A Multivariate Load Forecasting Method and System Based on Multi-Task Learning

Abstract

This invention relates to a multi-task learning-based method and system for multi-variable load forecasting. The method includes the following steps: collecting historical load data, meteorological data, and date data and integrating them into comprehensive data; preprocessing the comprehensive data to obtain a multi-dimensional input tensor; inputting the multi-dimensional input tensor into a pre-trained load forecasting model for joint forecasting to obtain load forecasting results. The load forecasting model is a CNN-GRU-MTL deep learning model, including an input layer, a CNN layer, a GRU layer, and an output layer connected in sequence. An MTL hard-shared framework is set on the CNN and GRU layers. The CNN layer is used to extract local spatiotemporal features of the comprehensive data, the GRU layer is used to perform time-series modeling of the local spatiotemporal features, and the MTL hard-shared framework is used to achieve feature sharing of load data on the CNN and GRU layers, enabling simultaneous forecasting of multi-variable load data. Compared with existing technologies, this invention significantly improves the forecasting accuracy for multiple loads.

Description

Technical Field

[0001] This invention relates to the field of energy management and intelligent load forecasting technology, and in particular to a multi-task learning-based multivariate load forecasting method and system. Background Technology

[0002] With the optimization and adjustment of the global energy structure and the continuous advancement of environmental governance policies, integrated energy systems, due to their advantages such as multi-energy complementarity and efficient synergy, are gradually becoming an important development direction in the energy sector. Through the coordinated scheduling and optimized management of multiple energy sources such as electricity, heat, and cooling, integrated energy systems can achieve efficient energy transmission, storage, and utilization, improving the overall operating efficiency and economy of the system.

[0003] Load forecasting, as a fundamental supporting element in the operation of integrated energy systems, is of great significance for ensuring stable system operation, optimizing energy dispatch, and improving energy utilization efficiency. Currently, research on load forecasting for integrated energy systems mainly focuses on independent forecasting of single load types. Typical methods include time series analysis, traditional statistical regression models, and machine learning-based forecasting methods such as Support Vector Machines (SVM), Random Forests (RF), Convolutional Neural Networks (CNN), Long Short-Term Memory Networks (LSTM), and Gated Recurrent Units (GRU). While these methods have achieved certain results in single load forecasting, they suffer from insufficient forecasting accuracy when dealing with the strong coupling and complex dynamic characteristics of multiple loads (such as electricity, cooling, and heating).

[0004] Existing technologies attempt to improve prediction performance by introducing fusion models or optimization algorithms. For example, the grasshopper optimization algorithm is used to optimize SVM parameters, or fuzzy clustering combined with random forests is used for data grouping and prediction. Some studies also use CNN-LSTM structures to extract load features and perform time series modeling. However, when faced with the coupling characteristics of multiple types of loads, the above methods still have problems such as complex modeling, insufficient feature extraction, or insufficient utilization of mutual information between tasks.

[0005] To address the aforementioned issues, multi-task learning has been introduced into the field of multivariate load forecasting. This method, by introducing a shared layer mechanism, allows different tasks to share model parameters, effectively mining the correlations between various load types, thereby improving the model's prediction accuracy and generalization ability. However, current multivariate load forecasting methods still face the following challenges in application: On the one hand, the imbalance between tasks affects the stability of model training. For example, patent application CN111950793A discloses a comprehensive energy system load forecasting method considering the coupling characteristics of multiple loads. This method uses an adaptive local iterative filtering decomposition method to decompose historical cold, heat, and electricity load time series, and reconstructs the periodic, fluctuation, and trend sequences of each load. Although this method has certain effects in load decomposition modeling and error correction, the overall modeling process is relatively complex, involving multiple rounds of sequence decomposition, reconstruction, and multi-stage modeling, resulting in a long model training and deployment process. Furthermore, the depth and breadth of shared features between tasks are limited, making it susceptible to the decomposition quality, thus affecting the overall prediction accuracy and generalization ability. On the other hand, the configuration of model hyperparameters often relies on human experience and lacks an adaptive adjustment mechanism, resulting in poor adaptability of the model in different application scenarios and difficulty in ensuring the consistency and robustness of prediction results. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art by providing a multi-task learning-based method and system for predicting multiple loads, which significantly improves the prediction accuracy for multiple loads.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] A multivariate load forecasting method based on multi-task learning includes the following steps:

[0009] S1. Collect historical load data, meteorological data, and date data; integrate the historical load data with the meteorological data and date data into comprehensive data; and perform data preprocessing on the comprehensive data to obtain a multidimensional input tensor.

[0010] S2. The multidimensional input tensor is input into the pre-trained load prediction model for joint prediction to obtain the load prediction result. The load prediction model is a CNN-GRU-MTL deep learning model, which includes an input layer, a CNN layer, a GRU layer and an output layer connected in sequence. The CNN layer and the GRU layer are equipped with an MTL hard-shared framework. The CNN layer is used to extract the local spatiotemporal features of the comprehensive data, the GRU layer is used to perform time-series modeling of the local spatiotemporal features, and the MTL hard-shared framework is used to realize the sharing of load data features on the CNN layer and the GRU layer, and to simultaneously predict multidimensional load data.

[0011] Furthermore, the load history data includes various load history data from cooling load, heating load, and electrical load.

[0012] Furthermore, the meteorological data includes multiple parameters such as air temperature, relative humidity, atmospheric pressure, wind speed, and perceived temperature.

[0013] Furthermore, the date data includes year, month, day, weekday, holiday identifier, and workday identifier.

[0014] Furthermore, the specific steps of the data preprocessing include:

[0015] Outlier analysis is performed on the comprehensive data. Outliers in the data are identified using a box plot method based on a monthly sliding window, and the outliers are replaced using the median substitution method.

[0016] For missing values ​​in the synthesis, interpolation of nearby time values ​​is used to fill in the missing values ​​to maintain the integrity and continuity of the synthesized data;

[0017] The comprehensive data, after outlier analysis and missing value imputation, is normalized using the maximum-minimum standardization method to obtain a multidimensional input tensor.

[0018] Furthermore, during the training process of the load forecasting model, a Bayesian optimization algorithm is used to adaptively adjust the hyperparameters of the load forecasting model. Specific steps include:

[0019] S201. Using the prediction error of the load forecasting model on the validation set as the objective function, construct a probabilistic surrogate model based on a Gaussian process.

[0020] S202. Simulate the distribution characteristics of the prediction error in the hyperparameter space using a probabilistic surrogate model;

[0021] S203. Based on the distribution characteristics, determine the next hyperparameter combination to be evaluated using the expected improvement acquisition function;

[0022] S204. Train the load prediction model under the hyperparameter combination to be evaluated, record the performance indicators, and update the probabilistic surrogate model.

[0023] S205. Iterate through steps S202 to S204 until the preset maximum number of iterations is met or the expected improvement in the acquisition function converges, and obtain the hyperparameter combination with the minimum prediction error.

[0024] Furthermore, the hyperparameters include: the kernel size and number of convolutional kernels in the CNN layer, the number of neurons in the GRU layer, the learning rate of the model, and the training batch size.

[0025] Furthermore, the joint loss function of the load forecasting model is:

[0026]

[0027]

[0028]

[0029] In the formula, For the joint loss function, These are all the trainable parameters for the load forecasting model. For the number of tasks ( (including cooling load, heating load and electrical load). For the first The weighting coefficients for each task. For the sample size, For the first The first task The predicted value for each sample, For the first The first task The true value of each sample These are the weighting coefficients of the relevance regularization term. For the task and tasks The dynamic correlation coefficient between them For the first extracted by the MTL hard shared framework Feature embedding representation of each task For the first extracted by the MTL hard shared framework Feature embedding representation of each task for Regularization term weights, For the Sigmoid function, For activation function, , , and For trainable parameters, The cosine similarity is between task features.

[0030] Furthermore, the prediction performance of the load forecasting model is analyzed using evaluation indicators, including root mean square error and mean absolute percentage error.

[0031] According to another aspect of the present invention, a multi-task learning-based multivariate load forecasting system is provided, comprising:

[0032] The multidimensional input tensor acquisition module is used to collect historical load data, meteorological data, and date data, integrate the historical load data with the meteorological data and date data into comprehensive data, and perform data preprocessing on the comprehensive data to obtain multidimensional input tensors;

[0033] The load prediction module is used to jointly predict the load prediction results by inputting the multidimensional input tensor into the pre-trained load prediction model. The load prediction model is a CNN-GRU-MTL deep learning model, which includes an input layer, a CNN layer, a GRU layer and an output layer connected in sequence. The CNN layer and the GRU layer are equipped with an MTL hard-shared framework. The CNN layer is used to extract the local spatiotemporal features of the comprehensive data, and the GRU layer is used to perform temporal modeling on the local spatiotemporal features. The MTL hard-shared framework is used to realize the sharing of load data features on the CNN layer and the GRU layer, and to simultaneously predict multi-dimensional load data.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] 1. This invention constructs a load forecasting model based on a combination of CNN, GRU, and MTL, introduces a multi-task learning mechanism, achieves unified feature extraction by sharing a CNN-GRU network, and sets multiple task branches in the output layer, thereby taking into account the coupling and differences of different load forecasting tasks. This enables joint modeling and forecasting of cooling load, heating load, and electrical load in integrated energy stations, effectively utilizes the synergistic change patterns among loads, improves generalization ability, reduces the total number of model parameters, improves computational efficiency, avoids information silos between tasks, enhances learning robustness, and significantly improves the forecasting accuracy for multiple loads.

[0036] 2. This invention uses a Bayesian optimization algorithm to adaptively adjust multiple key hyperparameters of the constructed load forecasting model. By improving the expected value acquisition function, it comprehensively considers the current optimal solution and potential improvement space, dynamically balances the exploration and utilization in the parameter space, and selects the hyperparameter combination most likely to improve model performance in each iteration. This avoids the subjectivity and uncertainty brought about by manual experience-based parameter tuning, and significantly improves the stability, adaptability and overall load forecasting accuracy of the forecasting model. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating a multi-task learning-based multivariate load forecasting method proposed in this invention.

[0038] Figure 2 This is a schematic diagram of the load forecasting model. Detailed Implementation

[0039] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0040] The following English abbreviations are involved:

[0041] Convolutional Neural Network (CNN)

[0042] Gated Recurrent Unit (GRU)

[0043] Multi-Task Learning (MTL)

[0044] Root Mean Square Error (RMSE)

[0045] Mean Absolute Percentage Error (MAPE)

[0046] Example 1

[0047] This embodiment provides a multivariate load forecasting method based on multi-task learning, such as... Figure 1 As shown, it includes the following steps:

[0048] S1. Collect historical load data, meteorological data, and date data. Integrate the historical load data with the meteorological and date data into comprehensive data. Perform data preprocessing on the comprehensive data to obtain a multidimensional input tensor.

[0049] Historical operating data for cooling, heating, and electrical loads are obtained from the integrated energy station's monitoring system, historical data platform, or relevant databases. This data is sampled at one-hour intervals, covering multiple energy supply seasons to ensure the model's broad applicability and generalization ability. In addition to load data, closely related meteorological data, including but not limited to temperature, relative humidity, atmospheric pressure, wind speed, and perceived temperature, must also be collected simultaneously. This data can be obtained from meteorological stations or third-party meteorological service platforms. Date information, including year, month, day, weekday, holiday markers, and workday markers, is also collected to capture the impact of periodic and human-induced factors on load variations.

[0050] The collected data is first cleaned. The specific steps of data preprocessing include:

[0051] Outlier analysis was performed on the comprehensive data. A box plot method based on a monthly sliding window was used to identify outliers in the data, and the median substitution method was used to replace the outliers. The specific steps are as follows:

[0052] Divide the data window into monthly units. Within each time window, calculate the first and third quartiles of the load data, and then calculate the interquartile range. The interquartile range is:

[0053]

[0054] In the formula, Interquartile range, It is the third quartile. It is the first quartile.

[0055] Based on the principle of box plot method, upper and lower limits are set for outlier judgment, with the upper limit being:

[0056]

[0057] The lower limit is:

[0058]

[0059] In the formula, The upper limit, It is the third quartile. Interquartile range, As the lower limit, It is the first quartile.

[0060] Any data point falling outside the lower and upper limits is identified as an outlier. For identified outliers, a median substitution method is used to correct them, that is, the median of the data in that month is used to replace the outlier, in order to maintain the stability and representativeness of the data distribution to the greatest extent.

[0061] To address missing values ​​in the synthesis, interpolation of nearby time values ​​is used to fill in the missing values, thus maintaining the integrity and continuity of the synthesized data.

[0062] The comprehensive data after outlier analysis and missing value imputation is normalized using the maximum-minimum standardization method, which linearly maps it to the [0,1] interval to ensure that each feature value has a consistent dimensional scale.

[0063] Through the above operations, a set of cleaned, completed and normalized multidimensional input tensors is finally obtained, providing a reliable input data foundation for the load forecasting model.

[0064] S2. Input the multidimensional input tensor into the pre-trained load prediction model for joint prediction to obtain the load prediction result.

[0065] A load forecasting model based on a combination of Convolutional Neural Networks (CNN), Gated Recurrent Units (GRU), and Multi-Task Learning (MTL) is constructed. The model structure incorporates a multi-task learning mechanism, achieving unified feature extraction through a shared CNN-GRU network. Multiple task branches are set in the output layer to balance the coupling and differences between various load forecasting tasks, enabling joint modeling and forecasting of cooling, heating, and electrical loads in integrated energy stations. The load forecasting model is a CNN-GRU-MTL deep learning model, as shown below. Figure 2 As shown, it includes an input layer, a CNN layer, a GRU layer, and an output layer connected in sequence. The CNN layer and the GRU layer are equipped with an MTL hard-shared framework. The CNN layer is used to extract local spatiotemporal features of the comprehensive data, the GRU layer is used to perform time series modeling of the local spatiotemporal features, and the MTL hard-shared framework is used to realize the sharing of load data features on the CNN layer and the GRU layer, and to perform simultaneous prediction of multivariate load data.

[0066] The input layer of the load forecasting model receives normalized historical load data and its corresponding meteorological characteristics (including temperature, humidity, air pressure, etc.) and date characteristics (including holiday and workday identifiers, etc.), and transforms them into a multidimensional input tensor in a unified format.

[0067] Subsequently, the input tensor is fed into a shared convolutional neural network (CNN) layer. The CNN layer consists of two one-dimensional convolutional layers. By setting appropriate kernel size and stride, local features of the load data are extracted on the time axis, which can effectively identify short-term dynamic features such as periodic fluctuations and abrupt changes. All tasks share the parameters of the CNN layer, enabling the three load tasks of cold, heat, and electricity to jointly learn representative low-level features, improve the utilization of model parameters, and enhance the coupling and expressive ability between tasks.

[0068] After convolution, the output features are fed into a shared GRU layer. The GRU layer has a recurrent structure and a gating mechanism, which can capture the dynamic dependencies of the input data over long time dimensions. Compared with the traditional RNN structure, the GRU layer significantly reduces the number of training parameters while maintaining the ability to remember time, thereby reducing the risk of overfitting and improving training efficiency. By sharing GRU parameters among multiple tasks, each task can jointly perceive the global load change trend, further enhancing the ability to model features at multiple time scales.

[0069] Following the shared CNN and GRU layers, the model processes cold load prediction, heat load prediction, and electrical load prediction as parallel subtasks within its branching structure. Each subtask is equipped with an independent fully connected output layer to extract specific information relevant to that load type, enabling differentiated modeling. The multi-task learning strategy employs a hard-sharing mechanism, where multiple tasks share the front-end model structure (CNN-GRU) and each constructs an independent task output channel, establishing an effective collaborative mechanism between shared and independent feature representations. While sharing parameters, the model retains the personalized modeling capabilities for each load type, enabling the sharing of feature representations among electrical, heat, and cold loads, thereby improving prediction accuracy.

[0070] The entire model employs a joint loss function for end-to-end training. The loss function weights and sums the prediction errors of each subtask, achieving balance and mutual assistance among multiple tasks. This design not only improves the model's prediction accuracy across various load types but also fully explores the inherent correlations between loads, enhancing the model's generalization ability and robustness under different load conditions. The joint loss function of the load forecasting model is:

[0071]

[0072]

[0073]

[0074] In the formula, For the joint loss function, These are all the trainable parameters for the load forecasting model. For the number of tasks ( (including cooling load, heating load and electrical load). For the first The weighting coefficients for each task. For the sample size, For the first The first task The predicted value for each sample, For the first The first task The true value of each sample These are the weighting coefficients of the relevance regularization term. For the task and tasks The dynamic correlation coefficient between them For the first extracted by the MTL hard shared framework Feature embedding representation of each task For the first extracted by the MTL hard shared framework Feature embedding representation of each task for Regularization term weights, For the Sigmoid function, For activation function, , , and For trainable parameters, The cosine similarity is between task features.

[0075] During the training process of the load forecasting model, the hyperparameters of the load forecasting model are adaptively adjusted using the Bayesian optimization algorithm. The specific steps include:

[0076] S201. Using the prediction error of the load forecasting model on the validation set as the objective function, construct a probabilistic surrogate model based on a Gaussian process.

[0077] S202. Simulate the distribution characteristics of prediction error in hyperparameter space using a probabilistic surrogate model;

[0078] S203. Based on the distribution characteristics, determine the next hyperparameter combination to be evaluated through the expected improvement acquisition function;

[0079] S204. Train the load prediction model under the hyperparameter combination to be evaluated, record the performance indicators, and update the probabilistic surrogate model.

[0080] S205. Iterate through steps S202 to S204 until the preset maximum number of iterations is met or the expected improvement in the acquisition function converges, and obtain the hyperparameter combination with the minimum prediction error.

[0081] By improving the acquisition function and comprehensively considering the current optimal solution and potential improvement space, the exploration and utilization in the parameter space are dynamically balanced. In each iteration, the hyperparameter combination most likely to improve model performance is selected. Hyperparameters include: kernel size and number of convolutional kernels in CNN layers, number of neurons in GRU layers, model learning rate, and training batch size. The hyperparameter combination most likely to improve model performance selected in each iteration is trained and validated on a real model. This process continuously updates the surrogate model, gradually bringing it closer to the global optimum, and obtaining the hyperparameter configuration that optimizes the model's predictive performance. Finally, the obtained optimal hyperparameter combination is applied to the training phase of the final model, avoiding the subjectivity and uncertainty brought about by manual experience-based parameter tuning, and significantly improving the stability, adaptability, and overall load prediction accuracy of the prediction model.

[0082] By combining feature extraction, temporal modeling, and multi-task learning, this structure can achieve the following advantages over traditional single-task networks:

[0083] (1) Effectively utilize the synergistic change patterns among loads to improve generalization ability;

[0084] (2) Reduce the total number of model parameters and improve computational efficiency;

[0085] (3) Avoid information silos between tasks and enhance learning robustness;

[0086] (4) Experimental results have shown that it can significantly improve the accuracy of various load predictions.

[0087] After completing the model structure construction and hyperparameter optimization, the prediction performance of the load forecasting model was analyzed using evaluation indicators, including root mean square error and mean absolute percentage error. To verify the effectiveness of the multivariate load forecasting method based on multi-task learning proposed in this invention, operational data from a comprehensive energy station was selected as experimental data. Data types included historical hourly cooling, heating, and electrical load data, and meteorological data (including temperature, humidity, and air pressure) and date information (such as holiday and workday markers) were collected simultaneously. The data was divided into a training set (70%) and a test set (30%) in chronological order.

[0088] During model training, input features include load data from the past 48 hours and external weather and date information. The output is the predicted load values ​​for electricity, heat, and cold at the next moment. The model training uses the Adam optimizer with a batch size of 200 and a maximum training epoch of 256, and includes an early stopping mechanism to prevent overfitting. The optimal hyperparameter combination obtained through Bayesian optimization is as follows: kernel size, number of kernels, number of GRU neurons in the first layer, number of GRU neurons in the second layer, and learning rate are 3, 46, 45, 40, and 0.00735, respectively.

[0089] To verify the predictive advantages of the multi-task learning-based multivariate load prediction method proposed in this invention in multi-task scenarios, this experiment sets up two prediction models, CNN-GRU and CNN-LSTM-MTL, for comparison. The evaluation metrics of each model on the test set are shown in Table 1.

[0090] Table 1. Prediction Evaluation Indicators of the Model of this Invention and Other Models

[0091]

[0092] As shown in Table 1, the fitting curves of the CNN-GRU-MTL model based on Bayesian optimization in this invention are closer to the true values ​​in the prediction of cold, heat, and electricity loads, demonstrating the best fitting effect. Compared with other prediction models, this model can more accurately capture the changing trends of cold, heat, and electricity loads in different time periods, reducing prediction errors. Compared with the CNN-GRU single-task model, the model in this paper reduces the mean absolute percentage error (MAPE) by an average of 26.94% and the root mean square error (RMSE) by an average of 24.74%. This indicates that the proposed multi-task learning framework can not only achieve information sharing among loads but also effectively suppress overfitting and improve the model's generalization ability. Furthermore, after adjusting the hyperparameters using Bayesian optimization, the model achieves optimal performance on both the validation and test sets, indicating that the optimization method plays a key role in improving model performance.

[0093] The multivariate load forecasting method based on multi-task learning proposed in this invention not only has high forecasting accuracy, but also good model stability and generalization ability, and can be widely applied to load forecasting and intelligent scheduling of integrated energy systems.

[0094] Example 2

[0095] This embodiment provides a multi-task learning-based multivariate load forecasting system, including:

[0096] The multidimensional input tensor acquisition module is used to collect historical load data, meteorological data, and date data, integrate the historical load data with the meteorological data and date data into comprehensive data, and perform data preprocessing on the comprehensive data to obtain multidimensional input tensors;

[0097] The load forecasting module is used to input multi-dimensional input tensors into a pre-trained load forecasting model for joint forecasting to obtain load forecasting results. The load forecasting model is a CNN-GRU-MTL deep learning model, which includes an input layer, a CNN layer, a GRU layer, and an output layer connected in sequence. The CNN layer and the GRU layer are equipped with an MTL hard-shared framework. The CNN layer is used to extract local spatiotemporal features of the comprehensive data, and the GRU layer is used to perform time series modeling on the local spatiotemporal features. The MTL hard-shared framework is used to realize the sharing of load data features on the CNN layer and the GRU layer, and to simultaneously forecast multi-dimensional load data.

[0098] The rest is the same as in Example 1.

[0099] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A multivariate load forecasting method based on multi-task learning, characterized in that, Includes the following steps: S1. Collect historical load data, meteorological data, and date data; integrate the historical load data with the meteorological data and date data into comprehensive data; and perform data preprocessing on the comprehensive data to obtain a multidimensional input tensor. S2. The multidimensional input tensor is input into the pre-trained load prediction model for joint prediction to obtain the load prediction result. The load prediction model is a CNN-GRU-MTL deep learning model, including an input layer, a CNN layer, a GRU layer, and an output layer connected in sequence. An MTL hard-sharing framework is set on the CNN layer and the GRU layer. The CNN layer is used to extract local spatiotemporal features of the comprehensive data, and the GRU layer is used to perform temporal modeling on the local spatiotemporal features. The MTL hard-sharing framework is used to achieve feature sharing of load data on the CNN layer and the GRU layer, enabling simultaneous prediction of multidimensional load data. The joint loss function of the load prediction model is: In the formula, For the joint loss function, These are all the trainable parameters for the load forecasting model. For the number of tasks, It includes cooling load, heating load and electrical load. For the first The weighting coefficients for each task. For the sample size, For the first The first task The predicted value for each sample, For the first The first task The true value of each sample These are the weighting coefficients of the relevance regularization term. For the task and tasks The dynamic correlation coefficient between them For the first extracted by the MTL hard shared framework Feature embedding representation of each task For the first extracted by the MTL hard shared framework Feature embedding representation of each task for Regularization term weights, For the Sigmoid function, For activation function, , , and For trainable parameters, The cosine similarity is between task features.

2. The multi-task learning-based multivariate load forecasting method according to claim 1, characterized in that, The load history data includes various load history data from cooling load, heating load, and electrical load.

3. The multi-task learning-based multivariate load forecasting method according to claim 1, characterized in that, The meteorological data includes multiple parameters such as air temperature, relative humidity, atmospheric pressure, wind speed, and perceived temperature.

4. The multi-task learning-based multivariate load forecasting method according to claim 1, characterized in that, The date data includes year, month, day, weekday, holiday identifier, and workday identifier.

5. The multi-task learning-based multivariate load forecasting method according to claim 1, characterized in that, The specific steps of the data preprocessing include: Outlier analysis is performed on the comprehensive data. Outliers in the data are identified using a box plot method based on a monthly sliding window, and the outliers are replaced using the median substitution method. For missing values ​​in the comprehensive data, interpolation of values ​​from nearby time points is used to fill in the missing values, so as to maintain the integrity and continuity of the comprehensive data. The comprehensive data, after outlier analysis and missing value imputation, is normalized using the maximum-minimum standardization method to obtain a multidimensional input tensor.

6. The multi-task learning-based multivariate load forecasting method according to claim 1, characterized in that, During the training process of the load forecasting model, the hyperparameters of the load forecasting model are adaptively adjusted using a Bayesian optimization algorithm. Specific steps include: S201. Using the prediction error of the load forecasting model on the validation set as the objective function, construct a probabilistic surrogate model based on a Gaussian process. S202. Simulate the distribution characteristics of the prediction error in the hyperparameter space using a probabilistic surrogate model; S203. Based on the distribution characteristics, determine the next hyperparameter combination to be evaluated using the expected improvement acquisition function; S204. Train the load prediction model under the hyperparameter combination to be evaluated, record the performance indicators, and update the probabilistic surrogate model. S205. Iterate through steps S202 to S204 until the preset maximum number of iterations is met or the expected improvement in the acquisition function converges, and obtain the hyperparameter combination with the minimum prediction error.

7. The multi-task learning-based multivariate load forecasting method according to claim 6, characterized in that, The hyperparameters include: kernel size and number of kernels in the CNN layer, number of neurons in the GRU layer, learning rate of the model, and training batch size.

8. The multi-task learning-based multivariate load forecasting method according to claim 1, characterized in that, The predictive performance of the model is analyzed using evaluation metrics, including root mean square error and mean absolute percentage error.

9. A multi-task learning-based multivariate load forecasting system, characterized in that, include: The multidimensional input tensor acquisition module is used to collect historical load data, meteorological data, and date data, integrate the historical load data with the meteorological data and date data into comprehensive data, and perform data preprocessing on the comprehensive data to obtain multidimensional input tensors; The load forecasting module is used to jointly forecast the load forecasting model by inputting the multidimensional input tensor into the pre-trained load forecasting model to obtain the load forecasting result. The load forecasting model is a CNN-GRU-MTL deep learning model, including an input layer, a CNN layer, a GRU layer, and an output layer connected in sequence. An MTL hard-sharing framework is set on the CNN layer and the GRU layer. The CNN layer is used to extract local spatiotemporal features of the comprehensive data, and the GRU layer is used to perform temporal modeling on the local spatiotemporal features. The MTL hard-sharing framework is used to achieve feature sharing of load data on the CNN layer and the GRU layer, enabling simultaneous forecasting of multidimensional load data. The joint loss function of the load forecasting model is: In the formula, For the joint loss function, These are all the trainable parameters for the load forecasting model. For the number of tasks, It includes cooling load, heating load and electrical load. For the first The weighting coefficients for each task. For the sample size, For the first The first task The predicted value for each sample, For the first The first task The true value of each sample These are the weighting coefficients of the relevance regularization term. For the task and tasks The dynamic correlation coefficient between them For the first extracted by the MTL hard shared framework Feature embedding representation of each task For the first extracted by the MTL hard shared framework Feature embedding representation of each task for Regularization term weights, For the Sigmoid function, For activation function, , , and For trainable parameters, The cosine similarity is between task features.

Images