Comprehensive energy multi-element load forecasting method based on full-stage multi-task learning architecture
By adopting a full-stage multi-task learning architecture, the problems of insufficient feature representation and task conflict in multi-dimensional load forecasting in integrated energy systems are solved, and efficient and accurate load forecasting is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- EAST CHINA JIAOTONG UNIVERSITY
- Filing Date
- 2026-02-12
- Publication Date
- 2026-05-29
AI Technical Summary
Existing technologies for multi-load forecasting in integrated energy systems suffer from problems such as insufficient depth and breadth of feature representation, inefficient and unbalanced utilization of model capacity, and easy occurrence of task conflicts in multi-task learning.
We adopt a method based on a full-stage multi-task learning architecture. Through a multi-task feature extraction expert module, a MOE router trained with load balancing loss, and the HS-xLSTM framework, we perform feature extraction, dynamic filtering, and fusion. By combining feature attention, multi-scale causal dilation convolution, and temporal attention, we can improve the depth and breadth of multi-task features. We also resolve task conflicts through load balancing loss and parameter sharing strategies.
It improves the performance of multivariate load forecasting, enhances the depth and breadth of feature representation, ensures efficient utilization of model capacity, alleviates task conflicts, and improves forecast accuracy and stability.
Smart Images
Figure CN121684553B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology, and more specifically to a comprehensive energy multi-load forecasting method based on a full-stage multi-task learning architecture. Background Technology
[0002] Against the backdrop of global efforts to address climate change, the energy system is undergoing a profound transformation from fossil fuel-dominated to renewable energy-driven. Integrated Energy Systems (IES), as the core carrier of this energy revolution, integrate multiple energy sources such as electricity, cooling, and heating at the regional level, achieving multi-energy complementarity and cascaded utilization. This highly integrated model demonstrates enormous potential in improving overall energy efficiency and promoting renewable energy utilization, but its complex internal coupling relationships also pose unprecedented challenges to system operation and control. In IES, various energy sources are closely interconnected through combined cooling, heating, and power (CCHP), electricity-to-heat, and cooling equipment, forming a dynamic, interconnected network. Therefore, accurately predicting future energy demand changes within the system has become the primary prerequisite and key link in ensuring its safe, economical, and low-carbon operation.
[0003] To overcome the challenge of multivariate load co-forecasting in IES (Enhanced Environmental Load) systems, researchers have explored various technical approaches. Early studies focused primarily on traditional statistical models, such as Vector Autoregression (VAR), which, while capable of handling multivariate time series, suffer from inherent linear assumptions that make it difficult to capture the strong nonlinear dynamics prevalent in load data. Subsequently, machine learning methods, such as Support Vector Regression (SVR) and Shallow Artificial Neural Networks (ANNs), improved prediction accuracy to some extent due to their nonlinear mapping capabilities. However, these methods often rely on cumbersome manual feature engineering and have shallow model structures, making it difficult to learn deep abstract patterns from high-dimensional, long-term time series data. In recent years, deep learning, with its superior automatic feature extraction and complex pattern recognition capabilities, has brought revolutionary breakthroughs to this field and has gradually become the mainstream research method. Among them, Recurrent Neural Networks (RNNs) and their variants, Long Short-Term Memory Networks (LSTM) and Gated Recurrent Units (GRUs), have shown natural advantages in processing time series data due to their unique "memory" mechanism and have become the mainstream choice in the field of load forecasting.
[0004] However, directly applying existing deep learning models to IES multivariate load forecasting still faces many challenges, mainly in the following aspects: 1) insufficient depth and breadth of feature representation; 2) inefficient and unbalanced utilization of model capacity; and 3) task conflicts easily occur in multi-task learning. These problems affect the performance of multivariate load forecasting. Summary of the Invention
[0005] The purpose of this invention is to provide a comprehensive energy multi-dimensional load forecasting method based on a full-stage multi-task learning architecture, in order to solve the problems of insufficient depth and breadth of feature representation, inefficient and unbalanced utilization of model capacity, and easy task conflicts in multi-task learning in the existing technology, thereby improving the performance of multi-dimensional load forecasting.
[0006] A comprehensive energy multi-variable load forecasting method based on a full-stage multi-task learning architecture includes:
[0007] Step S1: Collect multivariate load datasets. First, use moving average to find outliers and missing values due to statistical errors, and then use linear interpolation to fill them. Then, based on seasonal characteristics and data coupling relationships, divide the filled datasets into summer datasets and winter datasets.
[0008] Step S2: Perform correlation analysis on the climate characteristics and loads in the summer and winter datasets respectively, calculate the correlation analysis index, and screen out the feature sequences that are strongly correlated with the electrical load, cooling load and heating load based on the correlation analysis index to form the screened dataset;
[0009] Step S3: The filtered dataset is processed using a multi-task feature extraction expert module. The multi-task feature extraction expert module is used to extract feature information from two dimensions: variable association and time dependence, through feature attention, multi-scale causal dilation convolution and temporal attention, while introducing residual connections and causal masks.
[0010] Step S4: Based on the processing results of the multi-task feature extraction expert module, a MOE router trained on load balancing loss is used for dynamic filtering and fusion.
[0011] Step S5: For the feature fusion results of the MOE router, the HS-xLSTM framework is used for prediction. In the HS-xLSTM framework, the electrical load prediction task and the cold load prediction task are processed by the mLSTM layer with hard parameter sharing; the hot load prediction task is processed by the mLSTM layer with soft parameter sharing. Finally, each task is further processed by the sLSTM layer with soft parameter sharing for deep time series modeling, and the final prediction result is obtained by the fully connected layer.
[0012] The comprehensive energy multi-element load forecasting method based on a full-stage multi-task learning architecture provided by the present invention has the following beneficial effects:
[0013] 1. This invention designs a multi-task feature extraction expert module. Within each expert network, feature attention, multi-scale causal dilation convolution, and temporal attention are chained together. This enables the multi-task feature extraction expert module to extract information simultaneously from two orthogonal dimensions: variable association and temporal dependency, greatly improving the depth and breadth of feature representation. Simultaneously, residual connections and causal masks are introduced to effectively alleviate the gradient vanishing problem in deep network training while strictly ensuring no data leakage occurs.
[0014] 2. This invention employs an MOE router trained based on load balancing loss. Unlike traditional static expert allocation, this invention uses a learnable router to dynamically filter and integrate features from shared experts and task-specific experts for each task. By introducing load balancing loss, it effectively avoids the problem of insufficient utilization of expert modules and ensures efficient and balanced utilization of model capacity.
[0015] 3. This invention establishes an HS-xLSTM framework to address the commonalities and individual characteristics of modeling different tasks. This framework employs structured parameter sharing for task relevance, allowing highly correlated tasks such as electrical load forecasting and cooling load forecasting to be processed using hard-parameter-shared mLSTM layers, which promotes knowledge transfer and generalization. For heat load forecasting, soft-parameter-shared mLSTM layers are used, which effectively isolates tasks with significant differences while promoting positive transfer between related tasks through shared layers, thus avoiding negative transfer interference. This hierarchical sharing strategy effectively alleviates task conflicts in multi-task learning while maintaining model efficiency, thereby improving the performance of multivariate load forecasting.
[0016] 4. This invention constructs a correlation analysis index that combines the mean absolute values of Pearson and Spearman correlation coefficients to assess the strength of the association. A consistency penalty factor is introduced using standard deviation to penalize differences in the results of different coefficients. This correlation analysis index not only quantifies the strength of the association between variables but also provides a more comprehensive and reliable assessment of the true association between variables, effectively avoiding the limitations of single analysis methods. Attached Figure Description
[0017] Figure 1 This is a flowchart illustrating the integrated energy multi-load forecasting method based on a full-stage multi-task learning architecture provided by the present invention. Detailed Implementation
[0018] To facilitate understanding of the present invention, a more complete description will be given below with reference to various embodiments. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.
[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.
[0020] Please see Figure 1 The embodiments of the present invention provide a comprehensive energy multi-variable load forecasting method based on a full-stage multi-task learning architecture, including steps S1-S5:
[0021] Step S1: Collect a multivariate load dataset. First, use moving average to find outliers and missing values due to statistical errors, and then use linear interpolation to fill them. Then, based on seasonal characteristics and data coupling relationships, divide the filled dataset into summer dataset and winter dataset.
[0022] The process begins by using a moving average method to screen the collected multidimensional time series data, identifying missing data and outliers that deviate from the normal range in the original dataset. The identified outliers and missing values are then removed from the dataset and supplemented using linear interpolation.
[0023] The dataset is then divided into winter and summer parts for discussion. Based on seasonal characteristics and data coupling relationships, the filled dataset is divided into summer dataset and winter dataset.
[0024] Step S2: Perform correlation analysis on the climate characteristics and loads in the summer and winter datasets respectively, calculate the correlation analysis index, and screen out the feature sequences that are strongly correlated with the electrical load, cooling load and heating load based on the correlation analysis index to form the screened dataset.
[0025] The formula for calculating the correlation analysis index is as follows:
[0026]
[0027]
[0028]
[0029] in, Representing variables and variables Correlation analysis indicators between them Representing variables and variables The basic correlation strength between them Represents the consistency penalty factor. For variables and variables The Pearson correlation coefficient between them For variables and variables The Spearman correlation coefficient between them.
[0030] Taking the electrical load sequence in the predicted target and the temperature sequence in the input features as an example, we first calculate the absolute values of their Pearson and Spearman correlation coefficients, and take their arithmetic mean as the basic correlation strength. Then, we introduce a consistency penalty factor. When the two coefficients are highly consistent, it indicates that the result has high confidence, and no penalty is imposed. Conversely, if the two are significantly different, a penalty is imposed on the final score, thus obtaining the correlation analysis index between the electrical load sequence and the temperature sequence.
[0031] Similarly, correlation analysis indices between electricity, cooling, and heating loads and other input characteristics can be calculated, and then feature sequences that are strongly correlated with electricity, cooling, and heating loads, i.e. feature sequences whose correlation analysis indices are greater than the threshold, can be selected.
[0032] Step S3: The filtered dataset is processed using a multi-task feature extraction expert module. The multi-task feature extraction expert module is used to extract feature information from two dimensions: variable association and time dependence, through feature attention, multi-scale causal dilation convolution and temporal attention, while introducing residual connections and causal masks.
[0033] To overcome the limitations of conventional MOE expert modules in capturing temporal features, this invention integrates feature attention, multi-scale causal dilated convolution, and temporal attention mechanisms into its architecture, thereby achieving deep feature extraction from both variable association and temporal dependence dimensions. Simultaneously, by introducing residual connections and causal masks, it strictly ensures no data leakage and improves the training stability of deep networks. First, the feature attention mechanism calculates the correlation within the input sequence, dynamically weighting and aggregating features at different times to capture global long-distance dependencies and filter out the most critical historical features for the current prediction. A causal mask is also applied to prevent future information leakage. Next, multi-scale dilated convolution processes the input in parallel with various kernels of different sizes and dilation rates, simultaneously covering multi-scale receptive fields from local details to macro trends, thus efficiently capturing temporal features of different frequencies without increasing depth. Finally, the temporal attention mechanism calculates self-attention weights in the temporal dimension, enabling dynamic filtering and focusing on key historical time steps with the greatest impact on the current moment across time distances, enhancing the modeling ability of long-sequence dependencies, and again applying a causal mask to prevent future information leakage.
[0034] The complete multi-task feature extraction expert module includes shared feature extraction experts, electrical load feature extraction experts, cooling load feature extraction experts, and heating load feature extraction experts. The electrical load feature extraction experts, cooling load feature extraction experts, and heating load feature extraction experts are task-specific feature extraction experts. The internal structure of the feature extraction experts (whether shared or task-specific) is consistent.
[0035] In this embodiment, step S3 specifically includes:
[0036] S31, for the filtered dataset, it first undergoes a point-by-point linear transformation using a one-dimensional convolution with a kernel size of 1. This maps the original low-dimensional input features to the high-dimensional hidden layer feature space within the model, achieving feature dimension alignment and preliminary fusion, resulting in the data after one-dimensional convolutional dimensionality expansion. ;
[0037] S32, data after dimensionality expansion via one-dimensional convolution. The data are respectively fed into the task-sharing feature extraction expert, electrical load feature extraction expert, cold load feature extraction expert, and hot load feature extraction expert in the multi-task feature extraction expert module. A self-attention mechanism is used to calculate the correlation matrix between feature dimensions. The query matrix, key matrix, and value matrix are then calculated. A Softmax activation function is applied to calculate the feature attention weight matrix. Finally, scaling and residual joins are used to obtain the feature attention-enhanced data. To dynamically adjust the weight distribution of different meteorological and load characteristics;
[0038] S33, Data after feature attention enhancement The data is fed into a multi-scale dilated convolution module, which employs a multi-branch parallel structure, integrating convolutional kernels of different sizes and combining them with dilated convolution. This allows the model to significantly expand its receptive field without substantially increasing computational complexity, thereby simultaneously capturing both short-term local fluctuations and long-term periodic patterns, resulting in the multi-scale convolution output. The expression is;
[0039]
[0040] in, This represents the concatenation of multiple convolutional kernels of different scales. The kernel size is the convolutional kernel size. Indicates the kernel size as The dilated convolution kernel weights, yes At time step The input vector, For time steps, For the in-kernel indexing step, It is the expansion factor;
[0041] S34, the output of multi-scale convolution by the temporal attention mechanism. The attention scores between different time steps are calculated to identify and focus on the historical moments that have the greatest impact on the current prediction target. Finally, residual connections are used to obtain the output of the task-shared feature extraction expert. Output of electrical load feature extraction expert Output of Cooling Load Feature Extraction Expert Output of heat load feature extraction expert .
[0042] Step S4: Based on the processing results of the multi-task feature extraction expert module, a MOE router trained on load balancing loss is used for dynamic filtering and fusion.
[0043] Specifically, step S4 includes:
[0044] Step S41: The MOE router first receives the data after one-dimensional convolutional dimension expansion. And perform global average pooling along the time dimension to obtain a global context feature vector. This step compresses the high-dimensional feature matrix containing the time dimension into a feature vector containing global context information, aiming to eliminate the influence of local time fluctuations and provide a global view for routing decisions.
[0045] Step S42: Input the global context feature vector into a router network composed of MLPs. Through nonlinear mapping, calculate the nonnormalized confidence score of the current input sample for each expert in the multi-task feature extraction expert module. The nonnormalized confidence score is used to reflect the degree of matching of each expert in processing the current sample.
[0046] Step S43: The non-normalized confidence scores are transformed into a probability distribution using the Softmax function to determine the activation weights of all four experts, thereby constructing a sparse routing weight matrix. To prevent bias in expert utilization during the construction of the sparse routing weight matrix, this invention introduces a load balancing loss into the total loss function. This loss is calculated by... The average workload of each expert in the current batch The final calculated load balancing loss value Load balancing loss value It will be added to the total training loss and used for backpropagation to guide the MOE router to explore and utilize the capabilities of all experts during training.
[0047] Among them, load balancing loss value The expression is:
[0048]
[0049]
[0050] in, For the first The average workload undertaken by each expert in the current batch This represents the average usage rate across all experts. The total number of samples, Indicates the first The first sample selection The probability of an expert.
[0051] Step S44: Based on the sparse routing weight matrix, use batch matrix multiplication to... , , , Weighted linear aggregation is performed, a process that enables the dynamic fusion of different feature subspaces to obtain feature fusion results for multivariate load forecasting tasks. Feature fusion results Including the final characteristic tensor of the electricity load forecasting task The final characteristic tensor of the cooling load prediction task The final characteristic tensor of the heat load prediction task .
[0052] Among them, feature fusion results The expression is:
[0053]
[0054]
[0055] in, This represents the complete output of the feature extraction experts. This is for data splicing operations.
[0056] Step S5: For the feature fusion results of the MOE router, the HS-xLSTM framework is used for prediction. In the HS-xLSTM framework, the electrical load prediction task and the cold load prediction task are processed by the mLSTM layer with hard parameter sharing; the hot load prediction task is processed by the mLSTM layer with soft parameter sharing. Finally, each task is further processed by the sLSTM layer with soft parameter sharing for deep time series modeling, and the final prediction result is obtained by the fully connected layer.
[0057] The HS-xLSTM framework employs a hybrid sharing mechanism based on the correlation differences among load types: for strongly correlated electricity and cooling load forecasting tasks, hard parameter sharing is implemented in the mLSTM layer to force the extraction of common long-term trends; for significantly different heat load forecasting tasks, a separate mLSTM layer combined with soft parameter sharing constraints enables flexible knowledge modeling. All tasks then pass through an sLSTM layer constrained by soft sharing to refine short-order dynamic features, and finally, the prediction results are obtained via a fully connected layer.
[0058] In this embodiment, step S5 specifically includes:
[0059] Step S51, for the feature fusion result By performing linear projection through the task adaptation layer of each task, the feature dimensions are aligned to the hidden layer dimensions required by the HS-xLSTM framework.
[0060] Step S52: Within the HS-xLSTM framework, for both the electricity load forecasting task and the cooling load forecasting task, the final feature tensor of the electricity load forecasting task is... The final characteristic tensor of the cooling load prediction task By incorporating a hard parameter-shared mLSTM layer, the matrix memory mechanism of the hard parameter-shared mLSTM layer is used to mine the long-sequence time dependencies common to tasks.
[0061] Step S53: Within the HS-xLSTM framework, a soft-parameter-sharing mLSTM layer is used to generate the final feature tensor for the heat load prediction task. The process facilitates positive migration between related tasks through the shared layer, while effectively isolating tasks with significant differences to avoid negative migration interference.
[0062] In step S54, finally, the electrical load forecasting task, the cooling load forecasting task, and the heating load forecasting task are performed by deep time series modeling using sLSTM layers with soft parameter sharing, and the final prediction results are obtained by fully connected layers.
[0063] The following is a specific example illustrating this. The multivariate load dataset used is the IES electricity, cooling, and heating load data from Arizona State University, Tempe Campus, from 00:00 on April 1, 2020 to 23:00 on February 27, 2021. The data granularity is 1 hour, and the data length is 7992 records. Meteorological data was collected from the official website of the National Solar Radiation Database, including various climatic features such as temperature, wind speed, and relative humidity. Calendar data of months, days, and hours were also considered as input features. Considering the model's predictive performance under conditions of significant climatic variation, this application divides the original data into winter and summer portions for discussion. The summer dataset is from April to August 2020, with the last week of August used as the summer test set, the second-to-last week as the summer validation set, and the rest as the summer training set. Similarly, the winter dataset is from October 2020 to February 2021, with the last week of February 2021 used as the winter test set, the second-to-last week as the winter validation set, and the rest as the winter training set. To comprehensively evaluate the forecast results, root mean square error (RMSE), mean absolute error (MAE), and mean absolute error percentage (MAPE) were selected as evaluation indicators for each load.
[0064] Meanwhile, to fully reflect the coupling between data and give full play to the performance of the model, all three types of loads—electricity, cooling, and heating—were converted to kW at the beginning of the experiment. The conversion formula is as follows:
[0065] ;
[0066] To verify the effectiveness of this invention in the IES multivariate load prediction task, five existing time series prediction models, namely LSTM, XLSTM, Transformer, Autoformer, and CNN-BiLSTM-attention, were selected for comparison. The comparison results are shown in Table 1, focusing on three core prediction tasks: electrical load, cooling load, and heating load in winter and summer.
[0067] Table 1
[0068]
[0069] As shown in Table 1, this invention demonstrates significant performance advantages across all scenarios. In the summer scenario, the MAPE of this invention for electricity, cooling, and heating load prediction tasks are 2.201, 1.954, and 2.055, respectively, outperforming traditional models such as LSTM and Transformer. The MAPE for electricity, cooling, and heating loads is reduced by 39.75%, 28.84%, and 7.79% compared to the second-best performing Autoformer model, respectively. Further analysis of the winter scenario further highlights the advantages of this invention: in electricity, cooling, and heating load prediction, its MAPE is 2.504, 2.933, and 2.372, respectively, reducing the MAPE for electricity and cooling loads by 44.17%, 49.40%, and 29.86% compared to the second-best performing CNN-BiLSTM-Attention model. Considering all three indicators, this invention maintains stable and leading prediction performance in both summer and winter seasons and across all three load prediction tasks, fully validating its applicability and superiority in time series load prediction scenarios.
[0070] The comprehensive energy multi-variable load forecasting method based on the full-stage multi-task learning architecture described above has the following beneficial effects:
[0071] 1. This invention designs a multi-task feature extraction expert module. Within each expert network, feature attention, multi-scale causal dilation convolution, and temporal attention are chained together. This enables the multi-task feature extraction expert module to extract information simultaneously from two orthogonal dimensions: variable association and temporal dependency, greatly improving the depth and breadth of feature representation. Simultaneously, residual connections and causal masks are introduced to effectively alleviate the gradient vanishing problem in deep network training while strictly ensuring no data leakage occurs.
[0072] 2. This invention employs an MOE router trained based on load balancing loss. Unlike traditional static expert allocation, this invention uses a learnable router to dynamically filter and integrate features from shared experts and task-specific experts for each task. By introducing load balancing loss, it effectively avoids the problem of insufficient utilization of expert modules and ensures efficient and balanced utilization of model capacity.
[0073] 3. This invention establishes an HS-xLSTM framework to address the commonalities and individual characteristics of modeling different tasks. This framework employs structured parameter sharing for task relevance, allowing highly correlated tasks such as electrical load forecasting and cooling load forecasting to be processed using hard-parameter-shared mLSTM layers, which promotes knowledge transfer and generalization. For heat load forecasting, soft-parameter-shared mLSTM layers are used, which effectively isolates tasks with significant differences while promoting positive transfer between related tasks through shared layers, thus avoiding negative transfer interference. This hierarchical sharing strategy effectively alleviates task conflicts in multi-task learning while maintaining model efficiency, thereby improving the performance of multivariate load forecasting.
[0074] 4. This invention constructs a correlation analysis index that combines the mean absolute values of Pearson and Spearman correlation coefficients to assess the strength of the association. A consistency penalty factor is introduced using standard deviation to penalize differences in the results of different coefficients. This correlation analysis index not only quantifies the strength of the association between variables but also provides a more comprehensive and reliable assessment of the true association between variables, effectively avoiding the limitations of single analysis methods.
Claims
1. A comprehensive energy multi-element load forecasting method based on a full-stage multi-task learning architecture, characterized in that, include: Step S1: Collect multivariate load datasets. First, use moving average to find outliers and missing values due to statistical errors, and then use linear interpolation to fill them. Then, based on seasonal characteristics and data coupling relationships, divide the filled datasets into summer datasets and winter datasets. Step S2: Perform correlation analysis on the climate characteristics and loads in the summer and winter datasets respectively, calculate the correlation analysis index, and screen out the feature sequences that are strongly correlated with the electrical load, cooling load and heating load based on the correlation analysis index to form the screened dataset; Step S3: The filtered dataset is processed using a multi-task feature extraction expert module. The multi-task feature extraction expert module is used to extract feature information from two dimensions: variable association and time dependence, through feature attention, multi-scale causal dilation convolution and temporal attention, while introducing residual connections and causal masks. Step S4: Based on the processing results of the multi-task feature extraction expert module, a MOE router trained on load balancing loss is used for dynamic filtering and fusion. Step S5: For the feature fusion results of the MOE router, the HS-xLSTM framework is used for prediction. In the HS-xLSTM framework, the electrical load prediction task and the cold load prediction task are processed by the mLSTM layer with hard parameter sharing; the hot load prediction task is processed by the mLSTM layer with soft parameter sharing. Finally, each task is further processed by the sLSTM layer with soft parameter sharing for deep time series modeling, and the final prediction result is obtained by the fully connected layer. Specifically, step S4 includes: Step S41: The MOE router first receives the data after one-dimensional convolutional dimension expansion. And perform global average pooling along the time dimension to obtain the global context feature vector; Step S42: Input the global context feature vector into a router network composed of MLPs. Through nonlinear mapping, calculate the nonnormalized confidence score of the current input sample for each expert in the multi-task feature extraction expert module. The nonnormalized confidence score is used to reflect the degree of matching of each expert in processing the current sample. Step S43: The non-normalized confidence scores are transformed into a probability distribution using the Softmax function to determine the activation weights of all four experts, thereby constructing a sparse routing weight matrix. During the construction of the sparse routing weight matrix, a load balancing loss is introduced into the total loss function. This loss is calculated by... The average workload of each expert in the current batch To finally calculate the load balancing loss value. Load balancing loss value It will be added to the total training loss and used for backpropagation to guide the MOE router to explore and utilize the capabilities of all experts during training. Step S44: Based on the sparse routing weight matrix, use batch matrix multiplication to process the output of the task-shared feature extraction expert. Output of electrical load feature extraction expert Output of Cooling Load Feature Extraction Expert Output of heat load feature extraction expert Weighted linear aggregation is performed to obtain the feature fusion results for the multivariate load forecasting task. Feature fusion results Including the final characteristic tensor of the electricity load forecasting task The final characteristic tensor of the cooling load prediction task The final characteristic tensor of the heat load prediction task .
2. The comprehensive energy multi-dimensional load forecasting method based on a full-stage multi-task learning architecture according to claim 1, characterized in that, In step S2, the formula for calculating the correlation analysis index is: in, Representing variables and variables Correlation analysis indicators between them Representing variables and variables The basic correlation strength between them Represents the consistency penalty factor. For variables and variables The Pearson correlation coefficient between them For variables and variables The Spearman correlation coefficient between them.
3. The comprehensive energy multi-dimensional load forecasting method based on a full-stage multi-task learning architecture according to claim 1, characterized in that, Step S3 specifically includes: S31, For the filtered dataset, it is first subjected to a point-by-point linear transformation using a one-dimensional convolution with a kernel size of 1 to achieve feature dimension alignment and preliminary fusion, resulting in data after one-dimensional convolutional expansion. ; S32, data after dimensionality expansion via one-dimensional convolution. The data are respectively fed into the task-sharing feature extraction expert, electrical load feature extraction expert, cold load feature extraction expert, and hot load feature extraction expert in the multi-task feature extraction expert module. A self-attention mechanism is used to calculate the correlation matrix between feature dimensions. The query matrix, key matrix, and value matrix are then calculated. A Softmax activation function is applied to calculate the feature attention weight matrix. Finally, scaling and residual joins are used to obtain the feature attention-enhanced data. To dynamically adjust the weight distribution of different meteorological and load characteristics; S33, Data after feature attention enhancement The data is fed into a multi-scale dilated convolution module, which employs a multi-branch parallel structure, integrates convolution kernels of different sizes, and combines them with dilated convolution to obtain the multi-scale convolution output. The expression is; in, This represents the concatenation of multiple convolutional kernels of different scales. The kernel size is the convolutional kernel size. Indicates the kernel size as The dilated convolution kernel weights, yes At time step The input vector, For time steps, For the in-kernel indexing step, It is the expansion factor; S34, the output of multi-scale convolution by the temporal attention mechanism. The attention scores are calculated across different time steps, and finally, the output of the task-shared feature extraction expert is obtained using residual connections. Output of electrical load feature extraction expert Output of Cooling Load Feature Extraction Expert Output of heat load feature extraction expert .
4. The comprehensive energy multi-dimensional load forecasting method based on a full-stage multi-task learning architecture according to claim 3, characterized in that, In step S44, the feature fusion result The expression is: in, This represents the complete output of the feature extraction experts. For data concatenation operations, Indicates the first The first sample selection The probability of an expert.
5. The comprehensive energy multi-dimensional load forecasting method based on a full-stage multi-task learning architecture according to claim 4, characterized in that, In step S43, the load balancing loss value The expression is: in, For the first The average workload undertaken by each expert in the current batch This represents the average usage rate across all experts. The total number of samples.
6. The comprehensive energy multi-dimensional load forecasting method based on a full-stage multi-task learning architecture according to claim 5, characterized in that, Step S5 specifically includes: Step S51, for the feature fusion result By performing linear projection through the task adaptation layer of each task, the feature dimensions are aligned to the hidden layer dimensions required by the HS-xLSTM framework. Step S52: Within the HS-xLSTM framework, for both the electricity load forecasting task and the cooling load forecasting task, the final feature tensor of the electricity load forecasting task is... The final characteristic tensor of the cooling load prediction task By incorporating a hard parameter-shared mLSTM layer, the matrix memory mechanism of the hard parameter-shared mLSTM layer is used to mine the long-sequence time dependencies common to tasks. Step S53: Within the HS-xLSTM framework, a soft-parameter-sharing mLSTM layer is used to generate the final feature tensor for the heat load prediction task. Process it; In step S54, finally, the electrical load forecasting task, the cooling load forecasting task, and the heating load forecasting task are performed by deep time series modeling using sLSTM layers with soft parameter sharing, and the final prediction results are obtained by fully connected layers.