Power demand prediction method based on complex network and graph neural network

Abstract

The invention relates to the technical field of data analysis, and discloses a power demand prediction method based on a complex network and a graph neural network, and the method comprises the steps: recognizing a coupling relation between multi-source time sequence feature data sets, and constructing a power demand complex network through the multi-source time sequence feature data sets and the coupling relation, the method comprises the following steps: performing feature learning on a power demand complex network through a graph neural network, generating a power demand prediction model, pruning the power demand prediction model to obtain a lightweight prediction model, and performing demand analysis on power data to be analyzed by using the lightweight prediction model to obtain a target prediction demand. And performing feature attribution analysis on the target prediction demand to obtain a feature contribution degree, generating a demand prediction reason according to the feature contribution degree, and generating a power demand prediction result according to the target prediction demand and the demand prediction reason. According to the method, the requirements of high precision, real-time performance and credibility on power demand prediction under the new-quality productivity development background are effectively met.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to a method for predicting electricity demand based on complex networks and graph neural networks. Background Technology

[0002] With the advancement of "dual carbon targets" and digital transformation, the power system is facing multiple challenges, including the integration of new energy sources, industrial restructuring, and cross-regional energy allocation. Against the backdrop of accelerating the development of new productive forces, electricity demand forecasting has become a crucial supporting technology for ensuring the safe operation of the power grid and meeting the electricity needs of emerging industries.

[0003] Existing electricity demand forecasting methods mostly employ time series modeling or machine learning approaches, achieving predictions by fitting historical load data or relevant economic indicators. However, traditional methods often struggle to characterize cross-regional and cross-industry coupling relationships, and their performance is inadequate when facing the diverse load characteristics under the development of new productive forces (such as rapidly fluctuating loads like data centers, energy storage, and electric vehicle charging). In recent years, complex network modeling and graph neural network (GNN) methods have been introduced into electricity demand forecasting. By constructing network relationships between regions, industries, and users, and utilizing the topological information in the graph structure, prediction accuracy is improved. These methods have certain advantages in improving the shortcomings of traditional models, but they also introduce new problems.

[0004] On the one hand, the prediction methods themselves are limited. Due to the high computational complexity of complex networks and deep learning models, prediction results are often difficult to generate quickly within minutes or even seconds, resulting in insufficient real-time performance and making it difficult to meet the rapid response requirements of power dispatch. At the same time, complex models are highly dependent on computing resources, which makes it difficult to implement them in the engineering of power company dispatch systems. In addition, deep models usually lack interpretability, and the prediction results are difficult for dispatchers and regulatory authorities to directly accept, thus limiting their widespread adoption.

[0005] On the other hand, it impacts the development of the power system and new productive forces. If the forecast results lack interpretability, it will be difficult to quickly locate the cause of any deviations, potentially threatening the safety of power grid operation. The output of new energy power generation is highly volatile, requiring high-precision demand forecasting; otherwise, the black-box nature of the model will lead to conservative or lagging dispatch strategies, thus restricting the absorption of new energy. Meanwhile, emerging industries such as data centers and intelligent manufacturing are highly dependent on electricity demand, but if the forecast results lack transparency and credibility, dispatch departments will be hesitant to adopt them directly, inevitably affecting the power supply security of these industries and hindering the high-quality development of new productive forces.

[0006] In summary, although the existing power demand prediction methods based on complex networks and graph neural networks can improve the shortcomings of traditional time series models to some extent, they still have high model complexity, poor interpretability, and difficulty in direct application to power dispatch, which cannot meet the requirements of high precision, real-time performance, and reliability for power demand prediction under the background of new productivity development. SUMMARY

[0007] The present application provides a power demand prediction method based on complex networks and graph neural networks, which aims to solve the problems of high model complexity, poor interpretability, and difficulty in direct application to power dispatch, and to meet the requirements of high precision, real-time performance, and reliability for power demand prediction under the background of new productivity development.

[0008] In the first aspect, to achieve the above-mentioned purpose, the present application provides a power demand prediction method based on complex networks and graph neural networks, which specifically includes: obtaining a multi-source time series feature dataset, identifying the coupling relationship between the multi-source time series feature dataset, and constructing a power demand complex network using the multi-source time series feature dataset and the coupling relationship; performing feature learning on the power demand complex network through a graph neural network to generate a power demand prediction model; pruning the power demand prediction model to obtain a lightweight prediction model; obtaining power data to be analyzed, using the lightweight prediction model to analyze the demand of the power data to be analyzed, and obtaining a target prediction demand; performing feature attribution analysis on the target prediction demand to obtain a feature contribution degree, and generating a demand prediction reason according to the feature contribution degree; generating a power demand prediction result according to the target prediction demand and the demand prediction reason.

[0009] The present application provides accurate data support for the scheduling, planning, and trading of the power system by analyzing the target prediction demand. It changes the power resource allocation from a rough mode based on historical experience to a precise mode based on data-driven, effectively reduces the risk of power shortage or resource waste caused by prediction deviation, and improves the economic efficiency and safety of power grid operation. By analyzing the power data to be analyzed, the target prediction demand is obtained, and fast, accurate, and interpretable power demand prediction is realized, solving the problems of high model complexity, poor interpretability, and difficulty in direct application to power dispatch, effectively meeting the requirements of high precision, real-time performance, and reliability for power demand prediction under the background of new productivity development. BRIEF DESCRIPTION OF DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the description of the embodiments of the present application. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0011] Figure 1 A flowchart of a power demand prediction method based on a complex network and a graph neural network provided by an embodiment of the present application is shown in the figure. Figure 2 A module diagram of a power demand prediction system based on a complex network and a graph neural network provided by an embodiment of the present application is shown in the figure.

[0012] The purposes, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0013] In order to make the person skilled in the art better understand the technical solutions of the present disclosure, and to fully understand and implement the implementation process of the present disclosure how to apply technical means to solve technical problems and achieve corresponding technical effects, the technical solutions in the embodiments of the present disclosure will be described clearly and completely in the embodiments of the present disclosure with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, not all. The embodiments of the present disclosure and each feature in the embodiments can be combined with each other without conflict, and the technical solutions formed thereby are all within the protection scope of the present disclosure. Based on the embodiments in the present disclosure, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present disclosure.

[0014] It should be noted that the terms "first", "second" and the like in the specification and claims of the present disclosure and the above-described drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or a chronological sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present disclosure described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0015] The embodiment of the application provides a power demand prediction method based on a complex network and a graph neural network. The method can be executed by software or hardware installed in a terminal device or a server device. The server device includes but is not limited to a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be a stand-alone server or a cloud server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and demand prediction platforms.

[0016] Referring to Figure 1 FIG. 1 is a flowchart of a power demand prediction method based on a complex network and a graph neural network according to an embodiment of the application. In this embodiment, the power demand prediction method based on a complex network and a graph neural network includes the following steps. S1, obtaining a multi-source time series feature dataset, identifying coupling relationships between the multi-source time series feature dataset, and constructing a power demand complex network using the multi-source time series feature dataset and the coupling relationships.

[0017] In this embodiment, the multi-source time series feature dataset includes but is not limited to power load data (historical power consumption, regional load curve), weather data (temperature, humidity, wind speed, sunshine duration), industrial economic indicators (industrial output value, enterprise production plan, holiday effect), policy / marketing data (electricity price, energy-saving and emission-reduction policy), and the like. The coupling relationship refers to the correlation or mutual influence relationship between different features. The power demand complex network refers to a graph model that represents the multi-source features and coupling relationships in the power demand system.

[0018] In detail, the identification of the coupling relationships between the multi-source time series feature dataset includes the following steps. S11, preprocessing the multi-source time series feature dataset to obtain a standard time series feature dataset. The preprocessing in this embodiment includes time alignment, and the standard time series feature dataset is an aligned time series feature dataset. S12, calculating the comprehensive coupling degree between any two time series features of the standard time series feature dataset. The specific calculation process of the comprehensive coupling degree is as follows. Uniformly sampling the aligned time series feature dataset to obtain a uniform time series feature dataset. Normalizing the uniform time series feature dataset to obtain a standard time series feature dataset. S12: Select two time series feature data in the standard time series feature dataset one by one, and combine them into a time series feature group to be analyzed. S13: Calculate the correlation coefficient and dynamic time warping distance between the time series feature groups to be analyzed. S14: Weighted fusion of the correlation coefficient and the dynamic time warping distance to obtain the comprehensive coupling degree between the time series feature groups to be analyzed.

[0019] S13: According to the preset dynamic threshold, the comprehensive coupling degree is judged to determine the initial relationship between the two time series feature data; S14: Multi-step lag analysis is performed on the time series feature data with the initial relationship, and the final coupling relationship is generated in combination with the initial relationship. The specific steps are as follows: Obtain a time window, and generate a dynamic threshold according to all the comprehensive coupling degrees in the time window and a preset adjustable parameter; Determine whether the comprehensive coupling degree is greater than or equal to the dynamic threshold; If the comprehensive coupling degree is less than the dynamic threshold, a weak correlation relationship is determined as the initial relationship between the time series feature groups to be analyzed; If the comprehensive coupling degree is greater than or equal to the dynamic threshold, a strong correlation relationship is determined as the initial relationship between the time series feature groups to be analyzed; Perform multi-step lag analysis on the time series feature groups to be analyzed to obtain lag information; Generate the coupling relationship between the time series feature groups to be analyzed according to the initial relationship and the lag information.

[0020] In detail, for multi-source time series feature data sets from different sources, since the sampling time stamps of each data may be inconsistent or missing, a time alignment processing method is adopted to map all feature data to a unified time axis. For example, for the case of hourly sampling of meteorological data and 15-minute sampling of power load data, each feature can be unified to the same time node through interpolation, filling or truncation, etc. to obtain an aligned time series feature dataset.

[0021] Select a target sampling interval (such as 15 minutes or 1 hour), and perform resampling operation on the aligned time series feature dataset, including average method, interpolation method or window aggregation method, to ensure that each feature expresses dynamic changes under the same time scale, and obtain a unified time series feature dataset.

[0022] Z-score standardization or Min-Max normalization method is adopted to map each feature data to a unified numerical interval (such as [0, 1] or standard normal distribution with mean 0 and variance 1), so as to ensure the comparability of different features under the same scale, and obtain a standard time series feature dataset.

[0023] The correlation coefficient represents the strength of linear correlation between two time series feature groups, and the calculation formula is as follows:

[0024] wherein, represents the correlation coefficient between the time series feature groups to be analyzed, represents the time series length of the time series feature group to be analyzed, represents the mean value of the i-th time series feature data in the time series feature group to be analyzed, represents the i-th time series feature data in the time series feature group to be analyzed at time point t, represents the j-th time series feature data in the time series feature group to be analyzed at time point t, represents the mean value of the j-th time series feature data in the time series feature group to be analyzed.

[0025] The dynamic time warping distance represents the shape similarity between two time series features, allowing local stretching of the time axis (for example, the peak of the power load is staggered at different time periods), and the calculation formula is as follows:

[0026] wherein, represents the dynamic time warping distance, represents the i-th time series feature data in the time series feature group to be analyzed at time point t, represents the j-th time series feature data in the time series feature group to be analyzed at time point s, represents the distance between the i-th time series feature data in the time series feature group to be analyzed at time point t and the j-th time series feature data in the time series feature group to be analyzed at time point s, represents the matching path.

[0027] By weighting and fusing the correlation coefficient and the dynamic time warping (DTW) distance, the linear correlation and the nonlinear time series alignment characteristics are comprehensively considered, so as to obtain the comprehensive coupling degree between the time series feature groups to be analyzed, so as to more comprehensively and accurately reflect the internal relationship between different feature groups.

[0028] In detail, the time window is obtained, and a dynamic threshold is generated according to all the comprehensive coupling degrees in the time window and a preset adjustable parameter, comprising: obtaining the comprehensive coupling degrees between all the time series feature groups to be analyzed in the time window; calculating the mean value and the standard deviation of the comprehensive coupling degrees between all the time series feature groups to be analyzed; multiplying the preset adjustable parameter and the standard deviation to obtain an adjustable standard deviation; The mean value is added to the adjustable standard deviation to obtain a dynamic threshold value.

[0029] In detail, the mean value and the standard deviation of all comprehensive coupling values are calculated in a specified time window to reflect the current overall coupling level and fluctuation range of the feature group, the adjustable standard deviation can be used to adjust the sensitivity of the threshold value, and the dynamic threshold value can be used to filter out significant coupling relationships, so that only strong coupling feature pairs are reserved when constructing a complex network, and the calculation formula is as follows:

[0030] wherein, denotes the dynamic threshold value, denotes the mean value, denotes the adjustable parameter, denotes the standard deviation.

[0031] In detail, the power demand complex network is constructed by using the multi-source time series feature dataset and the coupling relationship, comprising: each standard time series feature data in the standard time series feature dataset is taken as a network node; when the initial relationship between the to-be-analyzed time series feature groups is a weak correlation relationship, no edge is established between the to-be-analyzed time series feature groups; when the initial relationship between the to-be-analyzed time series feature groups is a strong correlation relationship, a network edge is established between the to-be-analyzed time series feature groups; an initial demand network is constructed according to the network nodes and the network edges; an edge attribute of the network edge is generated according to the coupling relationship between the to-be-analyzed time series feature groups; the initial demand network is filled by using the edge attribute, and a power demand complex network is obtained.

[0032] In detail, each standardized time series feature data is regarded as a node in a complex network, and the strength of the association is determined according to the comprehensive coupling degree of each pair of to-be-analyzed time series feature groups: if it is a weak correlation relationship, no edge is established between the nodes; if it is a strong correlation relationship, a network edge is established between the corresponding nodes. The edge attribute (such as weight representing coupling strength, direction representing causal or lag relationship) of each edge is generated according to the coupling relationship between the feature groups, and these edge attributes are filled into the initial demand network, so that a complete power demand complex network is obtained.

[0033] In identifying the coupling relationship, the multi-source data is time-aligned, uniformly sampled and normalized to ensure the comparability of different source features, and then the comprehensive coupling degree is calculated by the weighted fusion of correlation coefficient, mutual information and dynamic time warping distance, and the strong and weak relationships and the hysteresis effect between the features are determined by the dynamic threshold, so as to accurately depict the complex dependence relationship between cross-features and cross-regions. In constructing the power demand complex network, each feature is taken as a node, the network edges are established or omitted according to the initial relationship, and the edge attributes are given by using the coupling relationship, so that the network not only reflects the connection between nodes, but also quantifies the connection strength and hysteresis characteristics. This method can comprehensively capture the multi-dimensional coupling relationship between time series features, improve the expression ability of the power demand prediction model to the complex load changes across regions and industries, and provide high-quality and structured input for subsequent graph neural network feature learning, which helps to improve the prediction accuracy and interpretability.

[0034] S2, learning features of the power demand complex network by a graph neural network to generate a power demand prediction model.

[0035] In the embodiment of the application, the standardized time series features of each network node are taken as initial node features input into a graph neural network (GNN), and through multi-layer graph convolution or message passing mechanism, each node receives information from neighbor nodes and fuses with its own features, and the node representation is updated layer by layer, so as to capture the coupling relationship and time series dependence across nodes in the network. After several layers of feature learning, the final representation vector of the node integrates local and global structure information and coupling dynamics, which is mapped into future power demand values by a graph-level or node-level prediction layer to form a power demand prediction model.

[0036] In detail, the learning features of the power demand complex network by a graph neural network to generate a power demand prediction model comprises: constructing an adjacency matrix and a node feature matrix according to the power demand complex network; performing multi-layer graph convolution on the power demand complex network by using a preset graph neural network model according to the adjacency matrix, the node feature matrix and a preset trainable weight, to obtain node updated feature embedding; performing linear transformation on the node updated feature embedding to obtain a power demand prediction value; obtaining a historical power demand true value according to the power demand complex network; calculating a mean absolute percentage error according to the power demand prediction value and the power demand true value; determining whether the mean absolute percentage error is greater than a preset error threshold; if the mean absolute percentage error is greater than the error threshold, fine-tuning the graph neural network model to obtain a power demand prediction model; If the average absolute percentage error is greater than the error threshold, the graph neural network model is taken as the power demand prediction model.

[0037] In detail, according to the power demand complex network, a node set and an edge set in the network are constructed into an adjacency matrix, and a standardized time sequence feature of each node is formed into a node feature matrix. The adjacency matrix, the node feature matrix and a preset trainable weight are input into a graph neural network model, and a node representation is iteratively updated through multi-layer graph convolution operation.

[0038] Each layer of nodes receives information from neighbor nodes and fuses its own features. After several layers of updates, a final node feature embedding is obtained, which reflects the local characteristics, neighbor information and overall coupling relationship of the network. The calculation formula is as follows:

[0039] wherein, denotes a normalized form of the adjacency matrix, denotes a node representation of the lth layer of the graph neural network model, denotes a trainable weight of the lth layer of the graph neural network model.

[0040] Each node's high-dimensional embedding is mapped to a corresponding power demand prediction value through linear transformation. The historical true load value is obtained from the corresponding monitoring system or database of the power demand complex network as a reference standard. The mean absolute percentage error (MAPE) is calculated, and the calculation formula is as follows:

[0041] wherein, denotes the total number of node updated feature embeddings, denotes the mth historical power demand true value, denotes the mth power demand prediction value.

[0042] The adjacency matrix and the node feature matrix are constructed so that the network structure and the node characteristics are fully utilized. The node embedding is updated through multi-layer graph convolution to fuse local and global information. The prediction value is generated through linear transformation, and the mean absolute percentage error (MAPE) is calculated by comparing with the historical true value to realize model precision evaluation and controllable optimization. Through error threshold judgment for fine-tuning or direct use of the model, the balance of precision and stability of the prediction model can be ensured. Not only the precision of cross-regional and cross-industry load prediction is improved, but also the reliability and applicability of the model are enhanced through adjustable fine-tuning mechanism and error monitoring, providing a high-precision, interpretable and applicable prediction tool for power dispatch and dynamic power management in the context of new productivity.

[0043] S3, pruning the power demand prediction model to obtain a lightweight prediction model.

[0044] In the embodiments of the present application, the importance of the weight matrix and node feature connection of each layer is evaluated, the contribution of each weight to the prediction output is calculated, and the weights with small contribution to the prediction are removed, thereby reducing the calculation amount and memory occupation of the model. After pruning and necessary fine-tuning, a lightweight prediction model is obtained.

[0045] In detail, the power demand prediction model is pruned to obtain a lightweight prediction model, including: generating candidate pruning objects and pruning granularity according to the power demand prediction model; obtaining edge weights of the candidate pruning objects, and generating weight importance of each candidate pruning object according to the edge weights; comparing the weight importance with a preset pruning threshold, and screening out candidate pruning objects with weight importance less than the pruning threshold; screening out the candidate pruning objects as target pruning objects; deleting the target pruning objects of the power demand prediction model according to the pruning granularity to obtain a lightweight prediction model.

[0046] In detail, the pruning objects are, for example, edges, nodes or weight matrix elements in a graph neural network, and the pruning granularity is, for example, a single weight, an entire connection or an entire neuron.

[0047] The edge weights corresponding to the candidate pruning objects (such as the connection edges between nodes) are extracted, and the edge weights usually reflect the coupling strength or correlation degree of the power demand features between two nodes. These edge weights are used as a basic index to calculate the weight importance of each candidate pruning object, for example, the absolute value, gradient information or sensitivity analysis can be used to measure the contribution to the prediction output.

[0048] The weight importance of each candidate pruning object is compared with a pruning threshold, and objects with importance lower than the pruning threshold are screened out. These objects are considered to have small contribution to the prediction accuracy of the model and are the targets that can be pruned first. According to the pruning granularity, the target pruning objects in the power demand prediction model are deleted or zeroed, thereby reducing the model parameters and calculation amount. After necessary fine-tuning or retraining, a lightweight prediction model is obtained, which improves the inference efficiency and resource utilization while ensuring the prediction accuracy.

[0049] The calculation complexity and memory occupation of the power demand prediction model are effectively reduced by systematic screening and deleting model parameters with less prediction contribution. Candidate pruning objects are generated according to the model structure, and the pruning granularity is determined, then the weight importance is calculated through the edge weight, the low-contribution objects are screened out as the target pruning objects, and finally the pruning is performed according to the pruning granularity. While retaining the key features and prediction ability to the greatest extent, a lightweight prediction model is generated, so as to improve the model inference speed, reduce resource consumption, and be more suitable for real-time and high-frequency load prediction applications in the power dispatching system, while maintaining the prediction accuracy and reliability.

[0050] S4, obtaining the to-be-analyzed power data, performing demand analysis on the to-be-analyzed power data by using the lightweight prediction model to obtain a target prediction demand.

[0051] In the embodiment of the application, the to-be-analyzed power data including the latest time sequence features and real-time load information of each node are input into the lightweight prediction model optimized by pruning, the future power demand prediction value of each node is generated through the graph convolution and linear mapping operation of the lightweight prediction model, the real-time input features are mapped into high-dimensional embedding and then into the prediction load that can be directly used, the rapid response to the load change of the current power system is realized, and the target prediction demand is obtained.

[0052] In detail, the demand analysis on the to-be-analyzed power data by using the lightweight prediction model to obtain the target prediction demand includes: aligning the to-be-analyzed power data by time stamp to obtain aligned power data; vectorizing the aligned power data to obtain a power feature vector; identifying the power feature coupling relationship of the power feature vector according to the power demand complex network in the lightweight prediction model; generating a power feature relationship graph according to the power feature vector and the power feature coupling relationship; performing graph convolution on the power feature relationship graph by using the lightweight prediction model to obtain a power feature node embedding sequence; extracting the time dependence relationship of the power feature node embedding sequence, and generating a time sequence feature according to the time dependence relationship; performing attention weighting on the power feature node embedding sequence and the time sequence feature to obtain a space-time joint feature; performing full connection and inverse normalization on the space-time joint feature to obtain the target prediction demand.

[0053] In detail, the collection time of each node and each feature of the power data to be analyzed is unified to the same time step to ensure the synchronization of different data sources and features in subsequent analysis, thereby obtaining aligned power data. The aligned power data is converted into a unified numerical vector form, and the multi-dimensional time series features of each node are integrated into a power feature vector.

[0054] According to the complex network structure of power demand in the lightweight prediction model, the node relationship between the power feature vectors is analyzed, the coupling and dependence relationship between different features is identified, the identified feature coupling relationship is used, the power feature vector is taken as a node, and an edge is established according to the coupling relationship, and a power feature relationship graph is constructed, so that the spatial correlation of the data is represented by a graph structure.

[0055] The power feature relationship graph is input into the lightweight prediction model, the node itself features and neighbor node information are fused through graph convolution operation, the power feature node embedding sequence is generated, and the high-dimensional representation of each node in the network is reflected. The node embedding sequence is analyzed along the time dimension, the time dependence in the historical data is captured, the time sequence features are generated, and the model understands the load change trend and periodicity. The spatial features and time sequence features of the node embedding sequence are weighted and fused through attention, and spatial-time joint features are obtained, so that the model can pay attention to the influence of key nodes and key time points on load prediction.

[0056] The spatial-time joint features are input into the full connection layer for mapping, and the output is subjected to inverse normalization processing, the model prediction value is converted into actual power units, and finally the target prediction demand is obtained, providing a directly usable prediction result for power grid dispatching and load management.

[0057] The uniformity and processability of multi-source time series data are ensured through timestamp alignment and feature vectorization, the feature coupling relationship is identified by using the complex network of power demand, the feature relationship graph is constructed, the spatial correlation is structured, the node embedding is generated through graph convolution, the time sequence features are extracted in combination with the time dependence, the spatial-time joint features are obtained through attention weighted fusion, and finally the target prediction demand is generated through full connection mapping and inverse normalization, while capturing the spatial coupling relationship between nodes and the time dependence of historical load, realizing fast, accurate and interpretable power demand prediction, providing efficient decision support for power grid dispatching, load management and multi-element electricity scene under new quality productivity.

[0058] S5, performing feature attribution analysis on the target prediction demand to obtain a feature contribution degree, and generating a demand prediction reason according to the feature contribution degree.

[0059] In the embodiments of the present application, the SHAP or LIME method is used to calculate the contribution of each feature to the prediction result, and to quantify the positive or negative effect of each factor in the load change. The contribution information is converted into understandable prediction reasons to generate explanatory statements, such as "the future one-hour load increase is mainly driven by the temperature drop (contribution 65%) and the electric vehicle charging peak (contribution 25%)".

[0060] In detail, the feature attribution analysis is performed on the target predicted demand to obtain a feature contribution degree, and a demand prediction reason is generated according to the feature contribution degree, comprising: randomly perturbing the aligned power data to generate a plurality of local perturbation power data; performing demand analysis on the local perturbation power data using the lightweight prediction model to obtain a local perturbation power prediction demand corresponding to each local perturbation power data; calculating the Euclidean distance between each local perturbation power data and the aligned power data; generating a perturbation weight of the local perturbation power data according to the Euclidean distance; constructing an initial linear model according to the local perturbation power data and the local perturbation power prediction demand; performing local weighted least squares on the initial linear model using the perturbation weight to obtain an intercept and a plurality of feature coefficients of the initial linear model; fitting a local linear model using the initial linear model, the intercept and the feature coefficients; extracting a feature contribution degree of the target predicted demand using the local linear model; sorting the feature contribution degrees in descending order, and selecting a key feature corresponding to a preset number of feature contribution degrees according to the sorted feature contribution degrees; obtaining a contribution direction and a contribution size of the selected key feature, and generating a demand prediction reason according to the key feature, the contribution direction and the contribution size.

[0061] In detail, the aligned power data is randomly perturbed to generate a plurality of local perturbation power data, each perturbation simulating a small change in the aligned power data, so as to analyze the sensitivity of the prediction result to the feature in the subsequent analysis. Each local perturbation power data is input into the lightweight prediction model for demand analysis to obtain the corresponding local perturbation power prediction demand, which is used to observe the response of the prediction result to the input change. The Euclidean distance between each local perturbation power data and the original aligned power data is calculated to measure the amplitude of the input perturbation and provide a quantitative basis for weight allocation.

[0062] The perturbation weight of each local perturbation power data is generated according to the Euclidean distance, the smaller the distance, the greater the weight of the perturbation sample, so that the local linear model pays more attention to the data points close to the original input. An initial linear model is fitted on these weighted perturbation samples to minimize the weighted squared error, and the intercept and feature coefficients are obtained. The fitted local linear model can approximately reflect the input-output relationship of the lightweight prediction model near the target data point. The calculation formula of the local linear model is as follows:

[0063] wherein, denotes the local perturbation power prediction demand output by the local linear model, denotes the intercept, denotes the dimension of the local perturbation power data, denotes the feature coefficient of the nth feature, denotes the nth local perturbation power data input.

[0064] The contribution of each input feature to the target prediction demand is extracted from the fitted local linear model, and the influence size of each feature in the prediction result is quantified. The feature contribution is sorted in descending order, and the feature with the highest contribution is selected as the key feature according to the preset number, which is used to explain the prediction result.

[0065] The contribution direction and contribution size of the selected key feature are obtained, and the demand prediction reason is generated combining these information, such as natural language or structured explanation, to explain the main driving factors of the prediction load rise or fall, and to provide understandable basis for scheduling decision.

[0066] Through local perturbation and linear fitting of the target prediction demand, the explainability of the complex prediction model is improved. By generating neighborhood samples through random perturbation and performing prediction, combining the Euclidean distance to calculate the perturbation weight, fitting the local linear model, quantifying the contribution of each feature to the prediction result, selecting the key feature and extracting the contribution direction and size, and generating the demand prediction reason, the prediction result of the lightweight prediction model can be converted into understandable reason explanation, so that the dispatch personnel can clearly understand the main driving factors behind the load change, thereby improving the transparency, credibility and operability of the decision, and providing reliable basis for power grid dispatching, load management and emergency response.

[0067] S6, generating a power demand prediction result according to the target prediction demand and the demand prediction reason.

[0068] In the embodiment of the present application, the target prediction demand is organized according to time sequence and node, and the contribution degree, contribution direction and generated explanatory reason of each key feature are attached. These information is encapsulated as the power demand prediction result.

[0069] In detail, the generating power demand prediction result according to the target prediction demand and the demand prediction reason comprises: performing confidence analysis on the target prediction demand and the demand prediction reason to obtain a comprehensive confidence score; judging whether the comprehensive confidence score is greater than a preset confidence threshold; if the comprehensive confidence score is greater than the confidence threshold, converting the demand prediction reason into an operable suggestion; summarizing the target prediction demand and the operable suggestion into the power demand prediction result; if the comprehensive confidence score is less than or equal to the confidence threshold, adding a risk label to the target prediction demand to generate a labeled prediction demand and a mitigation suggestion; converting the demand prediction reason into an operable suggestion; summarizing the labeled prediction demand, the operable suggestion and the mitigation suggestion into the power demand prediction result.

[0070] In detail, the comprehensive confidence score is calculated based on historical load prediction residual statistics (such as historical MAPE or MSE distribution) and attribution stability (such as consistency of feature contribution in multiple disturbance analysis), a prediction with small residual and stable attribution is given a higher confidence, and a confidence score is output to quantify the reliability of the prediction result, helping the dispatcher to judge the reliability of the prediction result.

[0071] judging whether the comprehensive confidence score is greater than the confidence threshold: if the comprehensive confidence score is higher than the confidence threshold, it means that the prediction result is reliable, the demand prediction reason is converted into an operable dispatching suggestion, and the target prediction demand is summarized to generate the final power demand prediction result. If the comprehensive confidence score is lower than or equal to the confidence threshold, it means that the prediction has uncertainty, a risk label needs to be added to the target prediction demand, and a labeled prediction demand is generated in combination with a mitigation measure, then the demand prediction reason is converted into an operable suggestion, and finally the labeled prediction demand, the operable suggestion and the mitigation measure are integrated into the power demand prediction result. This method ensures that the prediction result can be directly applied to dispatching, and in the case of uncertainty, it provides risk prompt and coping strategies, improving the safety and reliability of power grid dispatching.

[0072] By combining the target predicted demand with the demand prediction reason and confidence analysis, a precise and explainable power demand prediction result can be generated. In the case of high confidence, the system converts the prediction reason into an operable scheduling suggestion, enabling dispatch personnel to quickly make decisions; in the case of low confidence, risk labels are added and mitigation measures are provided to ensure the safe operation of the power grid. Not only does this improve the reliability and transparency of the prediction result, but it also takes into account the operability of decisions and risk prevention, providing comprehensive support for power grid dispatching, reserve capacity planning, and new energy consumption, significantly enhancing the application value of the prediction result in actual operation.

[0073] It should be understood that the size of the serial number of each step in the above embodiment does not mean the order of execution, and the execution order of each process should be determined according to its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present application.

[0074] As Figure 2 shown, is a functional module diagram of a power demand prediction system based on a complex network and a graph neural network provided by an embodiment of the present application.

[0075] In the embodiments of the present disclosure, a power demand prediction system based on a complex network and a graph neural network is provided, which corresponds one-to-one to the above-mentioned power demand prediction method based on a complex network and a graph neural network. As Figure 2 shown, the power demand prediction system based on a complex network and a graph neural network 100 includes a demand network construction module 101, a prediction model construction module 102, a prediction module pruning module 103, a power demand analysis module 104, a demand reason generation module 105, and a prediction result generation module 106. The detailed description of each functional module is as follows: The demand network construction module 101 is configured to obtain a multi-source time series feature dataset, identify the coupling relationship between the multi-source time series feature dataset, and construct a power demand complex network using the multi-source time series feature dataset and the coupling relationship; The prediction model construction module 102 is configured to perform feature learning on the power demand complex network through a graph neural network to generate a power demand prediction model; The prediction module pruning module 103 is configured to prune the power demand prediction model to obtain a lightweight prediction model; The power demand analysis module 104 is configured to obtain power data to be analyzed, perform demand analysis on the power data to be analyzed using the lightweight prediction model, and obtain a target predicted demand; The demand reason generation module 105 is configured to perform feature attribution analysis on the target predicted demand, obtain a feature contribution degree, and generate a demand prediction reason according to the feature contribution degree. The prediction result generation module 106 is configured to generate a power demand prediction result according to the target prediction demand and the demand prediction reason.

[0076] In an embodiment, the demand network construction module 101 performs the following steps in identifying the coupling relationship between the multi-source time series feature data sets: time aligning the multi-source time series feature data sets to obtain an aligned time series feature data set; unifying the sampling frequency of the aligned time series feature data set to obtain a unified time series feature data set; normalizing the unified time series feature data set to obtain a standard time series feature data set; selecting two time series feature data in the standard time series feature data set one by one to combine into a to-be-analyzed time series feature group; calculating the correlation coefficient and the dynamic time warping distance between the to-be-analyzed time series feature groups; weighting and fusing the correlation coefficient and the dynamic time warping distance to obtain a comprehensive coupling degree between the to-be-analyzed time series feature groups; obtaining a time window, and generating a dynamic threshold according to all the comprehensive coupling degrees in the time window and a preset adjustable parameter; judging whether the comprehensive coupling degree is greater than or equal to the dynamic threshold; if the comprehensive coupling degree is less than the dynamic threshold, determining a weak correlation relationship as an initial relationship between the to-be-analyzed time series feature groups; if the comprehensive coupling degree is greater than or equal to the dynamic threshold, determining a strong correlation relationship as the initial relationship between the to-be-analyzed time series feature groups; performing multi-step lag analysis on the to-be-analyzed time series feature group to obtain lag information; generating a coupling relationship between the to-be-analyzed time series feature groups according to the initial relationship and the lag information.

[0077] In an embodiment, the demand network construction module 101 performs the following steps in obtaining a time window and generating a dynamic threshold according to all the comprehensive coupling degrees in the time window and a preset adjustable parameter: obtaining the comprehensive coupling degrees between all the to-be-analyzed time series feature groups in the time window; calculating the mean and the standard deviation of the comprehensive coupling degrees between all the to-be-analyzed time series feature groups; multiplying the preset adjustable parameter and the standard deviation to obtain an adjustable standard deviation; adding the mean and the adjustable standard deviation to obtain a dynamic threshold.

[0078] In an embodiment, the demand network construction module 101 is configured to construct a power demand complex network by using the multi-source time-series feature dataset and the coupling relationship, for: taking each standard time-series feature data in the standard time-series feature dataset as a network node; when the initial relationship between the to-be-analyzed time-series feature groups is a weak correlation, no edge is established between the to-be-analyzed time-series feature groups; when the initial relationship between the to-be-analyzed time-series feature groups is a strong correlation, an edge is established between the to-be-analyzed time-series feature groups; constructing an initial demand network according to the network nodes and the network edges; generating an edge attribute of the network edge according to the coupling relationship between the to-be-analyzed time-series feature groups; filling the initial demand network by using the edge attribute to obtain a power demand complex network.

[0079] In an embodiment, the prediction model construction module 102 is configured to perform feature learning on the power demand complex network by using a graph neural network to generate a power demand prediction model, for: constructing an adjacency matrix and a node feature matrix according to the power demand complex network; performing multi-layer graph convolution on the power demand complex network by using a preset graph neural network model according to the adjacency matrix, the node feature matrix, and a preset trainable weight to obtain a node updated feature embedding; performing linear transformation on the node updated feature embedding to obtain a power demand prediction value; obtaining a historical power demand true value according to the power demand complex network; calculating a mean absolute percentage error according to the power demand prediction value and the power demand true value; determining whether the mean absolute percentage error is greater than a preset error threshold; if the mean absolute percentage error is greater than the error threshold, fine-tuning the graph neural network model to obtain a power demand prediction model; if the mean absolute percentage error is greater than the error threshold, taking the graph neural network model as a power demand prediction model.

[0080] In an embodiment, the prediction module pruning module 103 is configured to perform pruning on the power demand prediction model to obtain a lightweight prediction model, for: generating a candidate pruning object and a pruning granularity according to the power demand prediction model; obtaining an edge weight of the candidate pruning object, and generating a weight importance of each candidate pruning object according to the edge weight; comparing the weight importance with a preset pruning threshold to screen out a candidate pruning object with a weight importance less than the pruning threshold; taking the screened-out candidate pruning object as a target pruning object; deleting the target pruning object of the power demand prediction model according to the pruning granularity to obtain a lightweight prediction model.

[0081] In an embodiment, the power demand analysis module 104 performs demand analysis on the power data to be analyzed using the lightweight prediction model to obtain a target prediction demand, which is used for: timestamp alignment of the power data to be analyzed to obtain aligned power data; feature vectorization of the aligned power data to obtain a power feature vector; identification of a power feature coupling relationship of the power feature vector according to the power demand complex network in the lightweight prediction model; generation of a power feature relationship graph according to the power feature vector and the power feature coupling relationship; graph convolution of the power feature relationship graph using the lightweight prediction model to obtain a power feature node embedding sequence; extraction of a time dependence relationship of the power feature node embedding sequence, and generation of a time series feature according to the time dependence relationship; attention weighting of the power feature node embedding sequence and the time series feature to obtain a space-time joint feature; full connection and de-normalization of the space-time joint feature to obtain a target prediction demand.

[0082] In an embodiment, the demand reason generation module 105 performs feature attribution analysis on the target prediction demand to obtain a feature contribution degree, and generates a demand prediction reason according to the feature contribution degree, which is used for: random perturbation of the aligned power data to generate a plurality of local perturbation power data; demand analysis of the local perturbation power data using the lightweight prediction model to obtain a local perturbation power prediction demand corresponding to each of the local perturbation power data; calculation of the Euclidean distance of each of the local perturbation power data and the aligned power data; generation of a perturbation weight of the local perturbation power data according to the Euclidean distance; construction of an initial linear model according to the local perturbation power data and the local perturbation power prediction demand; performing local weighted least squares on the initial linear model by using the disturbance weight, to obtain an intercept and a plurality of feature coefficients of the initial linear model; fitting a local linear model by using the initial linear model, the intercept and the feature coefficients; extracting a feature contribution degree of the target predicted demand by using the local linear model; sorting the feature contribution degrees in descending order, and screening out key features corresponding to a preset number of feature contribution degrees according to the sorted feature contribution degrees; obtaining a contribution direction and a contribution size of the screened key features, and generating a demand prediction reason according to the key features, the contribution direction and the contribution size.

[0083] In an embodiment, the prediction result generation module 106 generates a power demand prediction result according to the target predicted demand and the demand prediction reason, for: performing confidence analysis on the target predicted demand and the demand prediction reason to obtain a comprehensive confidence score; determining whether the comprehensive confidence score is greater than a preset confidence threshold; if the comprehensive confidence score is greater than the confidence threshold, converting the demand prediction reason into an operable suggestion; summarizing the target predicted demand and the operable suggestion into the power demand prediction result; if the comprehensive confidence score is less than or equal to the confidence threshold, adding a risk label to the target predicted demand to generate a labeled prediction demand and a mitigation suggestion; converting the demand prediction reason into an operable suggestion; summarizing the labeled prediction demand, the operable suggestion and the mitigation suggestion into the power demand prediction result.

[0084] In the present application, the specific limitations of the power demand prediction system based on complex network and graph neural network can be referred to the limitations of the power demand prediction method based on complex network and graph neural network in the above, which will not be repeated here. Each module in the above power demand prediction system based on complex network and graph neural network can be realized by software, hardware and their combinations, in whole or in part. Each module can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so as to be called and executed by the processor to perform the operations corresponding to each module.

[0085] In the embodiments of the present application, it should be understood that the disclosed system can be implemented in other manners. For example, the division of the system embodiments described above is merely illustrative, and the division of the modules can be other division manners.

[0086] In addition, each function module in each embodiment of the present application can be integrated in a processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware, or in the form of hardware plus software function modules.

[0087] Therefore, from any viewpoint, the embodiments should be regarded as exemplary and non-limiting, and the scope of the present application is defined by the appended claims rather than the above description, and therefore all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present application. Any reference signs in the claims should not be regarded as limiting the claims to which they relate.

[0088] It is obvious for those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and the present application can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application.

[0089] It can be understood by those skilled in the art that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing relevant hardware, and the computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, the processes of the above-mentioned embodiments of each method can be included. In each embodiment provided by the present application, any reference to memory, storage, database or other medium can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0090] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the above-mentioned division of each functional unit and module is exemplified, and in actual application, the above-mentioned functions can be completed by different functional units and modules according to needs, that is, the internal structure of the system is divided into different functional units or modules to complete all or part of the functions described above.

[0091] In the embodiments provided by the present disclosure, it should be understood that the disclosed system and method can also be implemented in other ways. The system embodiments described above are only illustrative. For example, the flowcharts and block diagrams in the drawings show possible implementation architectures, functions and operations of the system, method and computer program product according to the embodiments of the present disclosure. In this regard, each block in the flowcharts or block diagrams can represent a module, a program segment or a part of code, which contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur in different orders from those shown in the drawings. For example, two consecutive blocks can actually be executed substantially in parallel, and sometimes they can be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and the combination of blocks in the block diagrams and / or flowcharts, can be implemented by a dedicated hardware-based system that performs the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions.

[0092] It should be noted that in the present disclosure, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or further includes elements inherent to such a process, method, article or device. Without more limitations, the element limited by the statement "including a" does not exclude the presence of additional identical elements in the process, method, article or device including the element.

[0093] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A power demand prediction method based on complex network and graph neural network, characterized in that, The method comprises: acquiring power data to be analyzed, performing demand analysis on the power data to be analyzed to obtain target predicted demand; performing feature attribution analysis on the target predicted demand to obtain feature contribution degree, and generating demand prediction reasons according to the feature contribution degree; generating a power demand prediction result according to the target predicted demand and the demand prediction reasons.

2. The complex network and graph neural network based power demand forecasting method of claim 1, wherein, The demand analysis on the power data to be analyzed is performed by a power demand prediction model, and the power demand prediction model is constructed as follows: acquiring a multi-source time series feature data set, identifying the coupling relationship between the multi-source time series feature data set, and constructing a power demand complex network based on the multi-source time series feature data set and the coupling relationship; performing feature learning on the power demand complex network to generate a power demand prediction model; performing pruning processing on the power demand prediction model to obtain a lightweight prediction model. 3.The power demand prediction method based on complex network and graph neural network according to claim 2, wherein, The identification of the coupling relationship between the multi-source time series feature data set comprises: performing preprocessing on the multi-source time series feature data set to obtain a standard time series feature data set; calculating the comprehensive coupling degree between any two time series feature data of the standard time series feature data set; judging the comprehensive coupling degree according to a preset dynamic threshold to determine the initial relationship between the two time series feature data; performing multi-step lag analysis on the time series feature data with the initial relationship, and generating the final coupling relationship in combination with the initial relationship.

4. The complex network and graph neural network based power demand forecasting method of claim 3, wherein, The generation step of the dynamic threshold comprises: acquiring the comprehensive coupling degree between all time series feature groups to be analyzed within a time window; calculating the mean and standard deviation of the comprehensive coupling degree; generating a dynamic threshold based on the product of the standard deviation and a preset adjustable parameter, and combining the mean.

5. The complex network and graph neural network based power demand forecasting method of claim 2, wherein, The construction of the power demand complex network based on the multi-source time series feature data set and the coupling relationship comprises: regarding each data in the standard time series feature data set as a network node; based on the initial relationship, establishing network edges between nodes with strong correlation to form an initial demand network; assigning edge attributes to the network edges according to the coupling relationship to generate the power demand complex network.

6. The complex network and graph neural network-based power demand prediction method of claim 5, wherein, The feature learning on the power demand complex network to generate the power demand prediction model comprises: constructing a graph structure data based on the power demand complex network, and performing feature embedding learning by using a graph neural network to obtain a power demand prediction value; calculating the error between the power demand prediction value and the historical true value; based on the comparison result of the error and a preset threshold, deciding whether to fine-tune the graph neural network to generate the power demand prediction model.

7. The complex network and graph neural network-based power demand prediction method of claim 2, wherein, The pruning processing on the power demand prediction model to obtain a lightweight prediction model comprises: determining candidate pruning objects of the power demand prediction model and corresponding pruning granularity; based on the comparison of the importance measurement of the candidate pruning objects and a preset pruning threshold, screening target pruning objects; according to the pruning granularity, deleting the target pruning objects to generate the lightweight prediction model.

8. The complex network and graph neural network-based power demand prediction method of claim 7, wherein, The acquisition of the power data to be analyzed, the demand analysis on the power data to be analyzed, and the obtaining of the target predicted demand comprise: processing the power data to be analyzed into input data compatible with the power demand complex network; Extracting spatio-temporal joint features of input data by using the lightweight prediction model; Generating target prediction demand based on the spatio-temporal joint features. 9.The power demand prediction method based on complex network and graph neural network according to claim 7, wherein, Generating a power demand prediction result according to the target prediction demand and the demand prediction reason, including: Generating a plurality of perturbation samples by perturbing input data, and obtaining corresponding prediction results; Calculating sample weights based on the difference between the perturbation samples and the original data; Using a weighted regression method to fit a local linear model to explain the target prediction demand; Extracting feature contribution based on the local linear model, and identifying key features and their contribution attributes to generate demand prediction reasons.

10. The complex network and graph neural network based power demand forecasting method of claim 9, wherein, The power demand prediction result generated according to the target prediction demand and the demand prediction reason further includes: Conducting confidence analysis on the target prediction demand and the demand prediction reason; Generating a power demand prediction result containing actionable recommendations based on the confidence analysis result; When the confidence is lower than the preset threshold, additional risk labeling and mitigation suggestions are generated.

Images