Regional short-term load integrated prediction method and device based on multi-dimensional fusion features and model lightening, equipment and medium

Abstract

The present application belongs to the technical field of power system load prediction, and discloses a regional short-term load integrated prediction method and device based on multi-dimensional fusion features and model lightening, equipment and medium; the method comprises: obtaining the original regional load sequence, meteorological measured data, meteorological forecast data, week data and holiday data of the region to be predicted, decomposing the original regional load sequence to obtain a plurality of regional load subsequences and a residual term; constructing multi-dimensional fusion features according to the plurality of regional load subsequences, the residual term, the meteorological measured data, the meteorological forecast data, the week data and the holiday data, inputting the multi-dimensional fusion features into a pre-trained lightened regional short-term load integrated prediction model to obtain a future regional short-term load prediction result. The present application can finely characterize and learn the periodicity and trend hidden in the load sequence, and can further improve the prediction accuracy compared with the existing prediction method.

Description

Technical Field

[0001] This invention belongs to the field of power system load forecasting technology, specifically relating to a regional short-term load integrated forecasting method, device, equipment, and medium based on multi-dimensional fusion features and lightweight model. Background Technology

[0002] To adapt to the needs of production and daily life in the new era, the power system is gradually developing towards a robust, self-healing, and highly efficient smart grid with large-scale renewable energy integration. Meanwhile, regional short-term load forecasting plays a crucial role as a key component of the smart grid's refined management, optimized dispatch decisions, and safe and stable operation. However, due to the strong volatility and randomness of regional loads caused by multiple coupled factors such as weather, weekday patterns, and holiday types, traditional statistical analysis-based regional short-term load forecasting methods struggle to fully exploit the inherent nonlinear relationships within regional load data, resulting in low forecast accuracy. Therefore, compared to traditional statistical analysis-based forecasting methods, artificial intelligence-based forecasting methods (especially deep learning) possess powerful representation learning and nonlinear mapping capabilities, enabling them to uncover hidden patterns in data distribution. Furthermore, their higher model complexity allows them to better represent massive amounts of data information, thus gradually becoming the main solution for improving the accuracy of regional short-term load forecasting.

[0003] However, existing AI-based regional short-term load forecasting methods often directly use load sequences to construct time-domain features as model input features, which fails to accurately represent and learn the periodicity and trends implicit in the load sequences, making it difficult to further improve forecast accuracy.

[0004] Furthermore, existing AI-based regional short-term load forecasting methods typically employ extremely complex deep neural network models, which require significant storage and computing resources for training and inference on cloud servers. Consequently, they are difficult to deploy and apply widely on local devices with limited storage and computing resources. Summary of the Invention

[0005] The purpose of this invention is to provide a method, apparatus, device, and medium for regional short-term load integrated forecasting based on multi-dimensional fusion features and lightweight models, in order to solve the technical problem that existing artificial intelligence-based regional short-term load forecasting methods cannot accurately characterize and learn the periodicity and trend implicit in the load sequence, thus making it difficult to further improve the forecasting accuracy.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a regional short-term load integrated forecasting method based on multi-dimensional fusion features and lightweight model, comprising: The original regional load sequence, measured meteorological data, meteorological forecast data, weekday data, and holiday data of the area to be predicted are obtained. The original regional load sequence is then decomposed to obtain multiple regional load subsequences and remainders. Based on the multiple regional load subsequences, residual items, meteorological measured data, meteorological forecast data, weekday data and holiday data, a multi-dimensional fusion feature is constructed. The multi-dimensional fusion feature is then input into a pre-trained lightweight regional short-term load integrated prediction model to obtain the future regional short-term load prediction results. The lightweight regional short-term load integrated prediction model is trained in the following way: Historical load sequences and corresponding meteorological measured data, meteorological forecast data, weekday data, and holiday data for the sample area are obtained. The historical load sequences are decomposed into signals and multidimensional fusion features are constructed for training. A deep neural network combined with ensemble learning is used to train an initial regional short-term load ensemble prediction model. The model parameters are jointly optimized based on the prediction errors of the original load sequence, regional load subsequence, and residual terms. The initial regional short-term load ensemble prediction model that meets the prediction accuracy requirements is pruned based on weight importance. The initial regional short-term load ensemble prediction model is used as the teacher network and the pruned model is used as the student network for knowledge distillation fine-tuning to obtain the lightweight regional short-term load ensemble prediction model.

[0007] A further improvement of the present invention is that: in the step of performing signal decomposition on the original regional load sequence to obtain multiple regional load sub-sequences and a remainder term, the signal decomposition adopts empirical mode decomposition or variational mode decomposition to decompose the original regional load sequence into N regional load sub-sequences and a remainder term.

[0008] A further improvement of the present invention is that the multidimensional fusion features include multiple regional load subsequences, remainders, meteorological measured data features, meteorological forecast data features, weekday type features, and holiday type features.

[0009] A further improvement of the present invention is that both the measured meteorological data and the forecast meteorological data include air pressure, air temperature, relative humidity, hourly precipitation, and maximum wind speed.

[0010] A further improvement of the present invention is that the initial regional short-term load integrated prediction model includes a fusion feature extraction module, a load component feature extraction module, an intermediate feature aggregation layer, a regional load subsequence prediction layer, a remainder prediction layer, a regional load subsequence and remainder prediction result aggregation layer, and an original regional load sequence prediction layer.

[0011] A further improvement of the present invention is that: the fusion feature extraction module is used to extract the deep implicit features of the multidimensional fusion features used for training using a fully connected deep neural network; and the load component feature extraction module is used to extract the deep implicit features of the load subsequences and remainders of each region sequentially using a convolutional neural network and a long short-term memory neural network.

[0012] A further improvement of this invention is that the step of training the initial regional short-term load ensemble prediction model using a deep neural network combined with ensemble learning specifically includes: The deep implicit features output by the fusion feature extraction module are input into the intermediate feature aggregation layer along with the deep implicit features of the load subsequences of each region and the deep implicit features of the remainder, respectively, to obtain the aggregated features corresponding to the load subsequences of each region and the aggregated features corresponding to the remainder. The aggregated features corresponding to the load subsequences of each region are input into the regional load subsequence prediction layer to obtain the prediction results of each regional load subsequence; the aggregated features corresponding to the remainder are input into the remainder prediction layer to obtain the remainder prediction results; the prediction results of each regional load subsequence and the remainder prediction results are input into the prediction result aggregation layer to obtain the aggregated prediction results, and the regional load prediction results are obtained based on the aggregated prediction results.

[0013] A further improvement of the present invention is that the step of jointly optimizing the model parameters based on the prediction errors of the original load sequence, the regional load subsequence, and the remainder specifically includes: constructing a joint loss function based on the prediction errors corresponding to the regional load prediction results, the prediction errors corresponding to the prediction results of each regional load subsequence, and the prediction errors corresponding to the prediction results of the remainder, and updating the model parameters of the initial regional short-term load integrated prediction model based on the joint loss function.

[0014] A further improvement of the present invention is that the joint loss function is a weighted combination of the root mean square error of the regional load forecast result, the root mean square error of the forecast results of each regional load subsequence, and the root mean square error of the forecast results of the remainder.

[0015] A further improvement of this invention is that the step of performing weight-based pruning on the initial regional short-term load integrated prediction model that meets the prediction accuracy requirements specifically includes: The weight tensors of each layer of each network module in the initial regional short-term load integrated prediction model are sorted in ascending order according to the absolute value of the weights and expanded into a one-dimensional weight vector. The importance score of each weight in the one-dimensional weight vector is calculated. The corresponding weights are set to zero according to the importance score from small to large until the preset sparsity of each layer is met.

[0016] A further improvement of this invention is that the step of performing knowledge distillation fine-tuning using the initial regional short-term load ensemble prediction model as the teacher network and the pruned model as the student network to obtain the lightweight regional short-term load ensemble prediction model specifically includes: Keeping the model parameters of the initial short-term load ensemble prediction model of the teacher network unchanged, the parameters of the pruned model of the student network are updated based on the training set so that the student network learns the prediction output and intermediate feature representation of the teacher network; the knowledge distillation fine-tuning constructs a total distillation loss based on feature distillation loss and logistic distillation loss, and optimizes and updates the model parameters of the student network according to the total distillation loss.

[0017] Secondly, the present invention provides a regional short-term load integrated forecasting device based on multi-dimensional fusion features and lightweight model, comprising: The data acquisition module is used to acquire the original regional load sequence, measured meteorological data, meteorological forecast data, weekday data and holiday data of the area to be predicted, and to perform signal decomposition on the original regional load sequence to obtain multiple regional load subsequences and remainders; The prediction module is used to construct multi-dimensional fusion features based on the multiple regional load sub-sequences, residual items, meteorological measured data, meteorological forecast data, weekday data and holiday data, and input the multi-dimensional fusion features into a pre-trained lightweight regional short-term load integrated prediction model to obtain the future regional short-term load prediction results. The lightweight regional short-term load integrated prediction model is trained in the following way: Historical load sequences and corresponding meteorological measured data, meteorological forecast data, weekday data, and holiday data for the sample area are obtained. The historical load sequences are decomposed into signals and multidimensional fusion features are constructed for training. A deep neural network combined with ensemble learning is used to train an initial regional short-term load ensemble prediction model. The model parameters are jointly optimized based on the prediction errors of the original load sequence, regional load subsequence, and residual terms. The initial regional short-term load ensemble prediction model that meets the prediction accuracy requirements is pruned based on weight importance. The initial regional short-term load ensemble prediction model is used as the teacher network and the pruned model is used as the student network for knowledge distillation fine-tuning to obtain the lightweight regional short-term load ensemble prediction model.

[0018] A further improvement of the present invention is that the signal decomposition adopts empirical mode decomposition or variational mode decomposition, which decomposes the original regional load sequence into N regional load subsequences and a remainder term.

[0019] A further improvement of the present invention is that the multidimensional fusion features include multiple regional load subsequences, remainders, meteorological measured data features, meteorological forecast data features, weekday type features, and holiday type features.

[0020] A further improvement of the present invention is that both the measured meteorological data and the forecast meteorological data include air pressure, air temperature, relative humidity, hourly precipitation, and maximum wind speed.

[0021] A further improvement of the present invention is that the initial regional short-term load integrated prediction model includes a fusion feature extraction module, a load component feature extraction module, an intermediate feature aggregation layer, a regional load subsequence prediction layer, a remainder prediction layer, a regional load subsequence and remainder prediction result aggregation layer, and an original regional load sequence prediction layer.

[0022] A further improvement of the present invention is that: the fusion feature extraction module is used to extract the deep implicit features of the multidimensional fusion features used for training using a fully connected deep neural network; and the load component feature extraction module is used to extract the deep implicit features of the load subsequences and remainders of each region sequentially using a convolutional neural network and a long short-term memory neural network.

[0023] A further improvement of this invention is that the step of training the initial regional short-term load ensemble prediction model using a deep neural network combined with ensemble learning specifically includes: The deep implicit features output by the fusion feature extraction module are input into the intermediate feature aggregation layer along with the deep implicit features of the load subsequences of each region and the deep implicit features of the remainder, respectively, to obtain the aggregated features corresponding to the load subsequences of each region and the aggregated features corresponding to the remainder. The aggregated features corresponding to the load subsequences of each region are input into the regional load subsequence prediction layer to obtain the prediction results of each regional load subsequence; the aggregated features corresponding to the remainder are input into the remainder prediction layer to obtain the remainder prediction results; the prediction results of each regional load subsequence and the remainder prediction results are input into the prediction result aggregation layer to obtain the aggregated prediction results, and the regional load prediction results are obtained based on the aggregated prediction results.

[0024] A further improvement of the present invention is that the step of jointly optimizing the model parameters based on the prediction errors of the original load sequence, the regional load subsequence, and the remainder specifically includes: constructing a joint loss function based on the prediction errors corresponding to the regional load prediction results, the prediction errors corresponding to the prediction results of each regional load subsequence, and the prediction errors corresponding to the prediction results of the remainder, and updating the model parameters of the initial regional short-term load integrated prediction model based on the joint loss function.

[0025] A further improvement of the present invention is that the joint loss function is a weighted combination of the root mean square error of the regional load forecast result, the root mean square error of the forecast results of each regional load subsequence, and the root mean square error of the forecast results of the remainder.

[0026] A further improvement of this invention is that the step of performing weight-based pruning on the initial regional short-term load integrated prediction model that meets the prediction accuracy requirements specifically includes: The weight tensors of each layer of each network module in the initial regional short-term load integrated prediction model are sorted in ascending order according to the absolute value of the weights and expanded into a one-dimensional weight vector. The importance score of each weight in the one-dimensional weight vector is calculated. The corresponding weights are set to zero according to the importance score from small to large until the preset sparsity of each layer is met.

[0027] A further improvement of this invention is that the step of performing knowledge distillation fine-tuning using the initial regional short-term load ensemble prediction model as the teacher network and the pruned model as the student network to obtain the lightweight regional short-term load ensemble prediction model specifically includes: Keeping the model parameters of the initial short-term load ensemble prediction model of the teacher network unchanged, the parameters of the pruned model of the student network are updated based on the training set so that the student network learns the prediction output and intermediate feature representation of the teacher network; the knowledge distillation fine-tuning constructs a total distillation loss based on feature distillation loss and logistic distillation loss, and optimizes and updates the model parameters of the student network according to the total distillation loss.

[0028] Thirdly, the present invention provides an electronic device including a processor and a memory, the processor being used to execute a computer program stored in the memory to implement the aforementioned regional short-term load integrated prediction method based on multi-dimensional fusion features and lightweight model.

[0029] Fourthly, the present invention provides a computer-readable storage medium storing at least one instruction, which, when executed by a processor, implements the regional short-term load integrated prediction method based on multi-dimensional fusion features and model lightweighting.

[0030] Compared with the prior art, the present invention has the following beneficial effects: This invention decomposes the original regional load sequence of the region to be predicted into multiple regional load sub-sequences and remainders. It further integrates measured meteorological data, forecast data, weekday data, and holiday data to construct multi-dimensional fusion features, enabling the model input to simultaneously characterize the periodicity, trend, randomness, and the impact of external environmental factors on the regional load. Simultaneously, by employing a deep neural network combined with ensemble learning to train the initial regional short-term load ensemble prediction model, and jointly optimizing the model parameters based on the prediction errors of the original load sequence, regional load sub-sequences, and remainders, the invention improves the model's ability to learn different load components and overall load variation patterns, thereby enhancing the accuracy of regional short-term load prediction. Furthermore, through weighted pruning and knowledge distillation fine-tuning, a lightweight regional short-term load ensemble prediction model is obtained, which reduces model complexity and computational resource consumption while maintaining prediction performance, improving the feasibility of deployment on local devices and inference efficiency. This invention can finely characterize and learn the inherent periodicity and trend within the load sequence, further improving prediction accuracy compared to existing prediction methods.

[0031] Furthermore, this invention decomposes the original regional load sequence into N regional load subsequences and a remainder term by employing empirical mode decomposition or variational mode decomposition. This enables the complex and non-stationary regional load sequence to be split into multiple components with distinct characteristics, allowing load components with different frequencies or variation patterns to be characterized separately. This facilitates subsequent models to learn the periodicity, trend, and random fluctuation characteristics in the load sequence separately, avoiding the feature aliasing problem caused by directly modeling the original load sequence as a whole, thereby improving the refinement and accuracy of regional load forecasting.

[0032] Furthermore, this invention uses multiple regional load subsequences, residual terms, meteorological measured data features, meteorological forecast data features, weekday type features, and holiday type features as multi-dimensional fusion features. This enables the model to characterize regional load change characteristics from both the load's own variation patterns and external influencing factors, making the model's input information more comprehensive. Specifically, regional load subsequences and residual terms are used to characterize the internal time-frequency domain characteristics of the load, while meteorological, weekday, and holiday features are used to characterize the impact of external coupling factors on the electricity load. This helps to improve the model's comprehensive perception capability of short-term regional load changes.

[0033] Furthermore, by using air pressure, temperature, relative humidity, hourly precipitation, and maximum wind speed as specific contents of meteorological measured data and meteorological forecast data, this invention can cover the main meteorological factors affecting regional electricity load changes. This allows the model to learn load change patterns using historical actual meteorological conditions and to make short-term load predictions by combining future meteorological forecast information, thereby improving the adaptability and reliability of regional short-term load predictions to meteorological changes.

[0034] Furthermore, this invention sets the initial regional short-term load integrated prediction model to include a fusion feature extraction module, a load component feature extraction module, an intermediate feature aggregation layer, a regional load subsequence prediction layer, a residual prediction layer, a regional load subsequence and residual prediction result aggregation layer, and an original regional load sequence prediction layer. This enables the model to form an integrated prediction structure that integrates multi-dimensional feature extraction, component prediction, result aggregation, and overall load prediction. This allows the model to learn the variation patterns of each regional load subsequence and residual separately, and to obtain the overall regional load prediction result through the aggregation of prediction results, thereby enhancing the model's ability to express complex regional load sequences.

[0035] Furthermore, this invention employs a fully connected deep neural network to extract deep implicit features from the multidimensional fusion features used in training, which can fully explore the nonlinear correlation between fusion features such as load, weather, weekday type, and holiday type; by using convolutional neural networks and long short-term memory neural networks to sequentially extract deep implicit features from the load subsequences and remainders of each region, it can take into account both the extraction of local change features and the learning of temporal dependencies, thereby improving the model's ability to model the variation patterns of regional load components.

[0036] Furthermore, this invention integrates the deep implicit features output by the fusion feature extraction module with the deep implicit features of each regional load subsequence and the deep implicit features of the remainder, enabling the full fusion of external influencing factors and the load component's own variation characteristics at the intermediate feature level. Furthermore, by obtaining the prediction results of each regional load subsequence and the remainder prediction results through the regional load subsequence prediction layer and the remainder prediction layer, and then obtaining the aggregated prediction result through the prediction result aggregation layer to form the regional load prediction result, this invention achieves independent prediction and overall integration of each load component. This helps reduce the error caused by a single model directly predicting the overall load, and improves the stability and accuracy of the prediction results.

[0037] Furthermore, this invention constructs a joint loss function based on the prediction errors corresponding to the regional load prediction results, the prediction errors corresponding to the prediction results of each regional load subsequence, and the prediction errors corresponding to the prediction results of the remainder, and updates the model parameters of the initial regional short-term load integrated prediction model accordingly. This allows the model training process to be constrained by both the overall load prediction error and the component prediction error, preventing the model from focusing only on the final output and ignoring the prediction accuracy of each load component. This improves the model's ability to collaboratively learn the relationship between the original regional load sequence, regional load subsequence, and remainder.

[0038] Furthermore, by setting the joint loss function as a weighted combination of the root mean square error of the regional load prediction results, the root mean square error of the prediction results of each regional load subsequence, and the root mean square error of the prediction results of the remainder, the present invention can measure the deviation of the overall prediction results, the prediction results of each load subsequence, and the prediction results of the remainder in a unified error metric. It also balances the influence of different prediction objectives on model training through weighting, thereby improving the pertinence and stability of model parameter optimization.

[0039] Furthermore, this invention sorts the weight tensors of each layer of each network module in the initial regional short-term load integrated prediction model in ascending order of absolute weight value and expands them into a one-dimensional weight vector. It calculates the importance score of each weight and sets the corresponding weights to zero according to the importance score from small to large until the preset sparsity of each layer is met. This can prioritize the removal of weight parameters that contribute less to the model prediction, thereby reducing the number of model parameters and the amount of computation while retaining as many key weights as possible that have a greater impact on the prediction performance. This improves the rationality and effectiveness of the lightweight model processing.

[0040] Furthermore, this invention maintains the model parameters of the initial regional short-term load ensemble prediction model (which serves as the teacher network) unchanged, and updates the parameters of the pruned model (which serves as the student network) based on the training set. This enables the pruned model to learn the prediction output and intermediate feature representation of the teacher network, compensating for the prediction performance loss caused by pruning. Simultaneously, by constructing a total distillation loss based on feature distillation loss and logistic distillation loss and optimizing the student network parameters, the student network can be constrained from both the intermediate feature level and the prediction output level. This allows the lightweight regional short-term load ensemble prediction model to maintain the prediction capability of the original model as much as possible while reducing model complexity, thereby improving inference efficiency and prediction reliability in local deployment scenarios. Attached Figure Description

[0041] The accompanying drawings, which form part of this specification, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart illustrating a regional short-term load integrated prediction method based on multi-dimensional fusion features and lightweight model according to an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the training of the lightweight regional short-term load integrated prediction model in an embodiment of the present invention. Figure 3 This is a flowchart illustrating another embodiment of the present invention: a regional short-term load integrated prediction method based on multi-dimensional fusion features and lightweight model. Figure 4This is a schematic diagram of the structure of a regional short-term load integrated prediction device based on multi-dimensional fusion features and lightweight model according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0042] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0043] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this invention is for describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0044] To address the limitations of existing AI-based regional short-term load forecasting methods, which often suffer from the inability to accurately represent and learn the inherent periodicity and trends within load sequences, and the difficulty in widespread deployment and application on local devices with limited storage and computing resources, this invention proposes a regional short-term load ensemble forecasting method based on multi-dimensional fusion features and a lightweight model. This method achieves accurate modeling of the periodicity, trends, and stochasticity of load sequences, significantly improving the accuracy of regional short-term load forecasting. Simultaneously, it achieves a lightweight regional short-term load ensemble forecasting model, enhancing the feasibility of deployment and application on local devices with limited storage and computing resources, as well as improving inference efficiency. Specifically, based on load and meteorological data, this invention constructs multi-dimensional fusion features, including load, meteorological, weekday type, and holiday type, as input features through signal decomposition technology from both time and frequency domain perspectives. A deep neural network, combined with ensemble learning, is used to establish a regional short-term load forecasting model. Furthermore, pruning and knowledge distillation techniques are combined to achieve more accurate regional short-term load forecasting, improving the feasibility of local model deployment and inference efficiency. Against the backdrop of the development and construction of new power systems, this invention can organically integrate multi-dimensional time and frequency domain features, accurately mine the periodicity, trend and randomness of load sequences, and effectively achieve lightweight prediction models, significantly improving the accuracy of regional short-term load forecasting and the efficiency of local inference calculations. It is of great significance for power grid optimization scheduling decisions, operation mode adjustments and power market transactions.

[0045] Please see Figure 1 As shown, this embodiment of the invention provides a regional short-term load integrated forecasting method based on multi-dimensional fusion features and lightweight model, including the following steps: Step 1: Using signal decomposition techniques, the original regional load sequence is decomposed into a series of regional load subsequences (i.e., mode functions) with distinct characteristics and a remainder term.

[0046] Step 2: Based on regional load data, meteorological measured data, weekday data, holiday data, and meteorological forecast data, construct multi-dimensional fusion features of load, meteorology, weekday type, and holiday type from the perspectives of time domain and frequency domain, establish a regional short-term load forecast sample set, and divide the sample set into a training set and a test set.

[0047] Step 3: Using the regional short-term load forecast training set, a deep neural network is employed, combined with ensemble learning, to comprehensively consider the prediction errors of the original regional load sequence, regional load sub-sequence, and residual terms, to train the regional short-term load forecast model and obtain the trained regional short-term load forecast model.

[0048] Step 4: Using a regional short-term load forecasting test set, evaluate the performance of the trained regional short-term load forecasting model based on prediction performance metrics such as root mean square error, mean absolute error, and mean absolute percentage error. If the prediction performance of the trained regional short-term load forecasting model meets the accuracy requirements of the application scenario, the trained regional short-term load forecasting model is lightweighted to obtain the original trained regional short-term load forecasting model; otherwise, optimize the model hyperparameters, retrain, and evaluate the regional short-term load forecasting model.

[0049] Step 5: Based on the importance of weights, pruning techniques are used to trim the original trained regional short-term load forecasting model, resulting in a pruned regional short-term load forecasting model. Further, based on the regional short-term load forecasting training set, knowledge distillation is used to use the original trained regional short-term load forecasting model as the teacher network and the pruned regional short-term load forecasting model as the student network. The teacher network is then used to fine-tune the student network, ultimately obtaining a lightweight regional short-term load forecasting model.

[0050] Step 6: Collect regional load sequences, measured meteorological data, weekday data, holiday data, and weather forecast data for the area to be predicted; decompose the regional load sequence of the area to be predicted into a series of distinct regional load subsequences. l i And the remaining term to be predicted, r; From the perspectives of time and frequency domains, a multi-dimensional fusion feature S=[ is constructed based on load, meteorology, weekday type, and holiday type. l 1, l 2, l 3,…, l N , r,M,P,W,HThe lightweight regional short-term load forecasting model is input to perform forecasting, and regional short-term load forecasting results based on multi-dimensional fusion features and model lightweighting are obtained.

[0051] in, l i Let N be the load subsequence of the i-th region, N be the number of regional load subsequences, M be the meteorological data characteristics, P be the meteorological forecast data characteristics, W be the weekday type characteristics, and H be the holiday type characteristics.

[0052] Please see Figures 1 to 2 As shown, this embodiment of the invention provides a regional short-term load integrated prediction method based on multi-dimensional fusion features and lightweight model, specifically including the following steps: Step 1: Obtain the original regional load sequence, weekday data, holiday data, actual meteorological data, and meteorological forecast data; and decompose the original regional load sequence into a series of regional load subsequences (i.e., modal functions) and a remainder term using signal decomposition techniques, as shown in Equation (1): (1) Where L is the original regional load sequence, l i Let be the i-th region load subsequence, N be the number of region load subsequences, and r be the remainder. In one specific embodiment, the signal decomposition technique used is empirical mode decomposition or variational mode decomposition.

[0053] Step 2: Based on regional load subsequences, residual terms, weekday data, holiday data, measured meteorological data, and meteorological forecast data, construct multi-dimensional fusion features of load, meteorology, weekday type, and holiday type from the time and frequency domains to establish a regional short-term load forecast sample set. The sample set was divided into a training set and a training set in an 8:2 ratio. and test set ; Where M represents the characteristics of actual meteorological data, P represents the characteristics of meteorological forecast data, W represents the characteristics of weekday types, and H represents the characteristics of holiday types.

[0054] In one specific implementation, the measured meteorological data and the forecast meteorological data include air pressure, temperature, relative humidity, hourly precipitation, and maximum wind speed.

[0055] Step 3: Please refer to Figure 2 As shown, the training set for regional short-term load forecasting... A short-term regional load forecasting model is trained by using deep neural networks and combining them with ensemble learning. This model comprehensively considers the prediction errors of the original regional load sequence, the regional load subsequence, and the remainder.

[0056] In one specific implementation, the deep neural network used is a fully connected deep neural network (FDNN), a convolutional neural network (CNN), or a long short-term memory neural network (LSTM).

[0057] In one specific embodiment, step S3 specifically includes: Step 3.1: Extract the training set for regional short-term load forecasting using the FDNN module. Integration characteristics The deep latent features are obtained by extracting the first deep latent feature, and the CNN module and LSTM module are used to extract the training set for short-term load prediction in the region in turn. Load subsequences of various regions and remainder The deep implicit features are obtained, and the deep implicit features of the load in the i-th region and the deep implicit features of the remainder are obtained.

[0058] Then, the first deep hidden feature and the deep hidden feature of the i-th region load are aggregated through the intermediate feature aggregation layer; the first deep hidden feature and the remaining deep hidden features are aggregated through the intermediate feature aggregation layer.

[0059] Step 3.2: Based on the aggregation features of the intermediate feature aggregation layer, the prediction results of each regional load subsequence are obtained through the regional load subsequence prediction layer and the residual term prediction layer, respectively. Prediction results of the remainder term .

[0060] Step 3.3: Prediction results based on load subsequences of each region Prediction results of the remainder term The aggregated prediction result is obtained through an aggregation layer of regional load subsequences and residual prediction results. .

[0061] Step 3.4: Based on the aggregated prediction results The final regional load forecast result is obtained through the original regional load sequence forecast layer. .

[0062] Step 3.5: Based on the aggregated prediction results and regional load forecast results Taking into account the prediction errors of the original regional load sequence, the regional load subsequence, and the remainder, the model parameters are optimized and updated according to the loss function shown in equation (2): (2) in, Regional load forecast results The corresponding true value, Prediction results for load subsequences in each region The corresponding true value, For the prediction results of the remainder The corresponding true value, Here, α, β, and γ are the root mean square error operator, and they are coefficients. and .

[0063] Step 3.6: Repeat steps 3.1 to 3.5 until the maximum training round T, then end the model training and obtain the trained regional short-term load prediction model.

[0064] Step 4: Using the regional short-term load forecasting test set The performance of the trained regional short-term load forecasting model is evaluated based on prediction performance metrics such as root mean square error, mean absolute error, and mean absolute percentage error. If the prediction performance of the trained regional short-term load forecasting model meets the accuracy requirements of the application scenario, the trained regional short-term load forecasting model is lightweighted; otherwise, the model hyperparameters are optimized, and the regional short-term load forecasting model is retrained and evaluated.

[0065] Step 5: Based on the weight importance, pruning techniques are used to trim the original trained regional short-term load forecasting model to obtain the pruned regional short-term load forecasting model. Further, based on the regional short-term load forecasting training set... This study employs knowledge distillation, using the original trained regional short-term load prediction model as the teacher network and the pruned regional short-term load prediction model as the student network. The teacher network is then used to fine-tune the student network, ultimately resulting in a lightweight regional short-term load prediction model. Typically, the knowledge distillation techniques used include logical distillation and feature distillation.

[0066] Step 5.1: Sort the weight tensors of each layer of each module (including the FDNN module, CNN module and LSTM module) in ascending order of absolute weight value and expand them into a one-dimensional weight vector W. Calculate the importance score of each weight, as shown in Equation (3): (3) in, To score the importance of the u-th element of the one-dimensional weight vector W, and Let u and v be the u-th and v-th elements of the one-dimensional weight vector W, respectively, and when hour, Heng was established.

[0067] Step 5.2: Based on the importance score of each weight, prune in ascending order (i.e., reset weights with lower importance scores to zero) until the required sparsity for each layer is met. This leads to the acquisition of a short-term load forecasting model for the pruned region.

[0068] Step 5.3: Training set based on regional short-term load forecasting The original trained regional short-term load prediction model is used as the teacher network, and the pruned regional short-term load prediction model is used as the student network. The student network is fine-tuned using the teacher network. That is, the parameters of the original well-trained regional short-term load prediction model remain unchanged, while taking into account both feature-based distillation loss and logic-based distillation loss. The parameters of the pruned regional short-term load prediction model are optimized and updated according to the loss functions shown in equations (4) to (6): (4) (5) (6) in, For feature-based distillation loss, For logic-based distillation loss, For total distillation losses, The regional load forecast results are from the pruned regional short-term load forecasting model. This represents the prediction result of the i-th regional load subsequence in the pruned regional short-term load forecasting model. The remaining terms of the short-term load forecasting model for the pruned region represent the forecast results. and .

[0069] Step 5.4: Repeat step 5.3 until the maximum number of training rounds M, then end the model fine-tuning and finally obtain a lightweight regional short-term load prediction model.

[0070] Please see Figure 3 As shown, this invention provides a regional short-term load integrated forecasting method based on multi-dimensional fusion features and lightweight model, comprising: S1. Obtain the original regional load sequence, actual meteorological data, meteorological forecast data, weekday data, and holiday data of the area to be predicted, and perform signal decomposition on the original regional load sequence to obtain multiple regional load sub-sequences and remainders; S2. Construct multi-dimensional fusion features based on the multiple regional load sub-sequences, residual items, meteorological measured data, meteorological forecast data, weekday data and holiday data, and input the multi-dimensional fusion features into a pre-trained lightweight regional short-term load integrated prediction model to obtain the future regional short-term load prediction results. The lightweight regional short-term load integrated prediction model is trained in the following way: Historical load sequences and corresponding meteorological measured data, meteorological forecast data, weekday data, and holiday data for the sample area are obtained. The historical load sequences are decomposed into signals and multidimensional fusion features are constructed for training. A deep neural network combined with ensemble learning is used to train an initial regional short-term load ensemble prediction model. The model parameters are jointly optimized based on the prediction errors of the original load sequence, regional load subsequence, and residual terms. The initial regional short-term load ensemble prediction model that meets the prediction accuracy requirements is pruned based on weight importance. The initial regional short-term load ensemble prediction model is used as the teacher network and the pruned model is used as the student network for knowledge distillation fine-tuning to obtain the lightweight regional short-term load ensemble prediction model.

[0071] In one specific embodiment, in the step of performing signal decomposition on the original regional load sequence to obtain multiple regional load sub-sequences and a remainder term, the signal decomposition employs empirical mode decomposition or variational mode decomposition, decomposing the original regional load sequence into N regional load sub-sequences and a remainder term: (1) Where L is the original regional load sequence, l i Let be the i-th region load subsequence, N be the number of region load subsequences, and r be the remainder. In one specific embodiment, the signal decomposition technique used is empirical mode decomposition or variational mode decomposition.

[0072] In one specific implementation, the multidimensional fusion features include multiple regional load subsequences, residual terms, meteorological measured data features, meteorological forecast data features, weekday type features, and holiday type features:

[0073] Where M represents the characteristics of actual meteorological data, P represents the characteristics of meteorological forecast data, W represents the characteristics of weekday types, and H represents the characteristics of holiday types.

[0074] In one specific embodiment, both the measured meteorological data and the forecast meteorological data include air pressure, air temperature, relative humidity, hourly precipitation, and maximum wind speed.

[0075] In one specific implementation, the initial regional short-term load integrated prediction model includes a fusion feature extraction module, a load component feature extraction module, an intermediate feature aggregation layer, a regional load subsequence prediction layer, a remainder prediction layer, a regional load subsequence and remainder prediction result aggregation layer, and an original regional load sequence prediction layer.

[0076] In one specific embodiment, the fusion feature extraction module is used to extract the deep implicit features of the training multidimensional fusion features using a fully connected deep neural network, and the load component feature extraction module is used to extract the deep implicit features of the load subsequences and remainders of each region sequentially using a convolutional neural network and a long short-term memory neural network.

[0077] In one specific embodiment, the step of training an initial regional short-term load ensemble prediction model using a deep neural network combined with ensemble learning specifically includes: The deep implicit features output by the fusion feature extraction module are input into the intermediate feature aggregation layer along with the deep implicit features of the load subsequences of each region and the deep implicit features of the remainder, respectively, to obtain the aggregated features corresponding to the load subsequences of each region and the aggregated features corresponding to the remainder. The aggregated features corresponding to the load subsequences of each region are input into the regional load subsequence prediction layer to obtain the prediction results of each regional load subsequence; the aggregated features corresponding to the remainder are input into the remainder prediction layer to obtain the remainder prediction results; the prediction results of each regional load subsequence and the remainder prediction results are input into the prediction result aggregation layer to obtain the aggregated prediction results, and the regional load prediction results are obtained based on the aggregated prediction results.

[0078] In one specific implementation, the step of jointly optimizing the model parameters based on the prediction errors of the original load sequence, regional load subsequences, and residual terms specifically includes: constructing a joint loss function based on the prediction errors corresponding to the regional load prediction results, the prediction errors corresponding to the prediction results of each regional load subsequence, and the prediction errors corresponding to the prediction results of the residual terms, and updating the model parameters of the initial regional short-term load integrated prediction model based on the joint loss function.

[0079] In one specific implementation, the joint loss function is: (2) in, Regional load forecast results The corresponding true value, Prediction results for load subsequences in each region The corresponding true value, For the prediction results of the remainder The corresponding true value, Here, α, β, and γ are the root mean square error operator, and they are coefficients. and .

[0080] In one specific implementation, the joint loss function is a weighted combination of the root mean square error of the regional load forecast results, the root mean square error of the forecast results of each regional load subsequence, and the root mean square error of the forecast results of the remainder.

[0081] In one specific implementation, the step of performing weighted importance-based pruning on the initial regional short-term load integrated forecasting model that meets the forecasting accuracy requirements specifically includes: The weight tensors of each layer of each network module in the initial regional short-term load integrated prediction model are sorted in ascending order according to the absolute value of the weights and expanded into a one-dimensional weight vector. The importance score of each weight in the one-dimensional weight vector is calculated. The corresponding weights are set to zero according to the importance score from small to large until the preset sparsity of each layer is met.

[0082] (3) in, To score the importance of the u-th element of the one-dimensional weight vector W, and Let u and v be the u-th and v-th elements of the one-dimensional weight vector W, respectively, and when hour, Heng was established.

[0083] In one specific embodiment, the step of performing knowledge distillation fine-tuning using the initial regional short-term load ensemble prediction model as the teacher network and the pruned model as the student network to obtain the lightweight regional short-term load ensemble prediction model specifically includes: Keeping the model parameters of the initial short-term load ensemble prediction model (serving as the teacher network) unchanged, the parameters of the pruned model (serving as the student network) are updated based on the training set, so that the student network learns the prediction output and intermediate feature representation of the teacher network; the knowledge distillation fine-tuning constructs a total distillation loss based on feature distillation loss and logistic distillation loss, and optimizes and updates the model parameters of the student network according to the total distillation loss. (4) (5) (6) in, For feature-based distillation loss, For logic-based distillation loss, For total distillation losses, The regional load forecast results are from the pruned regional short-term load forecasting model. This represents the prediction result of the i-th regional load subsequence in the pruned regional short-term load forecasting model. The remaining terms of the short-term load forecasting model for the pruned region represent the forecast results. and .

[0084] Please see Figure 4 As shown, this embodiment of the invention provides a regional short-term load integrated forecasting device based on multi-dimensional fusion features and lightweight model, comprising: The data acquisition module is used to acquire the original regional load sequence, measured meteorological data, meteorological forecast data, weekday data and holiday data of the area to be predicted, and to perform signal decomposition on the original regional load sequence to obtain multiple regional load subsequences and remainders; The prediction module is used to construct multi-dimensional fusion features based on the multiple regional load sub-sequences, residual items, meteorological measured data, meteorological forecast data, weekday data and holiday data, and input the multi-dimensional fusion features into a pre-trained lightweight regional short-term load integrated prediction model to obtain the future regional short-term load prediction results. The lightweight regional short-term load integrated prediction model is trained in the following way: Historical load sequences and corresponding meteorological measured data, meteorological forecast data, weekday data, and holiday data for the sample area are obtained. The historical load sequences are decomposed into signals and multidimensional fusion features are constructed for training. A deep neural network combined with ensemble learning is used to train an initial regional short-term load ensemble prediction model. The model parameters are jointly optimized based on the prediction errors of the original load sequence, regional load subsequence, and residual terms. The initial regional short-term load ensemble prediction model that meets the prediction accuracy requirements is pruned based on weight importance. The initial regional short-term load ensemble prediction model is used as the teacher network and the pruned model is used as the student network for knowledge distillation fine-tuning to obtain the lightweight regional short-term load ensemble prediction model.

[0085] Please see Figure 5 As shown, this embodiment of the invention provides an electronic device 100 for implementing a regional short-term load integrated prediction method based on multi-dimensional fusion features and model lightweighting; the electronic device 100 includes a memory 101, at least one processor 102, a computer program 103 stored in the memory 101 and executable on the at least one processor 102, and at least one communication bus 104.

[0086] The memory 101 can be used to store the computer program 103. The processor 102 implements the steps of the regional short-term load integrated prediction method based on multi-dimensional fusion features and model lightweighting described in the embodiment by running or executing the computer program stored in the memory 101 and calling the data stored in the memory 101. The memory 101 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the electronic device 100 (such as audio data), etc. In addition, the memory 101 may include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other non-volatile solid-state storage device.

[0087] The at least one processor 102 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor 102 may be a microprocessor or any conventional processor. The processor 102 is the control center of the electronic device 100, connecting various parts of the electronic device 100 via various interfaces and lines.

[0088] The memory 101 in the electronic device 100 stores multiple instructions to implement a regional short-term load integrated prediction method based on multi-dimensional fusion features and lightweight models. The processor 102 can execute the multiple instructions to achieve the following: The original regional load sequence, measured meteorological data, meteorological forecast data, weekday data, and holiday data of the area to be predicted are obtained. The original regional load sequence is then decomposed to obtain multiple regional load subsequences and remainders. Based on the multiple regional load subsequences, residual items, meteorological measured data, meteorological forecast data, weekday data and holiday data, a multi-dimensional fusion feature is constructed. The multi-dimensional fusion feature is then input into a pre-trained lightweight regional short-term load integrated prediction model to obtain the future regional short-term load prediction results. The lightweight regional short-term load integrated prediction model is trained in the following way: Historical load sequences and corresponding meteorological measured data, meteorological forecast data, weekday data, and holiday data for the sample area are obtained. The historical load sequences are decomposed into signals and multidimensional fusion features are constructed for training. A deep neural network combined with ensemble learning is used to train an initial regional short-term load ensemble prediction model. The model parameters are jointly optimized based on the prediction errors of the original load sequence, regional load subsequence, and residual terms. The initial regional short-term load ensemble prediction model that meets the prediction accuracy requirements is pruned based on weight importance. The initial regional short-term load ensemble prediction model is used as the teacher network and the pruned model is used as the student network for knowledge distillation fine-tuning to obtain the lightweight regional short-term load ensemble prediction model.

[0089] If the modules / units integrated in the electronic device 100 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, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, and a read-only memory (ROM).

[0090] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0091] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0092] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0093] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0094] 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 it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A regional short-term load ensemble forecasting method based on multi-dimensional fusion features and lightweight model, characterized in that, include: The original regional load sequence, measured meteorological data, meteorological forecast data, weekday data, and holiday data of the area to be predicted are obtained. The original regional load sequence is then decomposed to obtain multiple regional load subsequences and remainders. Based on the multiple regional load subsequences, residual items, meteorological measured data, meteorological forecast data, weekday data and holiday data, a multi-dimensional fusion feature is constructed. The multi-dimensional fusion feature is then input into a pre-trained lightweight regional short-term load integrated prediction model to obtain the future regional short-term load prediction results. The lightweight regional short-term load integrated prediction model is trained in the following way: Historical load sequences and corresponding meteorological measured data, meteorological forecast data, weekday data, and holiday data for the sample area are obtained. The historical load sequences are decomposed into signals and multidimensional fusion features are constructed for training. A deep neural network combined with ensemble learning is used to train an initial regional short-term load ensemble prediction model. The model parameters are jointly optimized based on the prediction errors of the original load sequence, regional load subsequence, and residual terms. The initial regional short-term load ensemble prediction model that meets the prediction accuracy requirements is pruned based on weight importance. The initial regional short-term load ensemble prediction model is used as the teacher network and the pruned model is used as the student network for knowledge distillation fine-tuning to obtain the lightweight regional short-term load ensemble prediction model.

2. The regional short-term load integrated forecasting method based on multi-dimensional fusion features and lightweight model as described in claim 1, characterized in that, In the step of performing signal decomposition on the original regional load sequence to obtain multiple regional load subsequences and a remainder term, the signal decomposition adopts empirical mode decomposition or variational mode decomposition to decompose the original regional load sequence into N regional load subsequences and a remainder term.

3. The regional short-term load integrated forecasting method based on multi-dimensional fusion features and lightweight model as described in claim 1, characterized in that, The multidimensional fusion features include multiple regional load subsequences, remainders, meteorological measured data features, meteorological forecast data features, weekday type features, and holiday type features.

4. The regional short-term load integrated forecasting method based on multi-dimensional fusion features and lightweight model as described in claim 1, characterized in that, Both the measured meteorological data and the forecast meteorological data include air pressure, air temperature, relative humidity, hourly precipitation, and maximum wind speed.

5. The regional short-term load integrated forecasting method based on multi-dimensional fusion features and lightweight model as described in claim 1, characterized in that, The initial regional short-term load integrated prediction model includes a fusion feature extraction module, a load component feature extraction module, an intermediate feature aggregation layer, a regional load subsequence prediction layer, a remainder prediction layer, a regional load subsequence and remainder prediction result aggregation layer, and an original regional load sequence prediction layer.

6. The regional short-term load integrated forecasting method based on multi-dimensional fusion features and lightweight model as described in claim 5, characterized in that, The fusion feature extraction module is used to extract the deep implicit features of the training multidimensional fusion features using a fully connected deep neural network. The load component feature extraction module is used to extract the deep implicit features of the load subsequences and remainders of each region sequentially using a convolutional neural network and a long short-term memory neural network.

7. The regional short-term load integrated forecasting method based on multi-dimensional fusion features and lightweight model as described in claim 6, characterized in that, The step of training an initial regional short-term load ensemble prediction model using a deep neural network combined with ensemble learning specifically includes: The deep implicit features output by the fusion feature extraction module are input into the intermediate feature aggregation layer along with the deep implicit features of the load subsequences of each region and the deep implicit features of the remainder, respectively, to obtain the aggregated features corresponding to the load subsequences of each region and the aggregated features corresponding to the remainder. The aggregated features corresponding to the load subsequences of each region are input into the regional load subsequence prediction layer to obtain the prediction results of each regional load subsequence; the aggregated features corresponding to the remainder are input into the remainder prediction layer to obtain the remainder prediction results; the prediction results of each regional load subsequence and the remainder prediction results are input into the prediction result aggregation layer to obtain the aggregated prediction results, and the regional load prediction results are obtained based on the aggregated prediction results.

8. The regional short-term load integrated forecasting method based on multi-dimensional fusion features and lightweight model as described in claim 1, characterized in that, The step of jointly optimizing the model parameters based on the prediction errors of the original load sequence, regional load subsequences, and residual terms specifically includes: constructing a joint loss function based on the prediction errors corresponding to the regional load prediction results, the prediction errors corresponding to the prediction results of each regional load subsequence, and the prediction errors corresponding to the prediction results of the residual terms, and updating the model parameters of the initial regional short-term load integrated prediction model based on the joint loss function.

9. The regional short-term load integrated forecasting method based on multi-dimensional fusion features and lightweight model as described in claim 8, characterized in that, The joint loss function is a weighted combination of the root mean square error of the regional load forecast results, the root mean square error of the forecast results of each regional load subsequence, and the root mean square error of the forecast results of the remaining terms.

10. The regional short-term load integrated forecasting method based on multi-dimensional fusion features and lightweight model as described in claim 6, characterized in that, The step of performing weight-based pruning on the initial regional short-term load integrated forecasting model that meets the forecasting accuracy requirements specifically includes: The weight tensors of each layer of each network module in the initial regional short-term load integrated prediction model are sorted in ascending order according to the absolute value of the weights and expanded into a one-dimensional weight vector. The importance score of each weight in the one-dimensional weight vector is calculated. The corresponding weights are set to zero according to the importance score from small to large until the preset sparsity of each layer is met.

11. The regional short-term load integrated forecasting method based on multi-dimensional fusion features and lightweight model as described in claim 10, characterized in that, The step of performing knowledge distillation fine-tuning using the initial regional short-term load ensemble prediction model as the teacher network and the pruned model as the student network to obtain the lightweight regional short-term load ensemble prediction model specifically includes: Keeping the model parameters of the initial short-term load ensemble prediction model of the teacher network unchanged, the parameters of the pruned model of the student network are updated based on the training set so that the student network learns the prediction output and intermediate feature representation of the teacher network; the knowledge distillation fine-tuning constructs a total distillation loss based on feature distillation loss and logistic distillation loss, and optimizes and updates the model parameters of the student network according to the total distillation loss.

12. A regional short-term load integrated forecasting device based on multi-dimensional fusion features and lightweight model, characterized in that, include: The data acquisition module is used to acquire the original regional load sequence, measured meteorological data, meteorological forecast data, weekday data and holiday data of the area to be predicted, and to perform signal decomposition on the original regional load sequence to obtain multiple regional load subsequences and remainders; The prediction module is used to construct multi-dimensional fusion features based on the multiple regional load sub-sequences, residual items, meteorological measured data, meteorological forecast data, weekday data and holiday data, and input the multi-dimensional fusion features into a pre-trained lightweight regional short-term load integrated prediction model to obtain the future regional short-term load prediction results. The lightweight regional short-term load integrated prediction model is trained in the following way: Historical load sequences and corresponding meteorological measured data, meteorological forecast data, weekday data, and holiday data for the sample area are obtained. The historical load sequences are decomposed into signals and multidimensional fusion features are constructed for training. A deep neural network combined with ensemble learning is used to train an initial regional short-term load ensemble prediction model. The model parameters are jointly optimized based on the prediction errors of the original load sequence, regional load subsequence, and residual terms. The initial regional short-term load ensemble prediction model that meets the prediction accuracy requirements is pruned based on weight importance. The initial regional short-term load ensemble prediction model is used as the teacher network and the pruned model is used as the student network for knowledge distillation fine-tuning to obtain the lightweight regional short-term load ensemble prediction model.

13. The regional short-term load integrated forecasting device based on multi-dimensional fusion features and lightweight model according to claim 12, characterized in that, The signal decomposition employs empirical mode decomposition or variational mode decomposition to decompose the original regional load sequence into N regional load subsequences and a remainder term.

14. The regional short-term load integrated forecasting device based on multi-dimensional fusion features and lightweight model according to claim 12, characterized in that, The multidimensional fusion features include multiple regional load subsequences, remainders, meteorological measured data features, meteorological forecast data features, weekday type features, and holiday type features.

15. The regional short-term load integrated forecasting device based on multi-dimensional fusion features and lightweight model according to claim 12, characterized in that, Both the measured meteorological data and the forecast meteorological data include air pressure, air temperature, relative humidity, hourly precipitation, and maximum wind speed.

16. The regional short-term load integrated forecasting device based on multi-dimensional fusion features and lightweight model according to claim 12, characterized in that, The initial regional short-term load integrated prediction model includes a fusion feature extraction module, a load component feature extraction module, an intermediate feature aggregation layer, a regional load subsequence prediction layer, a remainder prediction layer, a regional load subsequence and remainder prediction result aggregation layer, and an original regional load sequence prediction layer.

17. The regional short-term load integrated forecasting device based on multi-dimensional fusion features and lightweight model as described in claim 16, characterized in that, The fusion feature extraction module is used to extract the deep implicit features of the training multidimensional fusion features using a fully connected deep neural network. The load component feature extraction module is used to extract the deep implicit features of the load subsequences and remainders of each region sequentially using a convolutional neural network and a long short-term memory neural network.

18. The regional short-term load integrated forecasting device based on multi-dimensional fusion features and lightweight model according to claim 17, characterized in that, The step of training an initial regional short-term load ensemble prediction model using a deep neural network combined with ensemble learning specifically includes: The deep implicit features output by the fusion feature extraction module are input into the intermediate feature aggregation layer along with the deep implicit features of the load subsequences of each region and the deep implicit features of the remainder, respectively, to obtain the aggregated features corresponding to the load subsequences of each region and the aggregated features corresponding to the remainder. The aggregated features corresponding to the load subsequences of each region are input into the regional load subsequence prediction layer to obtain the prediction results of each regional load subsequence; the aggregated features corresponding to the remainder are input into the remainder prediction layer to obtain the remainder prediction results; the prediction results of each regional load subsequence and the remainder prediction results are input into the prediction result aggregation layer to obtain the aggregated prediction results, and the regional load prediction results are obtained based on the aggregated prediction results.

19. The regional short-term load integrated forecasting device based on multi-dimensional fusion features and lightweight model according to claim 12, characterized in that, The step of jointly optimizing the model parameters based on the prediction errors of the original load sequence, regional load subsequences, and residual terms specifically includes: constructing a joint loss function based on the prediction errors corresponding to the regional load prediction results, the prediction errors corresponding to the prediction results of each regional load subsequence, and the prediction errors corresponding to the prediction results of the residual terms, and updating the model parameters of the initial regional short-term load integrated prediction model based on the joint loss function.

20. The regional short-term load integrated forecasting device based on multi-dimensional fusion features and lightweight model according to claim 19, characterized in that, The joint loss function is a weighted combination of the root mean square error of the regional load forecast results, the root mean square error of the forecast results of each regional load subsequence, and the root mean square error of the forecast results of the remaining terms.

21. The regional short-term load integrated forecasting device based on multi-dimensional fusion features and lightweight model according to claim 17, characterized in that, The step of performing weight-based pruning on the initial regional short-term load integrated forecasting model that meets the forecasting accuracy requirements specifically includes: The weight tensors of each layer of each network module in the initial regional short-term load integrated prediction model are sorted in ascending order according to the absolute value of the weights and expanded into a one-dimensional weight vector. The importance score of each weight in the one-dimensional weight vector is calculated. The corresponding weights are set to zero according to the importance score from small to large until the preset sparsity of each layer is met.

22. The regional short-term load integrated forecasting device based on multi-dimensional fusion features and lightweight model according to claim 21, characterized in that, The step of performing knowledge distillation fine-tuning using the initial regional short-term load ensemble prediction model as the teacher network and the pruned model as the student network to obtain the lightweight regional short-term load ensemble prediction model specifically includes: Keeping the model parameters of the initial short-term load ensemble prediction model of the teacher network unchanged, the parameters of the pruned model of the student network are updated based on the training set so that the student network learns the prediction output and intermediate feature representation of the teacher network; the knowledge distillation fine-tuning constructs a total distillation loss based on feature distillation loss and logistic distillation loss, and optimizes and updates the model parameters of the student network according to the total distillation loss.

23. An electronic device, characterized in that, It includes a processor and a memory, the processor being used to execute a computer program stored in the memory to implement the regional short-term load integrated forecasting method based on multi-dimensional fusion features and model lightweighting as described in any one of claims 1 to 11.

24. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction that, when executed by a processor, implements the regional short-term load integrated forecasting method based on multi-dimensional fusion features and model lightweighting as described in any one of claims 1 to 11.

Images