Multi-element load hybrid prediction method, system and equipment of integrated energy system and medium

Abstract

The invention relates to the technical field of multi-element load prediction, and provides a multi-element load hybrid prediction method, system, device and medium for an integrated energy system, and the method comprises the steps: splicing a to-be-analyzed multi-element load sequence generated by processing a multi-element load sequence and an influence factor sequence, and the influence factor sequence to obtain a first multi-element load feature sequence; analyzing the first multi-element load characteristic sequence based on a correlation evaluation model to generate an associated load group and an independent load group, and splicing a reconstructed multi-element load sequence obtained by performing singular spectrum analysis on a to-be-analyzed multi-element load sequence with the to-be-analyzed multi-element load sequence and an influence factor sequence to obtain a second multi-element load characteristic sequence; and performing multi-element load hybrid prediction based on the associated load group and the independent load group to obtain a multi-element load prediction result according to the multi-element load prediction input characteristics obtained by the double attention mechanism fusion. According to the method, complex dynamic coupling characteristics among multiple loads can be fully captured, and the accuracy and robustness of load prediction are improved.

Description

Technical Field

[0001] This invention relates to the field of load forecasting technology for integrated energy systems, and in particular to a method, system, equipment, and medium for mixed forecasting of multiple loads in integrated energy systems. Background Technology

[0002] Multi-load forecasting is a key support for Integrated Energy Systems (IES) to improve energy efficiency and power supply reliability through multi-energy complementarity and synergistic optimization. However, the complex coupling relationships between various heterogeneous loads such as electricity, heat, and gas in IES make accurate multi-load forecasting difficult.

[0003] Existing multi-load forecasting methods, in order to effectively utilize the coupling relationships between multiple loads, typically perform correlation analysis on the loads beforehand using the maximum information coefficient (MIT), statically dividing the loads into "strongly correlated" and "weakly correlated" groups. Multi-task learning (MTL) is then employed to capture the coupling characteristics between loads within the "strongly correlated" group, and single-task learning (STL) is used to handle specific "weakly correlated" loads. However, the correlation of multiple loads in actual integrated energy systems is not static but is influenced by various factors such as seasonal variations, daily type differences, unforeseen events, and adjustments to operating modes, exhibiting dynamic characteristics. For example, two loads may be highly correlated in summer, but their correlation may significantly weaken or even become negatively correlated in winter. This can lead to unexpected new interferences between loads initially considered "weakly correlated" and other loads during certain periods, while the coupling relationship between loads initially considered "strongly correlated" may weaken at other times. In such cases, the pre-defined static grouping is simply unable to adapt to these dynamic changes, making MTL not only ineffective in synergistic effects but also introducing noise, thereby reducing forecast accuracy and model robustness. Summary of the Invention

[0004] The purpose of this invention is to provide a method for predicting multiple loads in an integrated energy system. By introducing a periodic dynamic load correlation assessment, the method fully captures the dynamic coupling characteristics changes among multiple loads and adaptively adjusts the load co-prediction strategy accordingly. This effectively avoids the negative impact of inaccurate load coupling relationship identification on the overall prediction performance and improves the accuracy and robustness of multiple load prediction in the integrated energy system.

[0005] To achieve the above objectives, it is necessary to provide a method, system, equipment, and medium for predicting multiple loads in an integrated energy system, addressing the aforementioned technical problems.

[0006] In a first aspect, embodiments of the present invention provide a method for predicting the mixed loads of an integrated energy system, the method comprising: The multivariate load series and influencing factor series are periodically acquired and preprocessed to generate the multivariate load series and influencing factor series to be analyzed. The multivariate load sequence to be analyzed is spliced ​​together with the influencing factor sequence in time alignment to obtain the first multivariate load characteristic sequence; Based on the first multivariate load characteristic sequence, load correlation analysis is performed based on a pre-constructed correlation assessment model to generate a load correlation matrix, and associated load groups and independent load groups are generated based on the load correlation matrix. The multivariate load sequence to be analyzed is subjected to singular spectrum decomposition and reconstruction to obtain a reconstructed multivariate load sequence. The reconstructed multivariate load sequence, the multivariate load sequence to be analyzed, and the influencing factor sequence are time-aligned and spliced ​​together to obtain a second multivariate load feature sequence. Each load feature sequence in the second multivariate load feature sequence is fused using a dual attention mechanism to obtain the multivariate load prediction input features. Based on the input characteristics of the multi-variable load forecast, multi-variable load mixed forecast is performed based on the associated load group and the independent load group to obtain the multi-variable load forecast result.

[0007] Furthermore, the relevance assessment model includes a feature extraction layer and a similarity analysis layer connected in sequence; the feature extraction layer includes two feature extraction sub-networks with shared weights; the feature extraction sub-network includes a convolutional layer and a recurrent coding layer.

[0008] Furthermore, the similarity analysis layer is used to perform cosine similarity analysis on the two load feature vectors output by the feature extraction layer to obtain the corresponding load correlation score.

[0009] Further, the step of generating associated load groups and independent load groups based on the load correlation matrix includes: Based on the load correlation matrix, loads whose load correlation scores with other loads are all less than a preset correlation score threshold are obtained, and the independent load group is generated; the preset correlation score threshold is adaptively adjusted and optimized based on a grid search method. The associated load group is generated by excluding the loads in the independent load group from the multi-load data.

[0010] Furthermore, the step of performing singular spectral decomposition and reconstruction on the multivariate load sequence to be analyzed to obtain the reconstructed multivariate load sequence includes: Each load sequence to be analyzed in the multivariate load sequence to be analyzed is subjected to singular spectral decomposition to obtain multiple corresponding frequency load subsequences; Based on the contribution of subsequence variance, the frequency load subsequences of each load sequence to be analyzed are screened and reconstructed to obtain the corresponding reconstructed load sequences. All the reconstructed load sequences are summarized to obtain the reconstructed multivariate load sequence.

[0011] Further, the step of performing multi-factor load mixed forecasting based on the associated load group and the independent load group according to the multi-factor load forecasting input characteristics to obtain the multi-factor load forecasting result includes: Based on all load prediction input features corresponding to the associated load group in the multivariate load prediction input features, collaborative prediction is performed based on a pre-built multi-task learning model to obtain the predicted value of each load in the associated load group. Each load prediction input feature corresponding to the independent load group in the multivariate load prediction input features is input into the corresponding single-task learning model for individual prediction to obtain the predicted value of each load in the independent load group. The predicted values ​​of each load in the associated load group and the predicted values ​​of each load in the independent load group are summarized to obtain the multi-variable load prediction result.

[0012] Further, the step of obtaining the predicted values ​​of each load in the associated load group by performing collaborative prediction based on a pre-built multi-task learning model using all load prediction input features corresponding to the associated load group from the multi-source load prediction input features includes: Based on a preset shared encoder, feature encoding is performed on each load prediction input feature corresponding to the associated load group to obtain the corresponding load encoding features; The corresponding collaborative coding features are obtained by weighted summing of all relevant load coding features corresponding to each load coding feature based on the corresponding load correlation scores in the load correlation matrix. Each of the load coding features and the corresponding co-coding features are concatenated and fused to obtain the corresponding load fusion feature. The load fusion feature is then input into the corresponding decoder for load prediction to obtain the predicted value of the corresponding load.

[0013] Secondly, embodiments of the present invention provide a multi-load hybrid forecasting system for integrated energy systems, the system comprising: The data acquisition module is used to periodically acquire multivariate load sequences and influencing factor sequences, perform preprocessing, and generate multivariate load sequences and influencing factor sequences to be analyzed. The first splicing module is used to align and splice the multivariate load sequence to be analyzed with the influencing factor sequence to obtain the first multivariate load characteristic sequence; The correlation analysis module is used to perform correlation analysis between loads based on the first multivariate load characteristic sequence and a pre-built correlation evaluation model, generate a load correlation matrix, and generate associated load groups and independent load groups based on the load correlation matrix. The second splicing module is used to perform singular spectrum decomposition and reconstruction on the multivariate load sequence to be analyzed to obtain a reconstructed multivariate load sequence, and to align and splice the reconstructed multivariate load sequence, the multivariate load sequence to be analyzed and the influencing factor sequence to obtain a second multivariate load feature sequence. The fusion analysis module is used to perform dual attention mechanism fusion analysis on each load feature sequence in the second multivariate load feature sequence to obtain multivariate load prediction input features; The load forecasting module is used to perform multi-load mixed forecasting based on the associated load group and the independent load group according to the multi-load forecasting input characteristics, and obtain multi-load forecasting results.

[0014] Thirdly, embodiments of the present invention also provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method.

[0015] Fourthly, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0016] This invention provides a method, system, device, and medium for hybrid forecasting of multiple loads in an integrated energy system. The method involves periodically acquiring multiple load sequences and influencing factor sequences, preprocessing them to generate sequences of multiple loads to be analyzed and sequences of influencing factors, then temporally aligning and concatenating the sequences of multiple loads to be analyzed and the sequences of influencing factors to obtain a first multiple load characteristic sequence. Based on the first multiple load characteristic sequence, a load correlation matrix is ​​generated using a pre-constructed correlation assessment model, and associated load groups and independent load groups are generated based on the load correlation matrix. Singular spectrum decomposition and reconstruction are then performed on the multiple load sequence to be analyzed to obtain a reconstructed multiple load sequence. The reconstructed multiple load sequence, the multiple load sequence to be analyzed, and the influencing factor sequences are temporally aligned and concatenated to obtain a second multiple load characteristic sequence. Each load characteristic sequence in the second multiple load characteristic sequence is then subjected to a dual-attention mechanism fusion analysis to obtain multiple load forecasting input features. Finally, based on the multiple load forecasting input features, hybrid forecasting of multiple loads is performed using associated load groups and independent load groups to obtain the multiple load forecasting result. Compared with existing technologies, this multi-load hybrid forecasting method for integrated energy systems, based on the introduction of periodic dynamic load correlation assessment to fully capture the dynamic coupling characteristics changes among multiple loads and adaptively adjust the load collaborative forecasting strategy accordingly, can effectively avoid the negative impact of inaccurate load coupling relationship identification on the overall forecasting performance and improve the accuracy and robustness of multi-load forecasting for integrated energy systems. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the multi-load hybrid prediction method for integrated energy systems in an embodiment of the present invention. Figure 2 This is a schematic diagram comparing the prediction results of the prediction method proposed in this invention with the prediction model based on LSTM or GRU multi-task learning architecture for predicting and analyzing the cooling load, gas load, heat load and electrical load in an integrated energy system in an embodiment of this invention. Figure 3 This is a schematic diagram of the structure of the integrated energy system multi-load hybrid prediction system in an embodiment of the present invention; Figure 4 This is an internal structural diagram of the computer device in an embodiment of the present invention; The attached figures are labeled as follows: 1. Data acquisition module; 2. First stitching module; 3. Correlation analysis module; 4. Second stitching module; 5. Fusion analysis module; 6. Load forecasting module. Detailed Implementation

[0018] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Obviously, the embodiments described below are only part of the embodiments of this invention and are used to illustrate the invention, but are not intended to limit the scope of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0019] The integrated energy system multi-load hybrid forecasting method provided by this invention can be understood as addressing the current application status of existing integrated energy system multi-load forecasting, which uses static grouping and struggles to dynamically capture the complex coupling relationships between multiple loads, resulting in poor load forecasting performance. This invention proposes a multi-load hybrid forecasting method that periodically and dynamically captures the complex coupling relationships between multiple loads and adaptively adjusts the load collaborative forecasting strategy. While ensuring the high efficiency of multi-load forecasting, it significantly improves the accuracy and robustness of load forecasting. The following embodiments will provide a detailed description of the integrated energy system multi-load hybrid forecasting method of this invention.

[0020] In one embodiment, such as Figure 1 As shown, a method for predicting multiple loads in an integrated energy system is provided, including the following steps: S11. Periodically acquire and preprocess the multi-variable load sequence and influencing factor sequence to generate the multi-variable load sequence and influencing factor sequence to be analyzed. The multi-variable load sequence can be understood as the historical load time series data of the integrated energy system acquired based on the current forecast time and forecast time window (the most recent historical time series data is traced back from the current forecast time. For example, in a scenario where forecasts are made once per hour, if the forecast time point is t, the forecast time window can be set to [t-24, t], that is, the data of the past 24 hours is used for subsequent dynamic analysis). It can include cooling load sequence, heating load sequence, gas load sequence, and electricity load sequence, etc. The size of the forecast time window can be determined according to the length of the data time interval required for the actual load forecast task. The corresponding influencing factor sequence can be understood as the load-related influencing factor sequence that will affect the actual load. It can include meteorological data sequence (including temperature, humidity, air pressure, and wind speed, etc.), timestamp sequence, holiday information sequence, and operating status sequence, etc., and the operating status includes the start and stop status of cooling / heating equipment.

[0021] In this embodiment, the process of obtaining the multivariate load sequence and influencing factor sequence to be analyzed can be understood as performing data cleaning, missing value imputation, outlier correction, time window alignment, and normalization on each sequence in the multivariate load sequence and influencing factor sequence to obtain a multidimensional sequence that can be used for subsequent analysis. The specific processing process can be referred to relevant existing technologies. It should be noted that the timestamp information in the influencing factor sequence can be represented by one-hot encoding for information such as hour, year, month, and day; holiday information is represented in binary form, such as 1 for holidays and 0 for non-holidays.

[0022] S12. The multivariate load sequence to be analyzed and the influencing factor sequence are time-aligned and spliced ​​together to obtain a first multivariate load feature sequence; wherein, the first multivariate load feature sequence includes multiple load feature sequences, and each load feature sequence can be understood as a data sequence obtained by horizontally splicing the data of each load sequence in the multivariate load sequence to be analyzed and the data of each factor sequence in the influencing factor sequence at the same time, and the complete feature data of each time in the data sequence includes both load feature value and value of each influencing factor.

[0023] S13. Based on the first multivariate load characteristic sequence, load correlation analysis is performed based on a pre-constructed correlation assessment model to generate a load correlation matrix. Then, based on the load correlation matrix, associated load groups and independent load groups are generated. The load correlation matrix can be understood as a matrix that dynamically maintains the results of the load correlation analysis. It is obtained by pairing load characteristic sequences in the first multivariate load characteristic sequence for correlation analysis. To ensure the efficiency and accuracy of the load correlation analysis, this embodiment preferably proposes a load correlation assessment model based on deep learning. Considering the characteristics of load changes being related to influencing factors, load influencing factors are introduced for correlation analysis to improve the comprehensiveness and reliability of the load correlation analysis.

[0024] In practical applications, the load feature sequences in the first multivariate load feature sequence are paired and input into the correlation evaluation model to evaluate the correlation between the two loads until the correlation evaluation results of all loads are obtained to generate the required load correlation matrix. Specifically, the correlation evaluation model includes a feature extraction layer and a similarity analysis layer connected in sequence. The feature extraction layer includes two feature extraction sub-networks with shared weights, each of which is used to extract feature vectors for one type of load feature sequence. The feature extraction sub-network includes convolutional layers and recurrent encoding layers to extract local and global temporal patterns of load task features. Among them, stacked one-dimensional convolutional neural networks (1D-CNN) can be used to enhance the local feature extraction capability. A sliding window method is used to extract the load change pattern along the time axis. The parameters can be adjusted and optimized according to the complexity of the actual task or the characteristics of the data. That is, the number of stacked one-dimensional convolutional neural networks can be set according to the actual application requirements. For example, when stacking two layers of one-dimensional convolutional neural networks, the kernel size can be set to 3 and 64 respectively. In addition, each convolutional layer is followed by a ReLU activation function and a normalization layer to improve the nonlinear modeling capability and improve the model convergence efficiency. The recurrent encoding layer can be set as an LSTM recurrent layer with multiple stacked layers, and each LSTM recurrent layer is a complete LSTM network layer, used to capture long-term dependencies in the load time series. The number of layers and the dimension of hidden states can be adjusted according to the task complexity and modeling requirements. For example, when a 2-layer stacked structure is used, the number of neurons in each hidden layer can be set to 128, and the corresponding activation function can be tanh. The state vector output by the last hidden layer is used to generate the required load feature vector.

[0025] The similarity analysis layer in the correlation assessment model can be understood as performing correlation analysis on the two load feature vectors output by the two feature extraction sub-networks in the feature extraction layer of the correlation assessment model to obtain a correlation score between 0 and 1, reflecting the correlation strength between the two load forecasting tasks in the current time period, and effectively capturing complex nonlinear dynamic coupling relationships. Specifically, the similarity analysis layer is used to perform cosine similarity analysis on the two load feature vectors output by the feature extraction layer to obtain the corresponding load correlation score.

[0026] This embodiment, by introducing multiple load influencing factors and employing a correlation assessment model to evaluate the correlation between loads, can efficiently and accurately capture the dynamic coupling relationship between loads, providing a reliable basis for subsequent load forecasting task grouping.

[0027] The load correlation matrix obtained through the above methods and steps will be dynamically adjusted based on the actual multivariate load sequence and influencing factor sequence acquired in each period to guide the dynamic grouping of load forecasting tasks. Specifically, the step of generating associated load groups and independent load groups based on the load correlation matrix includes: Based on the load correlation matrix, loads whose load correlation scores with other loads are all less than a preset correlation score threshold are obtained, and the independent load groups are generated. The preset correlation score threshold is adaptively adjusted and optimized based on a grid search method. In principle, the preset correlation score threshold can be set based on experience. However, considering that in actual applications, there may be situations where the actual application fails due to unreasonable setting of fixed empirical values, in order to ensure the reliability of dynamic load task grouping, this embodiment preferably sets the preset correlation score threshold to a certain empirical value (e.g., 0.3) in the initial application, and then adaptively adjusts and optimizes the preset correlation score threshold based on a grid search method in subsequent periodic prediction analysis. Specifically, the steps for optimizing the preset relevance score threshold using the grid search method may include: setting a candidate threshold set Θ={0.1,0.2,...,0.9}, and setting a training set and a validation set including multiple sets of multivariate load sequences, influencing factor sequences, and corresponding multivariate load prediction results; for each candidate threshold θ∈Θ, running a complete training and prediction process once on the training set and the validation set, and recording the average MSE (Mean Square Error) comprehensive evaluation index of the corresponding load prediction, so as to select the threshold with the best comprehensive index on the validation set as the preset relevance score threshold used for the division of associated load groups and independent load groups in the final prediction process.

[0028] The associated load group is generated by excluding the loads in the independent load group from the multi-load data.

[0029] In practical applications, the correlation scores of each load in the load correlation matrix are used to determine whether the correlation between each load and all other loads is weak or non-correlated. If the correlation scores of a load with all other loads are less than the preset correlation score threshold, it is considered that the load has a weak coupling correlation with other loads and needs to be predicted independently. It needs to be added to an independent load group and predicted separately in the current prediction period. If the correlation score of a load with any other load is greater than or equal to the preset correlation score threshold, it is considered that the load needs to be predicted collaboratively with other loads in the current prediction period and needs to be added to an associated load group.

[0030] This example captures the dynamic coupling characteristics of multiple loads through periodic analysis and intelligently and dynamically divides associated load groups and independent load groups. It can adaptively adjust the load coordination forecasting strategy based on the dynamic coupling characteristics of loads, effectively avoiding performance degradation caused by incorrect load forecasting task association and binding, thereby ensuring the reliability of multi-load group forecasting.

[0031] S14. Perform singular spectrum decomposition and reconstruction on the multivariate load sequence to be analyzed to obtain a reconstructed multivariate load sequence. Then, perform time alignment and splicing on the reconstructed multivariate load sequence, the multivariate load sequence to be analyzed, and the influencing factor sequence to obtain a second multivariate load feature sequence. The reconstructed multivariate load sequence can be understood as a multidimensional sequence obtained by performing singular spectrum analysis (SSA) on each load sequence to be analyzed in the multivariate load sequence to be analyzed and selecting important sub-sequences for reconstruction. It also includes multiple reconstructed load sequences.

[0032] Specifically, the step of performing singular spectral decomposition and reconstruction on the multivariate loading sequence to be analyzed to obtain the reconstructed multivariate loading sequence includes: Each load sequence to be analyzed in the multivariate load sequence to be analyzed is subjected to singular spectrum decomposition to obtain multiple corresponding frequency load subsequences. Singular spectrum decomposition can be understood as decomposing each load sequence to be analyzed into subsequences such as trend term, period term and residual term. The specific implementation process can be referred to the relevant existing technology, which will not be described in detail here.

[0033] Based on the variance contribution of each subsequence, the frequency load subsequences of each load sequence to be analyzed are screened and reconstructed to obtain the corresponding reconstructed load sequences. The variance contribution of each subsequence can be understood as the variance contribution of each subsequence to the original load sequence to be analyzed. In practical applications, the variance contribution of each frequency load subsequence relative to the corresponding load sequence to be analyzed is calculated, and then they are arranged in descending order of variance contribution. Subsequences with a cumulative variance contribution rate exceeding a certain proportion (e.g., 90%) are retained for sequence reconstruction to enhance the ability to capture the inherent variation characteristics of the load. For example, if the cumulative variance contribution rate of the first 4 subsequences reaches 90%, then the first 4 are reconstructed, and the remaining subsequences are discarded.

[0034] All the reconstructed load sequences are summarized to obtain the reconstructed multivariate load sequence.

[0035] This embodiment effectively separates trend and periodic terms by employing the singular spectral decomposition and reconstruction method, significantly reducing noise interference and thus improving the stationarity and predictability of the sequence, providing clearer structural features to support the subsequent learning process. Simultaneously, by splicing and fusing the reconstructed multivariate load sequence with the multivariate load sequence to be analyzed and the influencing factor sequence as input to the subsequent prediction model, the model's ability to capture the evolution of the load itself is enhanced through the reconstruction of the multivariate load sequence, while also sensing the influence of external factors on the load, achieving more comprehensive and accurate load prediction.

[0036] S15. Perform dual attention mechanism fusion analysis on each load feature sequence in the second multivariate load feature sequence to obtain multivariate load prediction input features.

[0037] The dual attention mechanism can be understood as a fusion mechanism composed of a temporal attention mechanism (TAM) and a feature attention mechanism (FAM). It dynamically weights each load feature sequence from both temporal and feature dimensions to capture important information in the time series or features. The temporal attention mechanism automatically identifies and weights key moments that significantly influence the prediction result by calculating the correlation between each time point and the prediction time, helping the model focus on information at key historical time points and thus better grasp load change trends. The feature attention mechanism, on the other hand, automatically assigns appropriate weights by evaluating the contribution of each feature in each load feature sequence to the prediction result, highlighting the role of important features and enhancing the subsequent prediction model's ability to discriminate input features. It should be noted that the processing procedures of both the temporal and feature attention mechanisms can be implemented with reference to existing technologies. After the features obtained from each load feature sequence are processed by the temporal and feature attention mechanisms respectively, they are concatenated and fused to obtain the corresponding load prediction input features, which then form the required multivariate load prediction input features.

[0038] S16. Based on the multi-load forecasting input features, perform multi-load hybrid forecasting based on the associated load group and the independent load group to obtain multi-load forecasting results; wherein, multi-load hybrid forecasting can be understood as a multi-load forecasting strategy that performs multi-task collaborative forecasting of all loads in the associated load group and single-task independent forecasting of each load in the independent load group; specifically, the step of performing multi-load hybrid forecasting based on the associated load group and the independent load group to obtain multi-load forecasting results includes: Based on all load prediction input features corresponding to the associated load group from the multivariate load prediction input features, collaborative prediction is performed based on a pre-built multi-task learning model to obtain the predicted values ​​of each load in the associated load group. In principle, the multi-task learning model can adopt existing multi-task prediction models. However, in order to better handle the complex and dynamic coupling relationship between multiple load prediction tasks, this embodiment preferably adopts a collaborative prediction model based on Temporal Convolutional Network (TCN) as the basic prediction structure. By learning the common change patterns and coupling relationship features of each related load in the associated load group through a shared TCN encoder, and then further learning its personalized features through the dedicated decoder of each load prediction task, the efficient modeling and analysis of the complex sequence of each load in the associated load group can be achieved.

[0039] Specifically, the step of obtaining the predicted values ​​of each load in the associated load group by performing collaborative prediction based on a pre-built multi-task learning model, according to all load prediction input features corresponding to the associated load group in the multivariate load prediction input features, includes: Based on the preset shared encoder, the load prediction input features corresponding to each load in the associated load group are respectively encoded to obtain the corresponding load encoding features. The preset shared encoder can be understood as a TCN encoder pre-trained based on the relevant load feature data. The specific process of using this encoder to encode the load prediction input features corresponding to each load in the associated load group can refer to the relevant application technology of the existing TCN encoder, which will not be detailed here.

[0040] All relevant load coding features corresponding to each of the aforementioned load coding features are weighted and summed based on the corresponding load correlation scores in the load correlation matrix to obtain the corresponding collaborative coding features. These collaborative coding features can be understood as shared features that each load in the associated load group needs to obtain from all loads with which it has a strong correlation. Considering that the load correlation scores between the same load and different loads are different, and the corresponding collaborative value also varies, this embodiment preferably adopts a weighted collaborative modeling mechanism driven by the load correlation matrix to obtain the collaborative coding features corresponding to each load forecasting task. That is, it guides different load forecasting tasks to share features to different degrees based on the corresponding load correlation scores in the load correlation matrix. The following describes the implementation process of the weighted collaborative modeling mechanism driven by the load correlation matrix using the collaborative coding feature acquisition process of a certain load i as an example: Assume that the load coded features obtained by feature encoding each load forecast input feature corresponding to the current associated load group S1 through a preset shared encoder are represented as: h1, h2, ..., h kFurthermore, the load correlation score between load i and related load j in the associated load group is expressed as follows: k represents the number of loads in associated load group S1.

[0041] The load correlation score between load j and load i is normalized using the following formula to obtain the data, which serves as the load encoding feature of load j and the shared weight required for the prediction task of load i: In the formula, This represents the load correlation score between load j and load i in the associated load group S1; This represents the load correlation score between load l and load i in the associated load group S1; The load coding characteristics of load j represent the shared weights required for the prediction task of load i.

[0042] Based on all the shared weights calculated using the above method, the load coding features of all loads related to load i in the associated load group are weighted and fused to obtain the collaborative coding features of load i as follows: In the formula, This represents the cooperative coding characteristics of load i.

[0043] The weighted collaborative modeling mechanism driven by the load correlation matrix provided in this embodiment can adaptively adjust the degree of feature sharing based on the dynamic correlation between loads, according to the principle that the greater the correlation, the greater the contribution of collaborative prediction. This ensures that the multi-task learning model can accurately perceive and use the correlation between loads to make reliable collaborative predictions, thereby effectively improving the accuracy of predictions for each associated load.

[0044] Each load coding feature and its corresponding co-coding feature are concatenated and fused to obtain a load fusion feature. This load fusion feature is then input into the corresponding decoder for load prediction to obtain the predicted value of the corresponding load. The load fusion feature can be understood as the feature obtained by concatenating the load coding feature of each load i in the associated load group with its corresponding co-coding feature. , can be represented as: .

[0045] After obtaining the load fusion characteristics of each load in the associated load group through the above methods and steps, these characteristics can be sent to the corresponding decoder for load forecasting task for processing and analysis to obtain the corresponding forecast results. Represented as: , The decoder corresponding to the load i prediction task can be pre-trained based on relevant load data.

[0046] Each load prediction input feature corresponding to the independent load group in the multivariate load prediction input features is input into its corresponding single-task learning model for individual prediction, obtaining the predicted value of each load in the independent load group. Preferably, the single-task learning model is an independent load task prediction model obtained by modeling an independent TCN structure, effectively avoiding interference with prediction results due to the introduction of low-correlation information, thus ensuring the accuracy of the prediction results for each load in the independent load group. It should be noted that the model structure of the single-task learning model corresponding to each load in the independent load group is the same, but the parameters may differ. All models can be pre-trained based on the TCN network using relevant load data; the specific training process can be implemented with reference to relevant existing technologies.

[0047] The predicted values ​​of each load in the associated load group and the predicted values ​​of each load in the independent load group are summarized to obtain the multi-variable load prediction result.

[0048] This embodiment employs the multi-task learning mechanism of TCN, enabling the multi-task learning model to fully utilize the dynamic correlation information between loads while also taking into account individual load differences. This effectively improves the accuracy of load collaborative forecasting. Furthermore, by adopting the single-task learning mechanism of TCN, it effectively avoids information interference between load forecasting tasks, allowing the single-task learning model to focus on the forecasting task of a single load and ensuring the reliability of forecasting for each independent load. This, in turn, effectively improves the overall forecasting performance of multi-load forecasting in the integrated energy system.

[0049] This invention, in its embodiments, periodically acquires and preprocesses multivariate load sequences and influencing factor sequences to generate the multivariate load sequence and influencing factor sequence to be analyzed. Then, it performs time-aligned concatenation of the multivariate load sequence and the influencing factor sequence to obtain a first multivariate load characteristic sequence. Based on the first multivariate load characteristic sequence, it performs load correlation analysis using a pre-built correlation assessment model to generate a load correlation matrix. Based on the load correlation matrix, it generates associated load groups and independent load groups. Finally, it performs singular spectrum decomposition and reconstruction on the multivariate load sequence to be analyzed to obtain a reconstructed multivariate load sequence. Finally, it performs time-aligned concatenation of the reconstructed multivariate load sequence, the multivariate load sequence to be analyzed, and the influencing factor sequence to obtain a second multivariate load characteristic sequence. The technical solution involves first analyzing the load characteristic sequences in the second multi-load characteristic sequence using a dual-attention mechanism to obtain the multi-load prediction input characteristics, and then performing multi-load hybrid prediction based on associated load groups and independent load groups to obtain the multi-load prediction results. This multi-load hybrid prediction mechanism, which incorporates periodic dynamic load correlation assessment to fully capture the dynamic coupling characteristics changes between multi-loads and adaptively adjusts the load collaborative prediction strategy accordingly, can effectively avoid the negative impact of inaccurate load coupling relationship identification on overall prediction performance. It improves the accuracy and robustness of multi-load prediction in integrated energy systems and has good prospects for promotion and engineering application value.

[0050] To verify the application performance of the proposed multi-load hybrid forecasting method for integrated energy systems, this invention uses actual operating data (covering cooling load, heating load, gas load, electricity load, and related meteorological factors and timestamp information) of an integrated energy system in a certain industrial park from January to December 2020. The preprocessed dataset was divided into 80% training set, 10% validation set, and 10% test set. A comparative experiment was conducted using the proposed forecasting method and a forecasting model based on a multi-task learning architecture (MTL) using LSTM (Long Short-Term Memory) or GRU (Gated Recurrent Unit). The results show... Figure 2 The analysis results are shown in Tables 1 and 2.

[0051] Table 1. Statistical Table of Root Mean Square Error Table 2. Statistical Table of Mean Absolute Percentage Error like Figure 2As shown, the overall trend of cooling load, gas load, heat load, and electrical load in the integrated energy system predicted by the method of this invention is significantly closer to the actual overall trend of the corresponding load than the prediction performance of multi-task learning models based on LSTM and GRU. Furthermore, as shown in Tables 1 and 2, the method of this invention achieves significant advantages in both root mean square error (RMSE) and mean absolute percentage error (MAPE), with average accuracy improvements of 68.26% and 83.60%, respectively. The method also exhibits stronger generalization ability and stability, making it suitable for complex scenarios involving the collaborative prediction of multiple load types.

[0052] It should be noted that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order requirement for the execution of these steps, and they can be executed in other orders.

[0053] In one embodiment, such as Figure 3 As shown, a multi-load hybrid forecasting system for integrated energy systems is provided, the system comprising: Data acquisition module 1 is used to periodically acquire multivariate load sequences and influencing factor sequences and perform preprocessing to generate multivariate load sequences and influencing factor sequences to be analyzed; The first splicing module 2 is used to align and splice the multivariate load sequence to be analyzed with the influencing factor sequence to obtain the first multivariate load characteristic sequence; The correlation analysis module 3 is used to perform correlation analysis between loads based on the first multivariate load characteristic sequence and a pre-built correlation evaluation model, generate a load correlation matrix, and generate associated load groups and independent load groups based on the load correlation matrix. The second splicing module 4 is used to perform singular spectrum decomposition and reconstruction on the multivariate load sequence to be analyzed to obtain a reconstructed multivariate load sequence, and to align and splice the reconstructed multivariate load sequence, the multivariate load sequence to be analyzed and the influencing factor sequence to obtain a second multivariate load feature sequence. The fusion analysis module 5 is used to perform dual attention mechanism fusion analysis on each load feature sequence in the second multivariate load feature sequence to obtain multivariate load prediction input features; The load forecasting module 6 is used to perform multi-load mixed forecasting based on the associated load group and the independent load group according to the multi-load forecasting input characteristics, and obtain multi-load forecasting results.

[0054] Specific limitations regarding the integrated energy system's multi-load hybrid forecasting system can be found in the limitations of the integrated energy system's multi-load hybrid forecasting method described above; the corresponding technical effects are equivalent and will not be repeated here. Each module in the aforementioned integrated energy system's multi-load hybrid forecasting system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, allowing the processor to call and execute the corresponding operations of each module.

[0055] Figure 4 An internal structural diagram of a computer device is shown in one embodiment. This computer device may specifically be a terminal or a server. Figure 4 As shown, the computer device includes a processor, memory, network interface, display, camera, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program can implement a multi-load hybrid forecasting method for integrated energy systems. The display screen can be an LCD screen or an e-ink display screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0056] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. Specific computing devices may include more or fewer components than those shown in the figure, or combine certain components, or have the same component arrangement.

[0057] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described above.

[0058] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0059] In summary, the multi-load hybrid forecasting method, system, equipment, and medium for integrated energy systems provided by the embodiments of the present invention can fully capture the dynamic coupling characteristics changes among multiple loads based on the introduction of periodic dynamic load correlation assessment and adaptively adjust the load collaborative forecasting strategy accordingly. This multi-load hybrid forecasting mechanism can effectively avoid the negative impact of inaccurate identification of load coupling relationships on the overall forecasting performance, improve the accuracy and robustness of multi-load forecasting for integrated energy systems, and has good prospects for promotion and engineering application value.

[0060] The various embodiments in this specification are described in a progressive manner. For directly identical or similar parts of the embodiments, refer to each other. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0061] The above-described embodiments are merely preferred embodiments of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the principles of the present invention, and these improvements and substitutions should also be considered within the scope of protection of the present invention. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A method for predicting multiple loads in a comprehensive energy system, characterized in that, The method comprises: The multivariate load series and influencing factor series are periodically acquired and preprocessed to generate the multivariate load series and influencing factor series to be analyzed. The multivariate load sequence to be analyzed is spliced ​​together with the influencing factor sequence in time alignment to obtain the first multivariate load characteristic sequence; Based on the first multivariate load characteristic sequence, load correlation analysis is performed based on a pre-constructed correlation assessment model to generate a load correlation matrix, and associated load groups and independent load groups are generated based on the load correlation matrix. The multivariate load sequence to be analyzed is subjected to singular spectrum decomposition and reconstruction to obtain a reconstructed multivariate load sequence. The reconstructed multivariate load sequence, the multivariate load sequence to be analyzed, and the influencing factor sequence are time-aligned and spliced ​​together to obtain a second multivariate load feature sequence. Each load feature sequence in the second multivariate load feature sequence is fused using a dual attention mechanism to obtain the multivariate load prediction input features. Based on the input characteristics of the multi-variable load forecast, multi-variable load mixed forecast is performed based on the associated load group and the independent load group to obtain the multi-variable load forecast result.

2. The integrated energy system multi-load hybrid forecasting method as described in claim 1, characterized in that, The relevance assessment model includes a feature extraction layer and a similarity analysis layer connected in sequence; the feature extraction layer includes two feature extraction sub-networks with shared weights; the feature extraction sub-network includes a convolutional layer and a recurrent encoding layer.

3. The integrated energy system multi-load hybrid forecasting method as described in claim 2, characterized in that, The similarity analysis layer is used to perform cosine similarity analysis on the two load feature vectors output by the feature extraction layer to obtain the corresponding load correlation score.

4. The integrated energy system multi-load hybrid forecasting method as described in claim 1, characterized in that, The step of generating associated load groups and independent load groups based on the load correlation matrix includes: Based on the load correlation matrix, loads whose load correlation scores with other loads are all less than a preset correlation score threshold are obtained, and the independent load group is generated; the preset correlation score threshold is adaptively adjusted and optimized based on a grid search method. The associated load group is generated by excluding the loads in the independent load group from the multi-load data.

5. The integrated energy system multi-load hybrid forecasting method as described in claim 1, characterized in that, The steps of performing singular spectral decomposition and reconstruction on the multivariate loading sequence to be analyzed to obtain the reconstructed multivariate loading sequence include: Each load sequence to be analyzed in the multivariate load sequence to be analyzed is subjected to singular spectral decomposition to obtain multiple corresponding frequency load subsequences; Based on the contribution of subsequence variance, the frequency load subsequences of each load sequence to be analyzed are screened and reconstructed to obtain the corresponding reconstructed load sequences. All the reconstructed load sequences are summarized to obtain the reconstructed multivariate load sequence.

6. The method for predicting multiple loads in a comprehensive energy system as described in claim 1, characterized in that, The step of performing multi-variable load mixed forecasting based on the multi-variable load forecasting input features, the associated load group, and the independent load group to obtain the multi-variable load forecasting result includes: Based on all load prediction input features corresponding to the associated load group in the multivariate load prediction input features, collaborative prediction is performed based on a pre-built multi-task learning model to obtain the predicted value of each load in the associated load group. Each load prediction input feature corresponding to the independent load group in the multivariate load prediction input features is input into the corresponding single-task learning model for individual prediction to obtain the predicted value of each load in the independent load group. The predicted values ​​of each load in the associated load group and the predicted values ​​of each load in the independent load group are summarized to obtain the multi-variable load prediction result.

7. The integrated energy system multi-load hybrid forecasting method as described in claim 6, characterized in that, The step of obtaining the predicted values ​​of each load in the associated load group by performing collaborative prediction based on a pre-built multi-task learning model using all load prediction input features corresponding to the associated load group from the multi-source load prediction input features includes: Based on a preset shared encoder, feature encoding is performed on each load prediction input feature corresponding to the associated load group to obtain the corresponding load encoding features; The corresponding collaborative coding features are obtained by weighted summing of all relevant load coding features corresponding to each load coding feature based on the corresponding load correlation scores in the load correlation matrix. Each of the load coding features and the corresponding co-coding features are concatenated and fused to obtain the corresponding load fusion feature. The load fusion feature is then input into the corresponding decoder for load prediction to obtain the predicted value of the corresponding load.

8. A multi-load hybrid forecasting system for integrated energy systems, characterized in that, The system includes: The data acquisition module is used to periodically acquire multivariate load sequences and influencing factor sequences, perform preprocessing, and generate multivariate load sequences and influencing factor sequences to be analyzed. The first splicing module is used to align and splice the multivariate load sequence to be analyzed with the influencing factor sequence to obtain the first multivariate load characteristic sequence; The correlation analysis module is used to perform correlation analysis between loads based on the first multivariate load characteristic sequence and a pre-built correlation evaluation model, generate a load correlation matrix, and generate associated load groups and independent load groups based on the load correlation matrix. The second splicing module is used to perform singular spectrum decomposition and reconstruction on the multivariate load sequence to be analyzed to obtain a reconstructed multivariate load sequence, and to align and splice the reconstructed multivariate load sequence, the multivariate load sequence to be analyzed and the influencing factor sequence to obtain a second multivariate load feature sequence. The fusion analysis module is used to perform dual attention mechanism fusion analysis on each load feature sequence in the second multivariate load feature sequence to obtain multivariate load prediction input features; The load forecasting module is used to perform multi-load mixed forecasting based on the associated load group and the independent load group according to the multi-load forecasting input characteristics, and obtain multi-load forecasting results.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

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

Images