Integrated energy system multi-load prediction method based on MVMD decomposition and hybrid model
The integrated energy system multi-load forecasting method based on MVMD decomposition and hybrid model solves the problems of insufficient mining of load coupling characteristics and strong subjectivity of component clustering, and realizes accurate forecasting of electricity, heat and cooling loads, improving forecast accuracy and adaptability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ANHUI UNIV OF SCI & TECH
- Filing Date
- 2026-01-06
- Publication Date
- 2026-04-17
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
In existing technologies, the load forecasting methods for integrated energy systems fail to fully exploit the coupling and correlation information between electricity, heat, and cooling loads, have insufficient multivariate signal decomposition capabilities, exhibit strong subjectivity in component clustering, and have poor frequency adaptability of the forecasting model, resulting in insufficient forecasting accuracy.
Multivariate variational mode decomposition (MVMD) is used to simultaneously decompose multiple load signals. The core parameters are determined by combining the Chinese pangolin optimization algorithm. Adaptive frequency component partitioning is achieved by sample entropy and K-means clustering. Differential prediction is performed using a hybrid model of support vector machine (SVM) and temporal convolutional network-gated recurrent unit (TCN-GRU).
It enables accurate prediction of electricity, heat, and cooling loads, improves prediction accuracy and adaptability, solves the problems of insufficient mining of multi-load coupling characteristics and strong subjectivity of component clustering, and meets the refined operation requirements of integrated energy systems.
Smart Images
Figure CN121886362A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of integrated energy system load forecasting, and in particular to a multi-load forecasting method for integrated energy systems based on MVMD decomposition and hybrid models. Background Technology
[0002] Integrated energy systems are crucial for achieving synergistic optimization of multiple energy sources, playing a significant role in improving energy efficiency and promoting the consumption of renewable energy. Within integrated energy systems, there are complex coupling relationships among electrical, thermal, and cooling loads. Their load characteristics are influenced by a combination of factors, including climate conditions and user behavior, exhibiting significant nonlinear and non-stationary features. Accurate multi-load forecasting is fundamental to the economic dispatch, optimized operation, and demand-side management of integrated energy systems, playing a key role in ensuring the safe and stable operation of the system and improving economic efficiency.
[0003] In existing technologies, load forecasting methods still have many shortcomings: First, traditional forecasting methods mostly adopt a single load independent forecasting mode, failing to fully consider the mutual influence and coordinated operation characteristics between electrical, heat, and cooling loads, and cannot effectively mine the inherent coupling correlation information of multiple loads; Second, signal decomposition methods also have limitations. Common empirical mode decomposition (EMD) is prone to mode aliasing problems when processing non-stationary load signals, and variational mode decomposition (VMD), although having better performance, is only suitable for univariate decomposition. Neither can effectively handle multivariate load signals; Third, when dividing the intrinsic mode function (IMF) components obtained after decomposition into high and low frequencies, existing methods mostly rely on manually setting thresholds for clustering, lacking scientific and objective quantitative evaluation indicators, and cannot adaptively identify the frequency characteristics and complexity differences of different IMF components, resulting in unreasonable high and low frequency component division, which in turn affects the adaptability and prediction accuracy of subsequent forecasting models. Finally, existing forecasting schemes mostly use a single model to complete the full component forecast. However, after the load data is decomposed, the high-frequency components often correspond to the characteristics of random fluctuations and sudden changes in the load, while the low-frequency components correspond to the characteristics of long-term trends and periodic patterns in the load. The learning ability of a single model is limited, making it difficult to accurately capture the differentiated change patterns of high-frequency fluctuations and low-frequency trends at the same time. This results in insufficient overall forecast accuracy and fails to meet the needs of refined operation of integrated energy systems. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides a multi-load prediction method for integrated energy systems based on MVMD decomposition and hybrid models. This invention solves the core problems in existing technologies, such as insufficient mining of multi-load coupling features, insufficient multivariate signal decomposition capabilities, strong subjectivity in component clustering, and poor frequency adaptability of prediction models.
[0005] To achieve the above objectives, this application provides the following solution: A comprehensive energy system multi-load forecasting method based on MVMD decomposition and hybrid model includes the following steps: Step 1: Data acquisition and preprocessing. Collect historical data on electrical load, heat load, and cooling load of the integrated energy system, as well as external characteristic data such as meteorological and temporal data. Clean and normalize the data. Step 2: Multiload sequence decomposition and component classification; Step 2 specifically includes: Step 2-1: First, the preprocessed multi-load data is simultaneously decomposed using the multivariate variational mode decomposition (MVMD) method. The core parameters of MVMD, namely the number of mode decompositions K and the penalty factor α, are determined by global optimization using the Chinese pangolin optimization algorithm CPO to ensure the accuracy and effectiveness of the decomposition results. Step 2-2: Extract the multiple sets of multivariable intrinsic mode functions (IMF) components obtained from the decomposition, and calculate the sample entropy value of each IMF component; Steps 2-3: Based on the sample entropy calculation results, the K-means clustering algorithm is used to adaptively classify all IMF components, dividing them into low-frequency IMF component groups and high-frequency IMF component groups; Step 3: Calculate the Spearman correlation coefficient between each feature in the external feature data and the original multi-load sequence; set a preset threshold for the correlation coefficient, and select external features with correlation coefficients greater than the preset threshold as the input feature set for the subsequent prediction model; Step 4: Hybrid model prediction. Adapted models are used for prediction of the low-frequency group and the high-frequency group respectively. Support vector machine (SVM) is used for prediction of the low-frequency group, and TCN-GRU, a hybrid model combining temporal convolutional network and gated recurrent unit, is used for prediction of the high-frequency group. Step 5: Reconstruct load forecast results. The forecast results of the low-frequency group and the high-frequency group obtained in Step 4 are superimposed and reconstructed to obtain the final load forecast results.
[0006] In step 1, the data cleaning specifically includes missing value processing and outlier processing; wherein, missing value processing uses linear interpolation to fill in missing values, and outlier processing uses the 3σ criterion to remove outliers; the normalization process uses the Min-Max normalization method to map the cleaned multi-load data and external feature data to the [0,1] interval, thereby obtaining a standardized multi-load dataset and a standardized external feature dataset.
[0007] In step 2, the specific steps are as follows: Step 2-1: For an input data x(t) = [x1(t), x2(t), …, xc(t)] containing C data channels, the Multivariate Variational Mode Decomposition (MVMD) decomposes it into K multivariable components uk,c(t). The constrained optimization problem is as follows:
[0008] Where uk,c(t)=[u1(t),u2(t),…,uc(t)] are the modal components. These are the center frequencies of each mode; To solve the above problem, the Lagrange function is introduced to transform the constrained problem into an unconstrained one:
[0009] in, As a penalty factor, For Lagrange operators; The core parameters of MVMD, namely the number of mode decompositions K and the penalty factor α, are determined through global optimization using the CPO algorithm. The specific optimization process includes: first, setting the optimization range of the number of mode decompositions K and the optimization range of the penalty factor α; then, taking the mean minimum envelope entropy of the IMF component corresponding to each load as the core optimization index; by minimizing this mean envelope entropy, ensuring that the modal purity and separation of the IMF component corresponding to different loads are both optimal. Step 2-2: After completing the MVMD decomposition of the multi-load data, calculate the sample entropy value for each obtained multivariate IMF component. The formula for calculating the sample entropy is:
[0010] in, For embedded dimensions, For similarity tolerance, The sequence length is represented by the sample entropy value. The higher the sample entropy value, the higher the complexity of the corresponding IMF component sequence and the higher the fluctuation frequency. Conversely, the lower the entropy value, the stronger the regularity of the sequence and the lower the fluctuation frequency. Steps 2-3: Based on the sample entropy calculation results of each IMF component, the K-means clustering algorithm is used to adaptively classify all IMF components, dividing them into two categories, corresponding to the low-frequency IMF component group and the high-frequency IMF component group, respectively.
[0011] Step 3 involves calculating the Spearman correlation coefficient for each feature in the standardized external feature data, comparing it to the original multi-load sequence. The formula for this coefficient is as follows:
[0012] in, For the sample size, For the first The difference between the rank of the external feature corresponding to each sample and the rank of the multiload sequence; A correlation threshold of 0.3 is set for the Spearman correlation coefficient. Features in the external feature data whose Spearman correlation coefficient with the original multiload sequence is greater than this threshold are retained, and the filtered features are used as input features for the subsequent prediction model.
[0013] In the component prediction process of the hybrid model described in step 4, suitable prediction models are used for the low-frequency component group and the high-frequency component group obtained by MVMD decomposition, respectively. The low-frequency component is predicted using the SVM model, and the high-frequency component is predicted using the TCN-GRU hybrid model. The specific implementation is as follows: Based on the low-frequency component group obtained from MVMD decomposition, an SVM prediction model is constructed. The training set input features include a feature subset obtained through feature selection and the historical values of the low-frequency component. The output is the low-frequency component value at the time to be predicted. The radial basis function (RBF) is used as the kernel function of the SVM, and the kernel function expression is as follows:
[0014] in and For the input vector, For kernel parameters; Based on the high-frequency component groups obtained from MVMD decomposition, a TCN-GRU hybrid model is constructed. This model has an input layer, a temporal convolutional network layer, a gated recurrent unit layer, and a fully connected layer as its core architecture: the input layer receives the temporal feature sequence composed of the high-frequency component groups obtained from MVMD decomposition; the TCN layer adopts a combination of causal convolution and dilated convolution to effectively extract the multi-scale temporal features of the high-frequency components; the GRU layer is used to capture the long-short-term temporal dependencies of the high-frequency component sequence; the fully connected layer uses a linear activation function to map the features extracted by the aforementioned layers to the prediction results; during the model training phase, the Adam optimizer is used to optimize the parameter convergence process, and Dropout regularization technology is introduced to effectively reduce the risk of model overfitting.
[0015] The load forecast result reconstruction step in step 5 is as follows: the low-frequency group forecast result obtained by the SVM model is superimposed and reconstructed with the high-frequency component group forecast result obtained by the TCN-GRU hybrid model, and finally the complete forecast result of the electric, heat and cooling loads of the integrated energy system is obtained.
[0016] In summary, the technical solutions conceived by this invention have the following beneficial effects compared with the prior art: This invention employs multivariate variational mode decomposition (MVMD) as a decomposition method for multi-load signals. This technique enables the simultaneous decomposition of multiple load signals, such as electrical load, thermal load, and cooling load, effectively avoiding the mode aliasing problem that easily occurs in traditional empirical mode decomposition methods. It also overcomes the limitation of traditional variational mode decomposition, which can only handle single-variable signals. MVMD has a noise reduction effect, improving the quality of the decomposed signal components. The intrinsic mode function components obtained after decomposition and noise reduction using MVMD accurately reflect the inherent characteristics of the original multi-load data, providing a foundation for the establishment of subsequent prediction models.
[0017] This invention automates component clustering by constructing an adaptive frequency component partitioning method based on sample entropy and K-means clustering. Specifically, sample entropy is used to quantify each intrinsic mode function component obtained from MVMD decomposition to construct an evaluation index. Then, the sample entropy calculation results of each IMF component are used as features input to the K-means clustering algorithm. The clustering algorithm is used to partition high-frequency components from low-frequency components, avoiding the partitioning bias problem caused by manually setting thresholds.
[0018] This invention employs a differentiated prediction method. For high-frequency load components, which exhibit random fluctuations, sudden changes, and nonlinearity, a hybrid model of Temporal Convolutional Network-Gated Recurrent Unit (TCN-GRU) is selected for prediction. TCN possesses local feature extraction capabilities, while GRU captures the dynamic dependencies of time-series data. The combination of these two models adapts to the rapid changes and nonlinear characteristics of high-frequency components, enabling the prediction of sudden load fluctuations. For low-frequency load components, which exhibit long-term trends and periodic patterns, Support Vector Machine (SVM) is used for learning. SVM demonstrates generalization ability and stability when processing trending data, enabling it to fit the changing trends and periodic patterns of low-frequency components and avoid overfitting. Attached Figure Description
[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0020] Figure 1 This is a schematic diagram of the MVMD decomposition and hybrid model prediction process provided by the present invention.
[0021] Figure 2 This is a graph showing the results of decomposing electrical load using the MVMD method provided by this invention.
[0022] Figure 3This is a graph showing the results of heat load decomposition using the MVMD method provided by this invention.
[0023] Figure 4 This is a diagram showing the results of decomposing the cooling load using the MVMD method provided by this invention. Detailed Implementation
[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the following description, in conjunction with the accompanying drawings and embodiments, will further clarify them. This invention will be further described in detail below. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the invention. Furthermore, the technical features involved in the various embodiments of the invention described below can be combined with each other as long as they do not conflict with each other.
[0025] Please see Figures 1 to 4 The present invention provides a method for predicting multiple loads in a comprehensive energy system based on MVMD decomposition and hybrid models, comprising the following steps: Step 1: Data acquisition and preprocessing. Collect historical data on electrical load, heat load, and cooling load of the integrated energy system, as well as external characteristic data such as meteorological and temporal data. Clean and normalize the data. In step 1: Historical data on electricity load, heat load, and cooling load of a university campus's integrated energy system over a year were collected, with a sampling interval of 1 hour, resulting in 8784 sets of load data. Simultaneously, meteorological characteristic data for the same period, including temperature, humidity, and time-related data, were collected to form the original dataset. The collected data underwent an integrity check, revealing a small number of missing values. Linear interpolation was used to fill in the missing values, ensuring the continuity of the data sequence. Outliers were removed using the 3σ criterion. First, the mean μ and standard deviation σ of each load sequence and meteorological characteristic sequence were calculated. Data deviating from the mean ± 3σ range were identified as outliers and removed, resulting in a cleaned dataset. The Min-Max normalization method was used to map the cleaned multi-load data and external characteristic data to the [0,1] interval. The normalization formula is:
[0026] in, For the cleaned data, , These are the minimum and maximum values of the corresponding data sequence, respectively.
[0027] Step 2: Multiload sequence decomposition and component classification; Step 2 specifically includes: Step 2-1: First, the preprocessed multi-load data is simultaneously decomposed using the multivariate variational mode decomposition (MVMD) method. The core parameters of MVMD, namely the number of mode decompositions K and the penalty factor α, are determined by global optimization using the Chinese pangolin optimization algorithm CPO to ensure the accuracy and effectiveness of the decomposition results. Step 2-2: Extract the multiple sets of multivariable intrinsic mode functions (IMF) components obtained from the decomposition, and calculate the sample entropy value of each IMF component; Steps 2-3: Based on the sample entropy calculation results, the K-means clustering algorithm is used to adaptively classify all IMF components, dividing them into low-frequency IMF component groups and high-frequency IMF component groups; In step 2, the specific steps are as follows: Step 2-1: For an input data x(t) = [x1(t), x2(t), …, xc(t)] containing C data channels, the Multivariate Variational Mode Decomposition (MVMD) decomposes it into K multivariable components uk,c(t). The constrained optimization problem is as follows:
[0028] Where uk,c(t)=[u1(t),u2(t),…,uc(t)] are the modal components. These are the center frequencies of each mode; To solve the above problem, the Lagrange function is introduced to transform the constrained problem into an unconstrained one:
[0029] in, As a penalty factor, For Lagrange operators; The core parameters of MVMD, the number of modal decompositions K and the penalty factor α, are determined through global optimization using the CPO algorithm. The specific optimization process includes: First, setting the optimization range of the number of modal decompositions K to [2, 12] and the optimization range of the penalty factor α to [1000, 3000]. Then, the minimum mean envelope entropy of the IMF components corresponding to each load is taken as the core optimization index. The smaller the envelope entropy, the higher the purity of the modal components and the better the separation effect. By minimizing the mean envelope entropy, the modal purity and separation of the IMF components corresponding to different loads are ensured to reach the optimal. The Chinese pangolin optimization algorithm CPO is used to globally optimize K and α. The population size of the CPO algorithm is set to 20 and the number of iterations is set to 100. The optimal parameter combination is obtained through iterative search: K=10, α=1127.0494. Step 2-2: After completing the MVMD decomposition of the multi-load data, the decomposition results are as follows: Figure 2 , Figure 3 , Figure 4 As shown, for each obtained multivariate IMF component, its sample entropy value is calculated, and the sample entropy parameter is set. =2, =0.2 × standard deviation =8784, the formula for calculating the sample entropy is:
[0030] in, For embedded dimensions, For similarity tolerance, The sequence length is represented by the sample entropy value. The higher the sample entropy value, the higher the complexity of the corresponding IMF component sequence and the higher the fluctuation frequency. Conversely, the lower the entropy value, the stronger the regularity of the sequence and the lower the fluctuation frequency. Steps 2-3: Based on the sample entropy calculation results of each IMF component, the K-means clustering algorithm is used to adaptively classify all IMF components into two categories, corresponding to the low-frequency IMF component group and the high-frequency IMF component group, respectively. The low-frequency component group represents the trend and periodic characteristics of the load, while the high-frequency component group represents the high-frequency fluctuation and randomness characteristics of the load.
[0031] Step 3: Calculate the Spearman correlation coefficient between each feature in the external feature data and the original multi-load sequence; set a preset threshold for the correlation coefficient, and select external features with correlation coefficients greater than the preset threshold as the input feature set for the subsequent prediction model.
[0032] In step 3: the Spearman correlation coefficient is calculated by taking each feature in the standardized external feature data and calculating its Spearman rank correlation coefficient with the original multiload sequence. The formula for calculating this coefficient is as follows:
[0033] in, For the sample size, For the first The difference between the rank of the external feature corresponding to each sample and the rank of the multiload sequence; A correlation threshold of 0.3 is set for the Spearman correlation coefficient. Features in the external feature data whose Spearman correlation coefficient with the original multiload sequence is greater than this threshold are retained, and the filtered features are used as input features for the subsequent prediction model.
[0034] Step 4: Hybrid model prediction. Adapted models are used for prediction of the low-frequency group and the high-frequency group respectively. Support vector machine (SVM) is used for prediction of the low-frequency group, and TCN-GRU, a hybrid model combining temporal convolutional network and gated recurrent unit, is used for prediction of the high-frequency group.
[0035] In step 4, the standardized multi-load dataset and the filtered feature set are divided into training and test sets in an 8:2 ratio. To comprehensively evaluate the predictive performance and computational efficiency of the experimental design model, this invention selects four prediction evaluation metrics: Mean Absolute Error (MAE), Mean Squared Error (MSE), Mean Absolute Percentage Error (MAPE), and Coefficient of Determination (R²). 2 The calculation formula is:
[0036]
[0037]
[0038]
[0039] in, It is the actual value. This is a predicted value.
[0040] For the components of the low-frequency IMF component group, an SVM prediction model is constructed. The training set input features of each model are the selected feature subset of the corresponding load and the historical value of the low-frequency component, and the output is the low-frequency component value at the time to be predicted. The radial basis function (RBF) is used as the kernel function of the SVM, and the kernel function expression is as follows:
[0041] in and For the input vector, For kernel parameters; The hyperparameters were optimized by combining grid search with 5-fold cross-validation. The search range of the penalty parameter C was 1 to 1000, and the search range of the kernel parameter γ was 0.001 to 1. The optimal parameters were finally obtained: C=100, γ=0.01. The optimized parameters were used to train a support vector machine model to predict the low-frequency components in the test set, and the prediction result sequence of the low-frequency components was obtained. For each component of the high-frequency IMF component group, a TCN-GRU hybrid model is constructed. The model architecture consists of an input layer, a TCN layer, a GRU layer, and a fully connected layer. The input layer receives the high-frequency component sequence and a temporal feature sequence composed of a selected feature subset, with a time window length of 24 time steps. The TCN layer employs a combination of causal convolution and dilated convolution, with a dilation factor sequence of [1,2,4,8], a kernel size of 3, and 64 filters per layer. The GRU layer captures the long-short-term temporal dependencies of the high-frequency component sequence, containing one GRU layer with 128 hidden units. The fully connected layer uses a linear activation function, outputting the high-frequency component value at the time to be predicted. The loss function is mean squared error (MSE), and Adam is used during model training. The optimizer optimizes the parameter convergence process, with a learning rate of 0.001 and 150 training epochs. Dropout regularization is introduced with a dropout rate of 0.2. At the same time, an early stopping strategy is adopted to prevent overfitting. After training the TCN-GRU model with the training set data, the high-frequency components in the test set are predicted to obtain the corresponding high-frequency component prediction result sequence.
[0042] Step 5: Reconstruct load forecast results. The forecast results of the low-frequency group and the high-frequency group obtained in Step 4 are superimposed and reconstructed to obtain the final load forecast results.
[0043] In step five: the low-frequency component prediction results and high-frequency component prediction results of the same load type obtained in step four are superimposed and summed to complete the reconstruction of the prediction results, and finally the complete prediction results of the electric, heat and cold loads of the integrated energy system are obtained.
[0044] The example is applied to the integrated energy system of a university campus. The campus includes 32 buildings such as teaching buildings, research buildings, and dormitories. The electricity, heat, and cooling loads are supplied by self-built cogeneration units, chiller units, and heating pipe networks. It is necessary to achieve accurate hourly multi-load forecasting to provide data support for the optimized scheduling of the campus energy system and the reduction of operating costs.
[0045] The proposed method for multi-load forecasting of integrated energy systems based on MVMD decomposition and hybrid models has multiple theoretical advantages: First, it employs MVMD multivariate synchronous decomposition technology, overcoming the limitations of traditional univariate decomposition methods. This method preserves the coupling correlation information between electrical, thermal, and cooling loads during the decomposition process, theoretically ensuring the effective mining of multi-load synergistic characteristics and avoiding the loss of cross-load correlation characteristics caused by independent decomposition. Second, it introduces the Chinese pangolin optimization algorithm to globally optimize the core parameters of MVMD, using the mean envelope entropy as the optimization objective. This algorithmically ensures the modal purity and separation of IMF components, solving the problems of traditional decomposition methods. The method parameters rely on human experience; thirdly, the complexity of IMF components is quantified by sample entropy, and K-means clustering is combined to achieve objective classification of high and low frequency components. Quantitative indicators replace subjective thresholds, improving the scientificity and adaptability of component division from the perspective of classification logic. On this basis, for the divided low frequency component group, the predicted value of each low frequency component is output by the SVM model, and for the high frequency component group, the predicted value of each high frequency component is output by the TCN-GRU model. Finally, the predicted values of all low frequency components and high frequency components of the same load are accumulated to obtain the final prediction result of the load, achieving accurate integration of multi-load prediction results.
[0046] At the level of prediction model and feature processing, the theoretical advantages of this invention are equally prominent: On the one hand, Spearman correlation analysis is used for feature selection, which can effectively capture the nonlinear correlation between load and external features, reducing the risk of input redundancy and missing key information from the feature engineering perspective, and providing a more accurate input basis for the model; on the other hand, a hybrid model architecture of low-frequency SVM and high-frequency TCN-GRU is constructed, achieving accurate matching of the characteristics of different frequency components from the perspective of model adaptability. SVM theoretically has strong generalization ability and computational efficiency when dealing with stable, low-complexity low-frequency trend components; while the TCN-GRU hybrid model combines the multi-scale local feature extraction capability of TCN with the long-short-term time series dependency capture capability of GRU, theoretically adapting to the complex fluctuation characteristics of high-frequency components. This divide-and-conquer model design breaks through the limitation of a single model being unable to take into account different frequency components from an architectural logic perspective, providing theoretical support for improving the accuracy of multi-load prediction.
Claims
1. A method for multi-load forecasting of integrated energy systems based on MVMD decomposition and mixed models, characterized in that, Includes the following steps: Step 1: Data acquisition and preprocessing. Collect historical data on electrical load, heat load, and cooling load of the integrated energy system, as well as external characteristic data such as meteorological and temporal data. Clean and normalize the data. Step 2: Multiload sequence decomposition and component classification; Step 2 specifically includes: Step 2-1: First, the preprocessed multi-load data is simultaneously decomposed using the multivariate variational mode decomposition (MVMD) method. The core parameters of MVMD, namely the number of mode decompositions K and the penalty factor α, are determined by global optimization using the Chinese pangolin optimization algorithm CPO to ensure the accuracy and effectiveness of the decomposition results. Step 2-2: Extract the multiple sets of multivariable intrinsic mode functions (IMF) components obtained from the decomposition, and calculate the sample entropy value of each IMF component; Steps 2-3: Based on the sample entropy calculation results, the K-means clustering algorithm is used to adaptively classify all IMF components, dividing them into low-frequency IMF component groups and high-frequency IMF component groups; Step 3: Calculate the Spearman correlation coefficient between each feature in the external feature data and the original multi-load sequence; set a preset threshold for the correlation coefficient, and select external features with correlation coefficients greater than the preset threshold as the input feature set for the subsequent prediction model; Step 4: Hybrid model prediction. Adapted models are used for prediction of the low-frequency group and the high-frequency group respectively. Support vector machine (SVM) is used for prediction of the low-frequency group, and TCN-GRU, a hybrid model combining temporal convolutional network and gated recurrent unit, is used for prediction of the high-frequency group. Step 5: Reconstruct load forecast results. The forecast results of the low-frequency group and the high-frequency group obtained in Step 4 are superimposed and reconstructed to obtain the final load forecast results.
2. The method of claim 1, wherein the method is based on a MVMD decomposition and a hybrid model. In step 1, the data cleaning specifically includes missing value processing and outlier processing; wherein, missing value processing uses linear interpolation to fill in missing values, and outlier processing uses the 3σ criterion to remove outliers; the normalization process uses the Min-Max normalization method to map the cleaned multi-load data and external feature data to the [0,1] interval, thereby obtaining a standardized multi-load dataset and a standardized external feature dataset.
3. The method of claim 1, wherein the method is based on a MVMD decomposition and a hybrid model. In step 2, the specific steps are as follows: Step 2-1: For an input data x(t) = [x1(t), x2(t), …, xc(t)] containing C data channels, the Multivariate Variational Mode Decomposition (MVMD) decomposes it into K multivariable components uk,c(t). The constrained optimization problem is as follows: ； where uk,c(t) = [u1(t), u2(t),...,uc(t)] is the modal component, is the center frequency of each mode. To solve the above problem, the Lagrange function is introduced to transform the constrained problem into an unconstrained one: ； wherein is a penalty factor, is a Lagrangian operator; The core parameters of MVMD, namely the number of mode decompositions K and the penalty factor α, are determined through global optimization using the CPO algorithm. The specific optimization process includes: first, setting the optimization range of the number of mode decompositions K and the optimization range of the penalty factor α; then, taking the mean minimum envelope entropy of the IMF component corresponding to each load as the core optimization index; by minimizing this mean envelope entropy, ensuring that the modal purity and separation of the IMF component corresponding to different loads are both optimal. Step 2-2: After completing the MVMD decomposition of the multi-load data, calculate the sample entropy value for each obtained multivariate IMF component. The formula for calculating the sample entropy is: ； wherein, is the embedding dimension, is the similarity tolerance, is the sequence length, the higher the sample entropy value, the higher the complexity and fluctuation frequency of the corresponding IMF component sequence, and vice versa. Steps 2-3: Based on the sample entropy calculation results of each IMF component, the K-means clustering algorithm is used to adaptively classify all IMF components, dividing them into two categories, corresponding to the low-frequency IMF component group and the high-frequency IMF component group, respectively.
4. The method of claim 1, wherein the method is based on a MVMD decomposition and a hybrid model. Step 3 involves calculating the Spearman correlation coefficient for each feature in the standardized external feature data, comparing it to the original multi-load sequence. The formula for this coefficient is as follows: ； wherein, is the number of samples, is the difference between the external feature rank and the multi-load sequence rank corresponding to the th sample. A correlation threshold of 0.3 is set for the Spearman correlation coefficient. Features in the external feature data whose Spearman correlation coefficient with the original multiload sequence is greater than this threshold are retained, and the filtered features are used as input features for the subsequent prediction model.
5. The integrated energy system multi-load forecasting method based on MVMD decomposition and hybrid model as described in claim 1, characterized in that, In the component prediction process of the hybrid model described in step 4, suitable prediction models are used for the low-frequency component group and the high-frequency component group obtained by MVMD decomposition, respectively. The low-frequency component is predicted using the SVM model, and the high-frequency component is predicted using the TCN-GRU hybrid model. The specific implementation is as follows: Based on the low-frequency component group obtained from MVMD decomposition, an SVM prediction model is constructed. The training set input features include a feature subset obtained through feature selection and the historical values of the low-frequency component. The output is the low-frequency component value at the time to be predicted. The radial basis function (RBF) is used as the kernel function of the SVM, and the kernel function expression is as follows: ； in and For the input vector, For kernel parameters; Based on the high-frequency component groups obtained from MVMD decomposition, a TCN-GRU hybrid model is constructed. The core architecture of this model consists of an input layer, a temporal convolutional network layer, a gated recurrent unit layer, and a fully connected layer. The input layer receives the temporal feature sequence composed of the high-frequency component groups obtained from MVMD decomposition. The TCN layer adopts a combination structure of causal convolution and dilated convolution to effectively extract the multi-scale temporal features of the high-frequency components. The GRU layer is used to capture the long and short-term temporal dependencies of high-frequency component sequences; the fully connected layer uses a linear activation function to map the features extracted by the previous layers into the prediction results. During the model training phase, the Adam optimizer is used to optimize the parameter convergence process, and Dropout regularization is introduced to effectively reduce the risk of model overfitting.
6. The integrated energy system multi-load forecasting method based on MVMD decomposition and hybrid model as described in claim 1, characterized in that, The load forecast result reconstruction step in step 5 is as follows: the low-frequency group forecast result obtained by the SVM model is superimposed and reconstructed with the high-frequency component group forecast result obtained by the TCN-GRU hybrid model, and finally the complete forecast result of the electric, heat and cooling loads of the integrated energy system is obtained.