the standard parameters of the target standard convolution layer; If the feature similarity is greater than or equal to the similarity threshold, the upper layer standard parameters are retained as the standard parameters of the target standard convolution layer.

5. The method for predicting power system energy supply based on transfer learning according to claim 1, characterized in that: The method further comprises the steps of: Inputting the regional real-time local time series features into a preset ARMA model, and outputting a first preliminary prediction result through the ARMA model; Inputting the power operation time series characteristics into a preset LSTM model, and outputting a second initial prediction result through the LSTM model; Calculating the difference of the initial prediction results by combining the first initial prediction result and the second initial prediction result; Based on the initial prediction result difference, the fusion parameters and the model parameters of the ARMA model and the LSTM model are adjusted through a parameter adjustment algorithm.

6. The method for predicting power system energy supply based on transfer learning according to claim 1, characterized in that: The generating of the regional power map network of the target area based on the regional power supply topology data, and extracting the regional power map network of the risk power supply area from the regional power map network according to the regional power supply topology data comprises the following steps: Traversing the regional power supply topology data, extracting all power supply nodes and power consumption nodes in the regional power supply topology data, and identifying the supply relationship between all the power supply nodes and all the power consumption nodes; Combining all the power supply nodes, all the power consumption nodes and the supply relationship to construct a regional power map network of the target area; Traversing the regional power supply topology data, and extracting all regional edge power supply nodes and regional edge power consumption nodes according to node attributes of topological nodes in the regional power supply topology data; The regional power map network of the risk power supply area is determined and intercepted in the regional power map network by using all the regional edge power supply nodes and all the regional edge power consumption nodes.

7. The method for predicting power system energy supply based on transfer learning according to claim 6, characterized in that: The combining of the regional power graph network and the area power graph network and extracting the power topology structure features by using a graph neural network comprises the following steps: Inputting the regional power graph network into a preset graph neural network, and using the graph convolution layer and the graph pooling layer in the graph neural network to sequentially perform convolution processing and pooling processing on the regional power graph network to obtain initial topological features; Inputting the initial topological features into a graph fully connected layer in the graph neural network, and updating graph fully connected parameters in the graph fully connected layer; Inputting the regional power graph network into the graph fully connected layer and fine-tuning the graph fully connected parameters; The initial topological features are re-input into the graph fully connected layer after parameter fine-tuning, and the power topological structure features are output through the graph fully connected layer after parameter fine-tuning.

8. A power system energy supply prediction system based on transfer learning, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method for predicting energy supply of a power system based on transfer learning as described in any one of claims 1 to 7 is implemented.

Images