A comprehensive energy system multi-element load prediction method, system, device and medium
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TAIYUAN UNIVERSITY OF TECHNOLOGY
- Filing Date
- 2026-04-29
- Publication Date
- 2026-07-24
Smart Images
Figure CN122453448A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of energy forecasting technology, and in particular to a method, system, device and medium for forecasting multiple loads in an integrated energy system. Background Technology
[0002] Integrated Energy Systems (IES) integrate multiple energy sources such as electricity, gas, cooling, and heating, and are a key form of achieving efficient energy utilization and renewable energy consumption. Accurate multi-load forecasting is a crucial foundation for system planning, optimized operation, and market transactions. However, various loads are affected by weather, time, and user behavior, exhibiting strong volatility and randomness, and there are complex physical couplings and demand relationships between them, which poses a significant challenge to high-precision forecasting.
[0003] Existing prediction methods have significant limitations in addressing the aforementioned challenges. Traditional statistical methods struggle to characterize the nonlinear and nonstationary nature of loads; mainstream machine learning models cannot fully capture the long-term temporal dependencies of sequences; and while single deep learning models such as LSTM and CNN each have their advantages, they often focus only on features in a single dimension of time or space, failing to adequately explore the spatiotemporal correlation features of deep coupling between multiple loads, thus limiting prediction accuracy and generalization ability.
[0004] Furthermore, improved methods for joint forecasting of multi-load loads also have inherent limitations. On the one hand, methods employing multi-task learning frameworks typically lack collaborative mining of local and global correlations in the load feature space, resulting in insufficient comprehensiveness in feature extraction. On the other hand, while graph convolutional network-based methods can model spatial relationships, their ability to learn weakly correlated features in sparse data scenarios is weak, and they mostly rely on static graph structures, making it difficult to dynamically adapt to changes in load correlations over time. In addition, these methods usually directly process the original load sequences, where noise can interfere with the model's learning of deep, effective features, affecting prediction performance.
[0005] Therefore, existing technologies have not yet been able to effectively solve key problems in multi-variable load forecasting of integrated energy systems, such as insufficient extraction of deep correlation features, inaccurate modeling of dynamic coupling relationships, and incomplete filtering of noise interference. Summary of the Invention
[0006] The purpose of this invention is to provide a method, system, device and medium for predicting multiple loads in an integrated energy system, which can more fully explore the deep spatiotemporal correlation characteristics of multiple loads, more accurately model the dynamic coupling relationship between loads, and effectively filter out noise interference in the original data, thereby improving prediction accuracy and generalization ability.
[0007] To achieve the above objectives, the present invention provides a multi-element load forecasting method for integrated energy systems, comprising the following steps: Step S1: Obtain historical multivariate load data and related meteorological data of the integrated energy system; calculate the correlation between each meteorological feature and each target load using the unified information coefficient method, and select strongly correlated features based on preset thresholds to form a filtered feature set; Step S2: The historical multivariate load data is denoised using a combined method of complete set empirical mode decomposition combining sample entropy and wavelet threshold denoising to obtain the denoised electric, cold and hot load sequences. Step S3: Using the filtered feature set and the denoised load sequence as input, construct and use a parallel model of graph convolutional network and bidirectional long short-term memory network for feature extraction; wherein, using graph convolutional network, based on the correlation adjacency matrix dynamically constructed from the input, the spatial features of multivariate loads are extracted; at the same time, the temporal features of multivariate loads are extracted using bidirectional long short-term memory network; the extracted spatial features and temporal features are concatenated to obtain fused spatiotemporal depth features. Step S4: Input the fused spatiotemporal depth features into the shared layer of the attention-based multi-task learning framework; In the multi-task learning framework, the loss function weights of each sub-task for predicting electricity, cooling, and heating loads are dynamically adjusted through the attention mechanism to perform joint training and learning. Step S5: Based on the learning results of the multi-task learning framework, the predicted values of electrical load, cooling load and heating load for future moments are synchronously output through their respective independent output layers.
[0008] Preferably, in step S1, the feature selection based on correlation analysis specifically involves: calculating the unified information coefficient between each meteorological feature and any one of the electricity, cooling, and heating loads; if the unified information coefficient between a certain meteorological feature and any target load is greater than the preset correlation coefficient threshold, then the feature is retained; otherwise, it is removed from the feature set.
[0009] Preferably, in step S2, the noise reduction process includes the following steps: Step S21: Using the complete set empirical mode decomposition method, the historical load sequences of electricity, cooling, and heating are decomposed into multiple intrinsic mode function components and residuals respectively; Step S22: Calculate the sample entropy value of each intrinsic mode function component; based on the preset sample entropy threshold, reconstruct the components with sample entropy values higher than the threshold as high-frequency components, and reconstruct the components with sample entropy values lower than or equal to the threshold as low-frequency components. Step S23: Perform wavelet threshold denoising on the high-frequency components, while keeping the low-frequency components unchanged; Step S24: Reconstruct the denoised high-frequency components and low-frequency components to obtain the denoised electrical, cold, and heat load sequences respectively.
[0010] Preferably, in step S23, the wavelet thresholding denoising process employs a soft-hard compromise threshold function: ; in, This is a threshold function that represents a trade-off between hard and soft approaches. For symbolic functions, These represent the detail coefficients after wavelet decomposition. This indicates that a threshold value is set. Represents the compromise factor, when When, it degenerates into a soft threshold; when When, it degenerates into a hard threshold; when At that time, a soft-hard trade-off threshold is achieved in the contraction amplitude.
[0011] Preferably, in step S3, the relevance adjacency matrix dynamically constructed from the input is constructed as follows: taking each feature and target load in the filtered feature set as nodes, the relevance metric between any two nodes is calculated as the weight of the corresponding edge, thereby forming an adjacency matrix; the adjacency matrix is dynamically calculated and updated according to different input time window data.
[0012] Preferably, in step S3, the graph convolutional network and the bidirectional long short-term memory network are parallel connection structures, both of which receive the same input sequence and extract features independently.
[0013] Preferably, in step S4, dynamically adjusting the weights of the loss function through the attention mechanism specifically includes: Step S41: Calculate the loss values for the electricity, cooling, and heating load prediction sub-tasks respectively; Step S42: Input the loss value of each subtask into the attention module, and calculate the corresponding dynamic weight coefficients through the Softmax function; Step S43: Use dynamic weight coefficients to perform a weighted summation of the loss values of each sub-task to obtain the overall loss function for multi-task learning, and use it for backpropagation and parameter update of the model.
[0014] This invention also provides a multi-element load forecasting system for integrated energy systems, comprising: The data acquisition and feature filtering module is used to acquire historical multi-variable load data and related meteorological data of the integrated energy system, and to calculate the correlation between each meteorological feature and each target load using the unified information coefficient method. Based on the preset threshold, strongly correlated features are filtered out to form a filtered feature set. The data denoising module is used to denoise historical multivariate load data by employing a combined method of complete set empirical mode decomposition combining sample entropy and wavelet threshold denoising, to obtain denoised electric, cold, and hot load sequences. The spatiotemporal feature extraction module takes the filtered feature set and the denoised load sequence as input and performs feature extraction through a parallel model of graph convolutional network and bidirectional long short-term memory network. Specifically, the graph convolutional network is used to extract the spatial features of the multivariate load based on the correlation adjacency matrix dynamically constructed from the input, while the bidirectional long short-term memory network is used to extract the temporal features of the multivariate load. The extracted spatial features and temporal features are then concatenated to obtain the fused spatiotemporal depth features. The multi-task learning module is used to input the fused spatiotemporal deep features into the shared layer of the attention-based multi-task learning framework. The loss function weights of each sub-task for predicting electricity, cooling, and heating loads are dynamically adjusted through the attention mechanism for joint training and learning. The prediction output module is used to synchronously output the predicted values of electrical load, cooling load, and heating load for future moments based on the learning results of the multi-task learning framework through their respective independent output layers.
[0015] The present invention also provides a computer device, including: a memory and a processor; the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described integrated energy system multi-load forecasting system.
[0016] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described integrated energy system multi-load forecasting system.
[0017] Therefore, the present invention employs the above-mentioned method, system, equipment, and medium for predicting multiple loads in an integrated energy system, and the beneficial technical effects are as follows: (1) This invention constructs a parallel model of graph convolutional network and bidirectional long short-term memory network as a shared layer for multi-task learning, extracts spatial and temporal features at the same time, and uses attention mechanism to dynamically optimize loss weights, effectively overcoming the problems of insufficient spatiotemporal feature mining and insufficient coupling relationship learning caused by fixed task weights in existing methods, thereby achieving more comprehensive and accurate extraction of deep correlation features of multivariate loads.
[0018] (2) This invention replaces the static graph structure relied upon in traditional graph convolutional networks by dynamically constructing a correlation adjacency matrix based on input data, enabling the model to adaptively reflect the relationship between load and influencing factors as time and operating conditions change. This solves the problems of weak modeling ability and poor generalization of existing methods when data is sparse or relationships change dynamically.
[0019] (3) This invention introduces a joint preprocessing step at the model input end, combining complete set empirical mode decomposition with sample entropy and wavelet threshold denoising, to specifically filter out high-frequency noise and irrelevant fluctuations in the original load sequence. This solves the problem of directly using the original noise data for predictive interference model learning, providing a cleaner and more stable input for subsequent deep feature extraction, and fundamentally improving prediction reliability. Attached Figure Description
[0020] Figure 1 Flowchart for CEEMD-SE-WTD processing; Figure 2 For graph convolutional network structures; Figure 3 It is a bidirectional long short-term memory network structure; Figure 4 This forms the framework for the IES multivariate load forecasting model. Figure 5 The results of multivariate load prediction for ablation experiments, including, Figure 5 (a) in the figure represents the 72-hour electrical load forecast. Figure 5 (b) shows the predicted cooling load for 72 hours. Figure 5 (c) in the figure represents the predicted heat load for 72 hours; Figure 6 To compare the results of the multivariate load prediction experiment with the model, among which, Figure 6 (a) in the figure represents the 72-hour electrical load forecast. Figure 6 (b) in the figure represents the predicted cooling load for 72 hours. Figure 6 (c) in the figure represents the predicted heat load over 72 hours. Detailed Implementation
[0021] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0022] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.
[0023] Example 1 like Figure 4 As shown, a multi-element load forecasting method for integrated energy systems includes the following steps: Data acquisition, correlation analysis, and feature selection.
[0024] Step S1: Obtain historical electricity, cooling, and heating load data and related meteorological data (such as temperature and humidity) of the integrated energy system. First, perform correlation analysis. Since there are complex linear and nonlinear relationships between loads and meteorological factors in the IES, this embodiment uses the Uniform Information Coefficient (UIC) for measurement.
[0025] Calculate the UIC value between each meteorological feature and each target load (electricity, cooling, and heating). The specific rule for feature selection is as follows: if the UIC of a certain meteorological feature with any target load (electricity, cooling, or heating) is greater than a preset correlation coefficient threshold, it is determined to be a strongly correlated feature and retained; otherwise, it is removed from the initial feature set. The final filtered feature set is used as input to the model.
[0026] This step aims to reduce irrelevant feature inputs, avoid model overfitting, and improve computational efficiency and robustness.
[0027] Step S2: Denoising of the load sequence based on the CEEMDAN-SE-WTD joint method.
[0028] To mitigate the interference of surface noise in historical load sequences on deep feature extraction, this embodiment employs a joint method combining complete ensemble empirical mode decomposition (CEEMDAN) with sample entropy and wavelet thresholding for denoising. The process is as follows: Figure 1 As shown.
[0029] Step S21, CEEMDAN Decomposition: The complete ensemble empirical mode decomposition method is used to decompose the historical load sequences of electricity, cooling, and heating into multiple intrinsic mode function components and residuals. This method effectively suppresses mode aliasing by adding adaptive white noise and performing multiple EMD decompositions and ensemble averagings. Its core steps include adding noise, EMD decomposition, calculating the IMF mean, and iteratively updating the residuals until the residuals are monotonic functions.
[0030] Step S22, IMF component classification based on sample entropy (SE): Calculate the sample entropy value of each intrinsic mode function component. Sample entropy is used to quantify the complexity of a time series; a higher entropy value indicates a more complex sequence and a greater likelihood of noise. Based on a preset sample entropy threshold, components with sample entropy values higher than the threshold are reconstructed as high-frequency components, and components with sample entropy values lower than or equal to the threshold are reconstructed as low-frequency components.
[0031] Step S23, Wavelet Threshold Denoising (WTD): Perform wavelet threshold denoising on high-frequency components, while keeping low-frequency components unchanged; Wavelet thresholding denoising uses a compromise threshold function between soft and hard thresholding: (1) Select appropriate wavelet basis functions and decomposition levels to decompose the original sequence into approximation coefficients and detail coefficients.
[0032] (2) For detail coefficients Applying a threshold function to suppress noise-dominated wavelet coefficients, the threshold is set to... Choose the soft-hard tradeoff threshold function, which is defined as follows: ; in, This is a threshold function that represents a trade-off between hard and soft approaches. For symbolic functions, These represent the detail coefficients after wavelet decomposition. This indicates that a threshold value is set. Represents the compromise factor, when When, it degenerates into a soft threshold; when When, it degenerates into a hard threshold, when At this time, a soft-hard trade-off threshold is achieved in the contraction amplitude, striking a balance between noise reduction effect and signal fidelity. Low-frequency components are left unprocessed.
[0033] Step S24, Sequence Reconstruction: Reconstruct the denoised high-frequency components and low-frequency components to obtain the denoised electrical, cold, and heat load sequences respectively.
[0034] Step S3: Spatiotemporal depth feature extraction based on the GCN-BiLSTM parallel model.
[0035] The filtered feature set obtained in step S1 and the denoised load sequence obtained in step S2 are used as model inputs, and spatiotemporal depth features are extracted through parallel graph convolutional networks and bidirectional long short-term memory networks.
[0036] Step S31, Dynamic Graph Construction and Spatial Feature Extraction (GCN): GCN is used to extract spatial correlation features between multivariate loadings and features. Its structure is as follows: Figure 2 As shown. First, a graph structure needs to be constructed for the time-series data. In this embodiment, each feature in the filtered feature set and the target electrical, cooling, and heating loads are used as graph nodes. The correlation metric (such as UIC) between any two nodes calculated in step S1 is used as the weight of the corresponding edge to construct a dynamic correlation adjacency matrix. This adjacency matrix can be recalculated based on the input data in different time windows, thereby dynamically reflecting the changes in the relationship between nodes. This adjacency matrix and node features are input into the GCN, and neighbor information is aggregated through the inter-layer propagation formula to extract and fuse spatial local and global features. The inter-layer propagation formula is as follows: ; Among them, the standardized adjacency matrix , Represents the adjacency matrix. Represents the identity matrix; express The degree matrix; express The feature matrix of the layer; express The learnable weight matrix of the layer, This represents a non-linear activation function; here, the ReLU function is used.
[0037] Step S32, Temporal Feature Extraction (BiLSTM): BiLSTM is used to extract long-term temporal dependency features of the payload sequence. Its structure is as follows: Figure 3 As shown, it consists of a forward LSTM and a backward LSTM, capable of simultaneously capturing both forward and backward information of the time series. Its hidden layer states are jointly determined by the forward and backward LSTMs, and information flow is managed through a gating mechanism, effectively avoiding the gradient vanishing / exploding problem, thus fully learning the complex temporal patterns of the load sequence.
[0038] ; in, and These represent the hidden layer states of the forward and backward LSTM at the current time. For the current input, and The output weights of the hidden layers for the forward and backward views are respectively given at the current time step. The weights of the hidden layer output at the current moment.
[0039] Step S33, Feature Fusion: The spatial feature matrix extracted by GCN is concatenated with the temporal feature sequence extracted by BiLSTM to form a deep feature representation that integrates spatiotemporal information. In this embodiment, GCN and BiLSTM adopt a parallel connection structure, both receiving the same input sequence simultaneously and extracting features independently, avoiding information loss that may occur with a serial structure.
[0040] Step S4: Joint learning and prediction based on the Attention-MTL framework.
[0041] The fused spatiotemporal depth features obtained in step S3 are input into the shared layer of the attention-based multi-task learning framework to jointly predict the electricity, cooling, and heating loads.
[0042] Step S41, Multi-task Learning Framework: This embodiment constructs a parameter-hard-shared MTL framework. This framework includes a shared layer (i.e., the GCN-BiLSTM parallel model in step S3) and multiple task-specific output layers. The shared layer learns a general feature representation of all loads, while each load (electric, cold, heat) has an independent output layer (fully connected layer) for final prediction.
[0043] Step S42: Attention mechanism optimizes loss weights: In order to balance the learning process of different prediction tasks and avoid one task dominating the training, this embodiment introduces an attention mechanism in MTL to dynamically adjust the weights of the loss functions of each task.
[0044] Specifically, in multi-task learning, let the task be... The dataset is ,in and These represent the number of tasks and the number of data samples, respectively. Indicates the first The sampling point at the th sampling point The labels in each task. The loss function for multi-task learning is a weighted sum of the loss functions of each task, defined as follows: ; in, Indicates the weighting coefficient. Represents the loss function. This represents shared parameters among multiple tasks. This represents unique parameters for each task.
[0045] In the design of the MTL loss function, the reasonable allocation of weights for each task can balance the importance of each task, avoiding problems such as some tasks dominating gradient backpropagation, preventing other tasks from learning and optimizing sufficiently, and inconsistent convergence speeds among different tasks. Therefore, this invention optimizes the weight allocation of the MTL loss function based on the Attention mechanism, dynamically adjusting task weights during the learning process, enabling the model to better adapt to the needs of different tasks. The principle is as follows: ; in, The weights of the loss function obtained for each task through the attention mechanism are defined, with the activation function of the attention mechanism being softmax. By dynamically adjusting the weights through the attention mechanism, the focus on different tasks can be adaptively adjusted, thereby learning richer feature representations and improving the model's generalization ability.
[0046] Step S5: Output the results.
[0047] Based on the training results of the multi-task learning framework in step S4, for new input data, after feature extraction in the shared layer, the predicted values of electrical, cooling, and heating loads will be synchronously calculated and output through their respective independent output layers. To prevent overfitting, a Dropout layer can be added before the fully connected layer.
[0048] To comprehensively and objectively measure the predictive performance of the prediction model, the root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and weighted mean absolute percentage error (WMAPE) are selected as evaluation metrics, defined as follows: ; ; ; ; in, This represents the total number of test samples. Indicates the sample index. and These represent the actual value and the predicted value, respectively. These represent the weighting factors for the three types of loads: electricity, cooling, and heating.
[0049] The invention will be further illustrated below with specific examples.
[0050] (1) Data preprocessing.
[0051] First, to avoid the negative impact of abnormal data input, missing and outlier values in the dataset are handled. Through detection, it is confirmed that there are no missing values in the dataset. Outliers are identified using the Z-score method. Considering the characteristic of large local fluctuations but a stable overall trend in the load sequence, piecewise cubic Hermite interpolation is used for imputation.
[0052] Secondly, to avoid the influence of different units of measurement and ensure that the input data has a unified measurement standard, the data is subjected to min-max normalization.
[0053] (2) Parameter settings.
[0054] This invention uses IES data from a certain campus from February 1, 2021 to February 27, 2021 as the test set, and divides the remaining data into training and validation sets in a 9:1 ratio. The experimental environment is built based on Python 3.9 and the PyTorch framework, and PyCharm is used to implement the simulation examples.
[0055] After multiple trials and adjustments, the training batch size was set to 64, the input sequence length to 24, the number of training epochs to 100, the learning rate to 0.001, and the hidden layer size to 128. For the shared layer model, the prediction performance of the model at different depths of GCN was compared through controlled variable experiments, and the results are shown in Table 1.
[0056] Table 1. Prediction results at different GCN depths
[0057] Analysis and comparison of the experimental results show that when the depth of each GCN layer is selected as 3, the WMAPE is the lowest and the prediction effect is the best. Therefore, the depth of each GCN layer is selected as 3. Next, the prediction performance of the model under different numbers of GCN and BiLSTM layers is compared, and the results are shown in Tables 2 and 3.
[0058] Table 2 Prediction results for different GCN layers
[0059] Table 3. Prediction results for different numbers of BiLSTM layers
[0060] Analysis of the controlled variable experiments revealed that WMAPE was smaller and the overall prediction performance was best when both the GCN and BiLSTM layers were set to 1. Therefore, the number of GCN and BiLSTM layers was set to 1. Furthermore, the dropout size was set to 0.3, and the Adam optimizer was used to train the model.
[0061] (3) Comparative analysis of joint forecasting and independent forecasting.
[0062] To verify the effectiveness of the multi-task learning joint prediction method of this invention compared with the independent prediction method for different types of loads, a single-task learning framework was constructed based on the parallel CSW-GCN-BiLSTM network model. Independent predictions were then performed on electricity, cooling, and heating loads, and the MTL joint prediction results were compared with the STL independent prediction results. Based on the actual proportions of electricity, cooling, and heating loads in the IES load, the weighting factors for electricity, cooling, and heating loads in WMAPE were set to 0.5, 0.4, and 0.1, respectively. The results are shown in Table 4.
[0063] Table 4 Results of Joint and Independent Forecasts of Multiple Load Factors
[0064] As shown in Table 4, when using the same model, multi-task learning significantly outperforms single-task learning in predicting IES multivariate loads, reducing the WMAPE by 1.19%. This is because there is a significant coupling relationship between electrical, cooling, and heating loads. By sharing underlying feature representations among multivariate loads, MTL can better learn the intrinsic correlations between them and capture the interdependencies between different load types. Furthermore, compared to single-task learning, multi-task learning reduces the risk of overfitting and enhances the model's robustness and generalization ability by jointly training multiple related tasks. Moreover, multi-task learning is significantly more efficient, with training time only half that of single-task learning. This is because MTL typically requires fewer computational resources than training multiple independent single-task models, and its sharing mechanism reduces the number of parameters, making model training more efficient.
[0065] (4) Analysis of ablation test results.
[0066] To verify the effectiveness of each component of the proposed prediction model based on parallel CSW-GCN-BiLSTM-MTL-Attention, an ablation experiment was conducted. Each component of the combined prediction model was removed sequentially, and the parallel GCN-BiLSTM-MTL-Attention model (P-GBMA), the CSW-BiLSTM-MTL-Attention model (CBMA), and the CSW-GCN-BiLSTM-MTL model with fixed loss function weights (P-CGBM) were constructed and compared with the proposed model (P-CGBMA) to explore the role of CEEMDAN-SE-WTD denoising on the load sequence, the effectiveness of using GCN to extract spatial features, and the necessity of introducing an attention mechanism in MTL to dynamically adjust the loss function weights. The ablation experiment results are shown in Table 5. To make the prediction results more intuitive, the 72-hour load prediction curves for each model are also shown in Table 5. Figure 5 As shown.
[0067] Table 5 Ablation Experiment Results
[0068] Analysis of the ablation experiment results shows that each step plays a crucial role in the combined model. Compared to P-GBMA, after applying CEEMDan-SE-WTD denoising to the historical load sequences, the WMAPE decreased by 0.88%, demonstrating that CSW denoising effectively filters noise from the load sequences, improves data quality, and optimizes learning performance. Compared to CBMA, adding GCN to the shared layer model to extract spatial features resulted in a 0.48% decrease in WMAPE, and the prediction accuracy for each load improved. This is because GCN can fuse local spatial features with higher-level global features, effectively capturing the complex relationships between different types of loads and between loads and meteorological features, leading to more comprehensive feature mining. Compared to P-CGBM, introducing an Attention mechanism to dynamically adjust the weights of the multi-task learning loss function significantly improved the cold load prediction accuracy, with a 0.31% decrease in WMAPE. This is because when using a fixed-weight loss function for multi-task learning, the model does not pay sufficient attention to the cold load training task. Introducing an Attention mechanism to dynamically adjust the weights better balances the learning process of different tasks, allowing each task to be fully optimized, thereby improving the overall prediction performance. This demonstrates that the introduction of each component of the combined model has led to performance improvements, and there is a good synergistic effect among the components of the model, which improves the performance of the model in mining load depth correlation features, thus proving the rationality of the model design.
[0069] (5) Comparative analysis of prediction models.
[0070] To further verify the effectiveness of the proposed IES multivariate load prediction model, the model of this invention was compared with various IES multivariate load joint prediction models based on multi-task learning frameworks in recent studies in the same test environment, including the PCC-LSTM-MTL model (PLM), MTL-GRU-Attention model (MGA), BiLSTM-MTL model (BM), and XGBoost-GRU-MTL model (XGM) from the literature. Simultaneously, a serial structure CSW-GCN-BiLSTM-MTL-Attention model (S-CGBMA) and a CSW-CNN-BiLSTM-MTL-Attention model (P-CCBMA) that extracts spatiotemporal features in parallel using CNN and BiLSTM were constructed and compared with the proposed model (P-CGBMA) to explore the advantages of parallel structures and the superiority of GCN over CNN in extracting IES multivariate load spatial features. The hyperparameters of each model were fully debugged. The prediction results for electricity, cooling, and heating loads obtained through the above prediction models are shown in Table 6, and their 72-hour prediction curves are shown in... Figure 6 As shown.
[0071] Table 6. Prediction results of various IES multivariate load forecasting models
[0072] comprehensive Figure 6 Comparative analysis with Table 6 shows that the proposed parallel CSW-GCN-BiLSTM-MTL-Attention prediction model outperforms other models in all evaluation metrics, and its prediction curve better matches the actual load curve. The other five models, built using a multi-task learning framework and considering the coupling characteristics between various load types, all show good prediction results, but the prediction error of the proposed model is significantly lower. Compared to the six models PLM, MGA, BM, XGM, S-GCBMA, and P-CCBMA, the WMAPE of the proposed model is reduced by 1.18%, 1.05%, 0.93%, 0.89%, 0.41%, and 0.28%, respectively, demonstrating higher prediction accuracy. Furthermore, compared to other models, the proposed model can more accurately predict the load curve trend when load fluctuates significantly, with a better fit. This is because the proposed model utilizes GCN to deeply mine spatial features and dynamically models node relationships through adjacency matrices, making it more suitable for the complex coupling characteristics of IES loads. Furthermore, the parallel extraction of spatiotemporal features by combining GCN and BiLSTM yields more comprehensive features compared to sequential structure feature extraction. CEEMDAN-SE-WTD is used to filter out superficial interference information in historical load sequences, enabling the model to extract more effective and deeper-level correlated features. Simultaneously, the model utilizes an attention mechanism to dynamically adjust the weights of the loss function, balancing the importance of each task in the MTL, fully exploring the complex coupling information between diverse loads, and enhancing the model's generalization ability.
[0073] (6) Diebold-Mariano test.
[0074] To further verify the superiority of the model of this invention compared with other models, the Diebold-Mariano test (DM test) is used to compare and analyze the prediction results of the model of this invention with those of the aforementioned models. The DM test constructs a loss difference sequence based on MSE to examine whether the differences in prediction errors among different models are caused by random factors.
[0075] First, using the model of this invention as the benchmark model, the prediction results for electricity, cooling, and heating loads were subjected to DM tests respectively. Simultaneously, due to the strong coupling between the IES multi-load components, the prediction results of the three loads need to be comprehensively considered during the test. Therefore, this invention innovatively proposes a comprehensive DM test method, which uses the weighted MSE of the three loads as the loss difference sequence to perform multi-load DM tests. The results are shown in Table 7. The weighting factors are set the same as in the WMAPE above, with electricity, cooling, and heating loads set to 0.5, 0.4, and 0.1, respectively.
[0076] Table 7 DM Test Results of Various IES Multivariate Load Forecasting Models
[0077] The test results show that, at the 5% significance level, the model of this invention demonstrates significant superiority over PLM, MGA, BM, XGM, and S-CGBMA models in predicting electricity, cooling, and heating loads, as well as in the overall prediction performance of multiple loads. While the model of this invention does not show a significant difference in electricity load prediction compared to P-CCBMA, it demonstrates significant superiority in predicting both cooling and heating loads, thus exhibiting significant superiority in the overall performance of multiple load prediction. In conclusion, the DM test further proves the necessity of the prediction model fully considering deep correlation characteristics, theoretically verifying that the improved prediction accuracy of the model of this invention is not accidental and has substantial significance.
[0078] Example 2 A multi-source load forecasting system for integrated energy systems, comprising: The data acquisition and feature filtering module is used to acquire historical multi-variable load data and related meteorological data of the integrated energy system, and to calculate the correlation between each meteorological feature and each target load using the unified information coefficient method. Based on the preset threshold, strongly correlated features are filtered out to form a filtered feature set. The data denoising module is used to denoise historical multivariate load data by employing a combined method of complete set empirical mode decomposition combining sample entropy and wavelet threshold denoising, to obtain denoised electric, cold, and hot load sequences. The spatiotemporal feature extraction module takes the filtered feature set and the denoised load sequence as input and performs feature extraction through a parallel model of graph convolutional network and bidirectional long short-term memory network. Specifically, the graph convolutional network is used to extract the spatial features of the multivariate load based on the correlation adjacency matrix dynamically constructed from the input, while the bidirectional long short-term memory network is used to extract the temporal features of the multivariate load. The extracted spatial features and temporal features are then concatenated to obtain the fused spatiotemporal depth features. The multi-task learning module is used to input the fused spatiotemporal deep features into the shared layer of the attention-based multi-task learning framework. The loss function weights of each sub-task for predicting electricity, cooling, and heating loads are dynamically adjusted through the attention mechanism for joint training and learning. The prediction output module is used to synchronously output the predicted values of electrical load, cooling load, and heating load for future moments based on the learning results of the multi-task learning framework through their respective independent output layers.
[0079] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0080] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.
[0081] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.
[0082] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.
[0083] Therefore, the present invention adopts the above-mentioned method, system, equipment and medium for predicting multiple loads in a comprehensive energy system, which can effectively improve the quality of input data through joint denoising preprocessing, and deeply explore the complex correlation between multiple loads by means of parallel spatiotemporal feature extraction network and dynamically optimized multi-task learning framework. Thus, while improving the prediction accuracy, it enhances the generalization ability and robustness of the model in different scenarios, and the overall architecture has high computational efficiency.
[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A multi-element load forecasting method for integrated energy systems, characterized in that, Includes the following steps: Step S1: Obtain historical multivariate load data and related meteorological data of the integrated energy system; calculate the correlation between each meteorological feature and each target load using the unified information coefficient method, and select strongly correlated features based on preset thresholds to form a filtered feature set; Step S2: The historical multivariate load data is denoised using a combined method of complete set empirical mode decomposition combining sample entropy and wavelet threshold denoising to obtain the denoised electric, cold and hot load sequences. Step S3: Using the filtered feature set and the denoised load sequence as input, construct and use a parallel model of graph convolutional network and bidirectional long short-term memory network for feature extraction; wherein, using graph convolutional network, based on the correlation adjacency matrix dynamically constructed from the input, the spatial features of multivariate loads are extracted; at the same time, the temporal features of multivariate loads are extracted using bidirectional long short-term memory network; the extracted spatial features and temporal features are concatenated to obtain fused spatiotemporal depth features. Step S4: Input the fused spatiotemporal depth features into the shared layer of the attention-based multi-task learning framework; In the multi-task learning framework, the loss function weights of each sub-task for predicting electricity, cooling, and heating loads are dynamically adjusted through the attention mechanism to perform joint training and learning. Step S5: Based on the learning results of the multi-task learning framework, the predicted values of electrical load, cooling load and heating load for future moments are synchronously output through their respective independent output layers.
2. The method for predicting multiple loads in an integrated energy system according to claim 1, characterized in that, In step S1, the feature selection based on correlation analysis is specifically as follows: calculate the unified information coefficient between each meteorological feature and any one of the electricity, cooling, and heating loads; if the unified information coefficient between a certain meteorological feature and any target load is greater than the preset correlation coefficient threshold, then the feature is retained; otherwise, it is removed from the feature set.
3. The method for predicting multiple loads in an integrated energy system according to claim 1, characterized in that, In step S2, the noise reduction process includes the following steps: Step S21: Using the complete set empirical mode decomposition method, the historical load sequences of electricity, cooling, and heating are decomposed into multiple intrinsic mode function components and residuals respectively; Step S22: Calculate the sample entropy value of each intrinsic mode function component; based on the preset sample entropy threshold, reconstruct the components with sample entropy values higher than the threshold as high-frequency components, and reconstruct the components with sample entropy values lower than or equal to the threshold as low-frequency components. Step S23: Perform wavelet threshold denoising on the high-frequency components, while keeping the low-frequency components unchanged; Step S24: Reconstruct the denoised high-frequency components and low-frequency components to obtain the denoised electrical, cold, and heat load sequences respectively.
4. The multi-element load forecasting method for an integrated energy system according to claim 3, characterized in that, In step S23, the wavelet thresholding denoising process uses a compromise threshold function between soft and hard thresholding: ； in, This is a threshold function that represents a trade-off between hard and soft approaches. For symbolic functions, These represent the detail coefficients after wavelet decomposition. This indicates that a threshold value is set. Represents the compromise factor, when When, it degenerates into a soft threshold; when When, it degenerates into a hard threshold; when At that time, a soft-hard trade-off threshold is achieved in the contraction amplitude.
5. The method for predicting multiple loads in an integrated energy system according to claim 1, characterized in that, In step S3, the relevance adjacency matrix dynamically constructed from the input is constructed as follows: taking each feature and target load in the filtered feature set as nodes, the relevance metric between any two nodes is calculated as the weight of the corresponding edge, thereby forming the adjacency matrix; this adjacency matrix is dynamically calculated and updated according to different input time window data.
6. The method for predicting multiple loads in an integrated energy system according to claim 1, characterized in that, In step S3, the graph convolutional network and the bidirectional long short-term memory network are parallel connection structures. Both receive the same input sequence at the same time and extract features independently.
7. The method for predicting multiple loads in an integrated energy system according to claim 1, characterized in that, In step S4, dynamically adjusting the weights of the loss function through the attention mechanism specifically includes: Step S41: Calculate the loss values for the electricity, cooling, and heating load prediction sub-tasks respectively; Step S42: Input the loss value of each subtask into the attention module, and calculate the corresponding dynamic weight coefficients through the Softmax function; Step S43: Use dynamic weight coefficients to perform a weighted summation of the loss values of each sub-task to obtain the overall loss function for multi-task learning, and use it for backpropagation and parameter update of the model.
8. A multi-element load forecasting system for integrated energy systems, characterized in that, A method for implementing a multi-source load forecasting method for an integrated energy system as described in any one of claims 1-7, comprising: The data acquisition and feature filtering module is used to acquire historical multi-variable load data and related meteorological data of the integrated energy system, and to calculate the correlation between each meteorological feature and each target load using the unified information coefficient method. Based on the preset threshold, strongly correlated features are filtered out to form a filtered feature set. The data denoising module is used to denoise historical multivariate load data by employing a combined method of complete set empirical mode decomposition combining sample entropy and wavelet threshold denoising, to obtain denoised electric, cold, and hot load sequences. The spatiotemporal feature extraction module takes the filtered feature set and the denoised load sequence as input and performs feature extraction through a parallel model of graph convolutional network and bidirectional long short-term memory network. Specifically, the graph convolutional network is used to extract the spatial features of the multivariate load based on the correlation adjacency matrix dynamically constructed from the input, while the bidirectional long short-term memory network is used to extract the temporal features of the multivariate load. The extracted spatial features and temporal features are then concatenated to obtain the fused spatiotemporal depth features. The multi-task learning module is used to input the fused spatiotemporal deep features into the shared layer of the attention-based multi-task learning framework. The loss function weights of each sub-task for predicting electricity, cooling, and heating loads are dynamically adjusted through the attention mechanism for joint training and learning. The prediction output module is used to synchronously output the predicted values of electrical load, cooling load, and heating load for future moments based on the learning results of the multi-task learning framework through their respective independent output layers.
9. A computer device, comprising: Memory and processor; The memory stores a computer program, characterized in that when the processor executes the computer program, it implements the steps of the multi-element load forecasting method for an integrated energy system as described in any one of claims 1-7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the multi-element load forecasting method for an integrated energy system as described in any one of claims 1-7.