Comprehensive energy system multi-energy load prediction method based on MSTL decomposition

Abstract

The invention provides a comprehensive energy system multi-energy load prediction method based on MSTL decomposition, and belongs to the technical field of electrical engineering, and the method comprises the steps: obtaining historical data of loads of various energy types; based on the historical data of each type of load, calculating the correlation between each type of load, and based on the correlation between each type of load, constructing a graph network; based on a multi-season-trend decomposition technology, performing time sequence decomposition on the historical data of each type of load to obtain a periodic feature sequence of multiple time scales in each piece of historical data; and inputting the graph network and each periodic feature sequence into a multi-energy load prediction model to obtain a multi-energy load prediction result. According to the comprehensive energy system multi-energy load prediction method based on MSTL decomposition provided by the invention, the spatial features and the time features of each type of load are integrated, the accurate load prediction result of the multi-type load is output, the prediction requirements of different seasons can be met, and the accuracy of multi-energy load prediction is improved.

Description

Technical Field

[0001] The present invention relates to the field of electrical engineering technology, and in particular to a multi-energy load forecasting method for an integrated energy system based on MSTL decomposition. Background Art

[0002] Integrated energy systems integrate multiple energy sources, including cooling, heating, electricity, and gas, to meet users' diverse energy needs. They are crucial for achieving complementary and flexible regulation of multiple energy sources, significantly improving the robustness of energy supply systems. Therefore, accurate load forecasting in integrated energy systems is a fundamental prerequisite for achieving optimized scheduling, coordinated planning, and energy conversion among multiple energy sources.

[0003] However, the strong coupling between multiple energy sources makes traditional methods of predicting the load size of a single energy type completely deviate from the actual needs of multi-energy complementarity. Furthermore, loads exhibit strong periodicity. Traditional forecasting methods only mine the surface periodicity of historical data and the underlying relationship between meteorological data and load size. These methods are unable to fully explore the multi-scale periodicity of multi-energy loads and the coupling relationships between multiple energy sources. Furthermore, due to the influence of climate temperature, the coupling strength and coupling relationships between multi-energy loads vary from season to season. Therefore, existing single, fixed forecasting models cannot meet the needs of year-round load forecasting, resulting in low accuracy in multi-energy load forecasting for integrated energy systems. Summary of the Invention

[0004] The present invention provides a multi-energy load forecasting method for an integrated energy system based on MSTL decomposition, which is used to solve the problem of low accuracy of multi-energy load forecasting for an integrated energy system in the prior art.

[0005] In a first aspect, the present invention provides a multi-energy load forecasting method for an integrated energy system based on MSTL decomposition, comprising:

[0006] Obtain historical data on loads of various energy types;

[0007] Based on the historical data of various types of loads, the correlation between various types of loads is calculated, and based on the correlation between various types of loads, a graph network is constructed;

[0008] Based on the multi-season-trend decomposition technology, the historical data of various types of loads are decomposed into time series to obtain the periodic characteristic sequences of multiple time scales in each historical data;

[0009] The graph network and each periodic feature sequence are input into a multi-energy load prediction model to obtain a multi-energy load prediction result output by the multi-energy load prediction model; the multi-energy load prediction model is obtained by model training based on graph network samples and periodic sequence samples.

[0010] In one embodiment, the calculation of correlations between load types based on historical data of load types, and the construction of a graph network based on the correlations between load types include:

[0011] Based on the historical data of each type of load, correlation calculation is performed between each type of load to obtain multiple correlation results; the correlation results include the correlation and correlation value between any two loads;

[0012] Identify each type of load as a node;

[0013] Determine each correlation as a connecting edge;

[0014] Determine each correlation value as the weight value of the connecting edge;

[0015] A graph network is constructed based on each node, each connecting edge, and the weight value of each connecting edge.

[0016] In one embodiment, when the historical data of each type of load is decomposed into time series based on the multi-season-trend decomposition technology to obtain periodic characteristic sequences of multiple time scales in each historical data, the following steps are performed for each type of load:

[0017] Decomposing the historical data corresponding to the current type of load into a time series to obtain multiple seasonal components, and determining the multiple seasonal components as the periodic characteristics of the current type of load at multiple time scales;

[0018] The historical data and the periodic features of the multiple time scales are spliced ​​together to obtain a periodic feature sequence of the multiple time scales in the historical data.

[0019] In one embodiment, the multi-energy load prediction model includes a spatial convolution layer, a temporal convolution layer, an adaptive spatial convolution layer, and an output layer;

[0020] The spatial convolution layer is used to extract spatial features from the graph network to obtain spatial features of each type of load;

[0021] The temporal convolution layer is used to extract temporal features from each periodic feature sequence to obtain the temporal features of each type of load;

[0022] The adaptive spatial convolution layer is used to aggregate the spatial features and temporal features of each type of load to obtain aggregated features of each type of load;

[0023] The output layer is used to perform load forecasting based on the aggregated features of each type of load to obtain a multi-energy load forecasting result.

[0024] In one embodiment, the multi-energy load forecasting model is trained according to the following method:

[0025] Obtain historical data samples and load labels for various types of load samples;

[0026] Based on the historical data samples of each type of load sample, the correlation between each type of load sample is calculated, and based on the correlation between each type of load sample, a graph network sample is constructed;

[0027] Based on the multi-season-trend decomposition technology, the historical data samples of various types of load samples are decomposed into time series to obtain periodic series samples of multiple time scales in each historical data sample;

[0028] Model training is performed based on the graph network samples, the periodic sequence samples, and the load labels of various types to obtain the multi-energy load prediction model.

[0029] The present invention provides a multi-energy load forecasting method for an integrated energy system based on MSTL decomposition. Based on the historical data of various types of loads, a graph network is constructed. The correlation and influence between various loads can be captured through the historical data, and the spatial feature mining of various loads can be realized. Furthermore, based on the multi-season-trend decomposition technology, the historical data are decomposed into time series to obtain periodic feature sequences of multiple time scales in each historical data. The periodic fluctuations and trends on different time scales can be extracted to realize the multi-scale periodic time feature mining of various loads. Furthermore, through the multi-energy load forecasting model, the spatial and temporal features of each type of load are integrated to output accurate load forecasting results for loads of multiple energy types, which can meet the forecasting needs of different seasons and improve the accuracy of multi-energy load forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction is given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0031] Figure 1 This is a flow chart of a multi-energy load forecasting method for an integrated energy system based on MSTL decomposition provided by the present invention;

[0032] Figure 2 It is a structural diagram of the graph network provided by the present invention;

[0033] Figure 3 It is a functional schematic diagram of the dynamic graph convolutional neural network provided by the present invention. DETAILED DESCRIPTION

[0034] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0035] The following combination Figure 1-Figure 3 The present invention describes a multi-energy load forecasting method for an integrated energy system based on MSTL decomposition.

[0036] Figure 1 It is a flow chart of a multi-energy load forecasting method for an integrated energy system based on MSTL decomposition provided by the present invention.

[0037] like Figure 1 As shown, the present invention provides a multi-energy load forecasting method for an integrated energy system based on MSTL decomposition, including but not limited to the following steps:

[0038] Obtain historical data on loads of various energy types;

[0039] Based on the historical data of various types of loads, the correlation between various types of loads is calculated, and based on the correlation between various types of loads, a graph network is constructed;

[0040] Based on the multi-season-trend decomposition technology, the historical data of various types of loads are decomposed into time series to obtain the periodic characteristic sequences of multiple time scales in each historical data;

[0041] The graph network and each periodic characteristic sequence are input into a multi-energy load prediction model to obtain a multi-energy load prediction result output by the multi-energy load prediction model.

[0042] It should be noted that the multi-energy load forecasting method for an integrated energy system based on MSTL decomposition provided in the embodiments of the present invention is applied in the energy field. It is primarily targeted at the multiple energy types integrated in the integrated energy system, such as cooling, heating, electricity, and gas, to predict the load demand of each energy type over a period of time in the future. This embodiment of the present invention describes a multi-energy load forecasting method for an integrated energy system based on MSTL decomposition, using a multi-energy load forecasting device as an example.

[0043] Specifically, the multi-energy load forecasting device obtains historical data of loads of multiple energy types.

[0044] Furthermore, the multi-energy load forecasting device determines the correlation between the various types of loads based on the historical data of the various types of loads. Furthermore, the multi-energy load forecasting device constructs a graph network based on the correlation between the various types of loads. It should be noted that the embodiments of the present invention are directed to loads of multiple energy types and, accordingly, also to multiple types of loads.

[0045] It should be noted that graph networks are machine learning models used to process graph data. In a graph network, nodes represent entities, and edges represent the relationships or connections between these entities. Therefore, graph networks can effectively capture the relationships between nodes and the information contained in the graph structure. By constructing a graph network, the multi-energy load forecasting device can more comprehensively consider the correlations between different energy sources and different loads, helping to improve the accuracy of multi-energy load forecasting.

[0046] Furthermore, based on the multi-season-trend decomposition technology, the historical data of each type of load are decomposed into time series to obtain the periodic characteristic sequences of multiple time scales in the historical data of each type of load.

[0047] It should be noted that the time series decomposition method uses Multiple Seasonal-Trend decomposition using Loess (MSTL). MSTL is a method for time series decomposition, primarily used to process time series data with multiple seasonal cycles. Through time series decomposition, periodic characteristics at different time scales can be obtained to form periodic characteristic sequences at different time scales. The multi-energy load forecasting device decomposes the time series of historical data to discover the periodic characteristic sequences at multiple time scales hidden in the historical data. Therefore, combining the changes in load at different time scales helps to improve the accuracy of multi-energy load forecasting.

[0048] Furthermore, the multi-energy load prediction device inputs the graph network and each periodic feature sequence into the multi-energy load prediction model to obtain a multi-energy load prediction result output by the multi-energy load prediction model, wherein the multi-energy load prediction model is obtained by model training based on graph network samples and periodic sequence samples.

[0049] The present invention provides a multi-energy load forecasting method for an integrated energy system based on MSTL decomposition. Based on the historical data of various types of loads, a graph network is constructed. The correlation and influence between various loads can be captured through the historical data, and the spatial feature mining of various loads can be realized. Furthermore, based on the multi-season-trend decomposition technology, the historical data are decomposed into time series to obtain periodic feature sequences of multiple time scales in each historical data. The periodic fluctuations and trends on different time scales can be extracted to realize the multi-scale periodic time feature mining of various loads. Furthermore, through the multi-energy load forecasting model, the spatial and temporal features of each type of load are integrated to output accurate load forecasting results for loads of multiple energy types, which can meet the forecasting needs of different seasons and improve the accuracy of multi-energy load forecasting. In one embodiment, after acquiring historical load data for multiple energy types, and before constructing a graph network and decomposing periodic feature sequences, all historical data can be preprocessed, including the following steps: cleaning all historical data to obtain multiple cleaned data sets; normalizing each cleaned data set to obtain multiple normalized data sets; identifying outliers and missing points in each normalized data set; removing outliers and interpolating missing points to obtain multiple historical data sets. Preprocessing all historical data helps improve the quality and reliability of the historical data, thereby enabling more accurate data analysis and modeling.

[0050] Furthermore, the calculation of the correlation between the various types of loads based on the historical data of the various types of loads, and the construction of a graph network based on the correlation between the various types of loads include:

[0051] Based on the historical data of each type of load, correlation calculation is performed between each type of load to obtain multiple correlation results; the correlation results include the correlation and correlation value between any two loads;

[0052] Identify each type of load as a node;

[0053] Determine each correlation as a connecting edge;

[0054] Determine each correlation value as the weight value of the connecting edge;

[0055] A graph network is constructed based on each node, each connecting edge, and the weight value of each connecting edge.

[0056] Specifically, the multi-energy load prediction device calculates the correlation between various types of loads based on all historical data to obtain multiple correlation results, wherein the correlation results include the correlation and correlation value between any two loads.

[0057] Therefore, it can be understood that the multi-energy load prediction device calculates the correlation between any two loads based on the historical data of various types of loads, obtains the correlation and correlation value between each two loads, and thereby characterizes the coupling relationship and coupling strength between each two loads.

[0058] It should be noted that commonly used correlation calculation methods include Pearson correlation coefficient and Spearman rank correlation coefficient, etc., which are not limited here, and the most appropriate calculation method can be selected according to the specific situation.

[0059] Furthermore, the multi-energy load forecasting device identifies each type of load as a node, each correlation as an edge, and each correlation value as an edge weight. Furthermore, the multi-energy load forecasting device constructs a graph network based on each node, each edge, and each edge weight. Therefore, in the constructed graph network, nodes represent loads, edges represent correlations between nodes, and edge weights represent quantified correlation values.

[0060] In one embodiment, referring to Figure 2 , Figure 2 This is a schematic diagram of the structure of the graph network provided by the present invention, in which the load sets {L1, L2, L3, L4} of the gas load, electric load, heat load and cooling load of the integrated energy system are constructed into a graph network. The load sets {L1, L2, L3, L4} are used as the node sets {N1, N2, N3, N4} of the graph network, and the correlations between the load sets {L1, L2, L3, L4} are used as the edge sets {E 1,2 、E 1,3 、E 1,4 、E 2,3 、E 2,4 、E 3,4}, the size of each connecting edge is determined by the quantized correlation value.

[0061] The embodiment of the present invention calculates the correlation between various types of loads based on all historical data, obtains multiple correlation results, and further constructs a graph network based on the correlation results, which can capture the associations and influences between various loads, realize spatial feature mining of various loads, and then combine the time characteristics of each type of load to determine the accurate load prediction results of loads of multiple energy types, thereby improving the accuracy of multi-energy load prediction.

[0062] Furthermore, based on the multi-seasonal-trend decomposition technology, the historical data of each type of load is decomposed into time series to obtain the periodic characteristic sequences of multiple time scales in each historical data. The following steps are performed for each type of load:

[0063] Decomposing the historical data corresponding to the current type of load into a time series to obtain a plurality of seasonal components, and determining the plurality of seasonal components as the periodic characteristics of the current type of load at a plurality of time scales;

[0064] The historical data and the periodic features of the multiple time scales are spliced ​​together to obtain a periodic feature sequence of the multiple time scales in the historical data.

[0065] Specifically, the multi-energy load forecasting device decomposes the historical data corresponding to the current type of load into a time series to obtain multiple seasonal components, and determines the multiple seasonal components as the periodic characteristics of the current type of load at multiple time scales.

[0066] It should be noted that the seasonal component reflects the cyclical changes in time series data that recur within a fixed period, while the random component reflects the irregular fluctuations in time series data that cannot be explained by trend and seasonality. Therefore, based on the seasonal component, cyclical characteristics at multiple time scales can be determined, such as daily, weekly, monthly, and annual cycles, thereby better capturing the cyclical changes in load at different time scales.

[0067] In one embodiment, the time series decomposition is performed based on the historical data corresponding to the current type of load L1 to obtain the periodic characteristics of the current type of load L1 at multiple time scales {T 1,1 、T 1,2 、T 1,3 ,…,T 1,m}, where m represents m time scales.

[0068] Furthermore, the multi-energy load forecasting device combines the historical data corresponding to the current type of load with the periodic characteristics of multiple time scales to obtain a periodic characteristic sequence of multiple time scales in the historical data of the current type of load.

[0069] In one embodiment, the historical data corresponding to the current type load L1 is T 1,0 , the periodic characteristics of various time scales are {T 1,1 、T 1,2 、T 1,3 ,…,T 1,m}, so the periodic characteristic sequences of various time scales obtained by splicing the two are {T 1,0 、T 1,1 、T 1,2 、T 1,3 ,…,T 1,m}.

[0070] The embodiment of the present invention performs time series decomposition based on the historical data corresponding to each type of load, obtains the periodic characteristics of each type of load at multiple time scales, further splices the historical data with the periodic characteristics of multiple time scales, and obtains a periodic feature sequence of multiple time scales. This strengthens the periodic characteristics of multiple time scales in the historical data of each type of load, can capture the periodic fluctuations and trends of the load at different time scales, and realizes the multi-scale periodic time feature mining of the load. Then, combined with the spatial characteristics of each type of load, it determines the accurate load forecast results of multiple energy types, thereby improving the accuracy of multi-energy load forecasting. In addition, it can meet the forecasting needs of different seasons while fully mining the potential multiple periodic characteristics of the load.

[0071] Furthermore, the multi-energy load prediction model includes a spatial convolution layer, a temporal convolution layer, an adaptive spatial convolution layer and an output layer;

[0072] The spatial convolution layer is used to extract spatial features from the graph network to obtain spatial features of each type of load;

[0073] The temporal convolution layer is used to extract temporal features from each periodic feature sequence to obtain the temporal features of each type of load;

[0074] The adaptive spatial convolution layer is used to aggregate the spatial features and temporal features of each type of load to obtain the aggregated features of each type of load;

[0075] The output layer is used to perform load forecasting based on the aggregated features of each type of load, and obtain load forecasting results for each type of load corresponding to each energy source.

[0076] Specifically, the multi-energy load prediction device inputs the graph network and each periodic feature sequence into the multi-energy load prediction model to obtain the multi-energy load prediction result output by the multi-energy load prediction model, that is, the load prediction result of loads of multiple energy types, wherein the multi-energy load prediction model includes a spatial convolution layer, a temporal convolution layer, an adaptive spatial convolution layer and an output layer.

[0077] Therefore, it can be understood that after the multi-energy load prediction device inputs the graph network and each periodic feature sequence into the multi-energy load prediction model, the spatial convolution layer extracts the spatial features of the graph network, obtains the spatial features of each type of load, and outputs the spatial features of each type of load.

[0078] Furthermore, the time convolution layer extracts the time features of each periodic feature sequence respectively, obtains the time features of each type of load, and outputs the time features of each type of load.

[0079] Furthermore, the adaptive spatial convolution layer aggregates the spatial features and temporal features of each type of load output to obtain the aggregated features of each type of load, and outputs the aggregated features of each type of load.

[0080] Furthermore, the output layer performs load prediction based on the aggregated features of each type of load outputted, obtains the load prediction result of each type of load, and outputs the load prediction result of each type of load as a model prediction result.

[0081] In one embodiment, the model may adopt a dynamic graph convolutional neural network (DyGCN) model, such as Figure 3 As shown, Figure 3 This is a functional schematic diagram of the dynamic graph convolutional neural network provided by the present invention, which is a deep learning model for processing dynamic graph data, and can model the representation learning on the dynamic graph as the aggregation of time and space information. The spatial convolution layer and the temporal convolution layer of the DyGCN model can realize the extraction of spatial information and spatiotemporal information in historical data. In addition, an adaptive mechanism is introduced in the spatial convolution layer to form an adaptive spatial convolution layer, which can update the convolution weights as the graph network changes, so that the spatial convolution layer can dynamically extract the graph network information of different time nodes. The bottom layer uses a dynamic graph convolutional neural network as an unsupervised learning method to extract abstract features from historical data, and the top layer uses multi-task learning as a supervised learning method to output the load prediction results of each type of load corresponding to each energy source. Multi-task learning is based on the inductive transfer mechanism, which obtains knowledge of other related problems by sharing information, and trains multiple tasks in parallel through a sharing mechanism, so as to fully utilize the potential characteristics of other types of loads in the final prediction stage.

[0082] The embodiment of the present invention obtains the spatial characteristics of each type of load by spatial feature extraction, obtains the temporal characteristics of each type of load by temporal feature extraction, and further improves the accuracy and efficiency of load analysis and prediction by effectively integrating the spatial characteristics and temporal characteristics of each type of load.

[0083] Furthermore, the multi-energy load forecasting model is trained according to the following method:

[0084] Obtain historical data samples and load labels for various types of load samples;

[0085] Based on the historical data samples of each type of load sample, the correlation between each type of load sample is calculated, and based on the correlation between each type of load sample, a graph network sample is constructed;

[0086] Based on the multi-season-trend decomposition technology, the historical data samples of various types of load samples are decomposed into time series to obtain periodic series samples of multiple time scales in each historical data sample;

[0087] Model training is performed based on the graph network samples, the periodic sequence samples, and the load labels of various types to obtain the multi-energy load prediction model.

[0088] Specifically, the multi-energy load forecasting device obtains historical data samples and load labels of multiple types of load samples.

[0089] Furthermore, the multi-energy load prediction device determines the correlation between each type of load samples based on the historical data samples of each type of load samples. Furthermore, the multi-energy load prediction device constructs a graph network sample based on the correlation between each type of load samples.

[0090] Furthermore, the multi-energy load forecasting device decomposes the historical data samples of each type of load sample into time series based on the multi-season-trend decomposition technology to obtain trend component samples, seasonal component samples and random component samples corresponding to each type of load sample.

[0091] Furthermore, the multi-energy load forecasting device determines a multi-period sequence sample of each type of load sample based on the trend component sample, seasonal component sample and random component sample corresponding to each type of load sample.

[0092] Furthermore, the multi-energy load forecasting device respectively splices the historical data samples corresponding to each type of load sample with the multi-period sequence samples to obtain periodic sequence samples of multiple time scales corresponding to the load samples.

[0093] Furthermore, the multi-energy load prediction device performs model training based on graph network samples, periodic sequence samples, and load labels of each type of load sample to obtain a multi-energy load prediction model.

[0094] It should be noted that before model training, the graph network samples and periodic sequence samples of each type of load sample will be divided into training set, validation set and test set according to the preset ratio, wherein the preset ratio is set according to the actual situation. The test set and validation set are input into the training model, and it is iterated multiple times according to the convergence effect to complete the training of the training model and the self-adjustment of the model parameters. Then, the validation set is input into the trained model to test the prediction accuracy and response speed of the trained model. The test set is further input into the trained model to test the prediction accuracy of the test set in the trained model and compare it with the prediction accuracy corresponding to the validation set to determine the degree of fit of the model and correct the model parameters to obtain the optimal prediction model.

[0095] The embodiment of the present invention obtains graph network samples and periodic sequence samples of each type of load sample through correlation calculation and MSTL decomposition, and further performs model training based on the graph network samples, periodic sequence samples and load labels of each type of load sample to obtain a multi-energy load prediction model, thereby realizing intelligent multi-energy load prediction and being able to output accurate load prediction results for all loads corresponding to each energy source, thereby improving the accuracy of multi-energy load prediction.

Claims

1. A multi-energy load forecasting method for an integrated energy system based on MSTL multi-season-trend decomposition, characterized in that: include: Obtain historical data on loads of various energy types; Based on the historical data of various types of loads, the correlation between various types of loads is calculated, and based on the correlation between various types of loads, a graph network is constructed; Based on the multi-season-trend decomposition technology, the historical data of various types of loads are decomposed into time series to obtain the periodic characteristic sequences of multiple time scales in each historical data; Inputting the graph network and each periodic characteristic sequence into a multi-energy load forecasting model to obtain a multi-energy load forecasting result output by the multi-energy load forecasting model; The multi-energy load forecasting model is obtained by training the model based on graph network samples and periodic sequence samples.

2. The method for multi-energy load forecasting of an integrated energy system based on MSTL decomposition according to claim 1 is characterized in that: The process of calculating the correlation between the load types based on the historical data of the load types and constructing a graph network based on the correlation between the load types includes: Based on the historical data of each type of load, correlation calculation is performed between each type of load to obtain multiple correlation results; the correlation results include the correlation and correlation value between any two loads; Identify each type of load as a node; Determine each correlation as a connecting edge; Determine each correlation value as the weight value of the connecting edge; A graph network is constructed based on each node, each connecting edge, and the weight value of each connecting edge.

3. The method for multi-energy load forecasting of an integrated energy system based on MSTL decomposition according to claim 1 is characterized in that: Based on the multi-seasonal-trend decomposition technology, the historical data of each type of load is decomposed into time series to obtain the periodic characteristic sequences of multiple time scales in each historical data. The following steps are performed for each type of load: Decomposing the historical data corresponding to the current type of load into a time series to obtain multiple seasonal components, and determining the multiple seasonal components as the periodic characteristics of the current type of load at multiple time scales; The historical data and the periodic features of the multiple time scales are spliced ​​together to obtain a periodic feature sequence of the multiple time scales in the historical data.

4. The method for multi-energy load forecasting of an integrated energy system based on MSTL decomposition according to claim 1 is characterized in that: The multi-energy load prediction model includes a spatial convolution layer, a temporal convolution layer, an adaptive spatial convolution layer and an output layer; The spatial convolution layer is used to extract spatial features from the graph network to obtain spatial features of each type of load; The temporal convolution layer is used to extract temporal features from each periodic feature sequence to obtain the temporal features of each type of load; The adaptive spatial convolution layer is used to aggregate the spatial features and temporal features of each type of load to obtain the aggregated features of each type of load; The output layer is used to perform load forecasting based on the aggregated features of each type of load to obtain a multi-energy load forecasting result.

5. The method for multi-energy load forecasting of an integrated energy system based on MSTL decomposition according to claim 1 is characterized in that: The multi-energy load forecasting model is trained according to the following method: Obtain historical data samples and load labels for various types of load samples; Based on the historical data samples of each type of load sample, the correlation between each type of load sample is calculated, and based on the correlation between each type of load sample, a graph network sample is constructed; Based on the multi-season-trend decomposition technology, the historical data samples of various types of load samples are decomposed into time series to obtain periodic series samples of multiple time scales in each historical data sample; Model training is performed based on the graph network samples, the periodic sequence samples, and the load labels of various types to obtain the multi-energy load prediction model.

Images