Industrial load prediction method

Abstract

The application provides an industrial load prediction method, and relates to the technical field of load prediction. The method comprises the following steps: obtaining a historical industrial load time sequence of an industrial subject to be measured in a target area; inputting the historical industrial load time sequence into a prediction model corresponding to the industrial subject to be measured to obtain an industrial load prediction result of the industrial subject to be measured; wherein the prediction model is a local model trained in a federated learning framework; the federated learning framework comprises a global model corresponding to the target area and local models corresponding to each industrial subject in the target area, and the global model and each local model are trained in a meta-gradient updating manner to obtain the trained local model. The application can improve the industrial load prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of load forecasting technology, and more particularly to an industrial load forecasting method. Background Technology

[0002] Against the backdrop of the power system's transformation towards intelligent and refined operation, regional industrial short-term load forecasting is a core foundation for the safe and stable operation of the power system, refined dispatching, and optimal energy allocation. It plays an indispensable and crucial role in scenarios such as grid dispatching decisions, demand response implementation, energy efficiency assessment, and renewable energy consumption. With the accelerated transformation of the power system towards intelligence and digitalization, and the rapid development of regional industrial economies, the scale of regional industrial loads continues to expand, and their fluctuation characteristics are becoming increasingly complex, placing higher demands on the accuracy, stability, and scenario adaptability of short-term load forecasting.

[0003] Compared to other load types such as residential and commercial loads, regional industrial loads are highly coupled with the overall industrial production layout, production plans, equipment operating conditions, and production rhythm of the region, resulting in load changes exhibiting significant non-stationarity, strong nonlinearity, and abrupt changes. Due to the diversity of regional industrial production processes and the frequent switching of operating conditions, traditional statistical models (such as ARIMA) and single data-driven models (such as LSTM and Transformer) struggle to accurately capture the dynamic evolution of regional industrial loads, and their prediction accuracy falls short of practical application requirements. Furthermore, limitations imposed by factors such as the accumulation cycle of regional industrial data and data acquisition conditions often result in a limited scale of historical industrial load data used for modeling, leading to insufficient model training and inadequate generalization ability. Under new scenarios such as new regional industrial capacity layouts and adjustments to production conditions, the performance of existing prediction models significantly deteriorates, failing to quickly adapt to the dynamic changes in load characteristics. Summary of the Invention

[0004] This invention provides an industrial load forecasting method to address the problem that forecasting models in related technologies cannot adapt to the dynamic changes in industrial load, resulting in low forecasting accuracy.

[0005] In a first aspect, embodiments of the present invention provide an industrial load forecasting method, comprising: Obtain the historical industrial load time series of the industrial entities to be measured in the target area; The historical industrial load time series is input into the prediction model corresponding to the industrial entity to be measured to obtain the industrial load prediction result of the industrial entity to be measured. The prediction model is a local model trained in a federated learning framework. The federated learning framework includes a global model corresponding to the target region and local models corresponding to each industrial entity in the target region. The global model and each local model are trained using a meta-gradient update method to obtain the trained local model.

[0006] In one possible implementation, the training process for each local model includes: Receive the global model parameters of the global model, and train each local model based on the global model parameters to obtain the local model parameters of each local model; Based on the parameters of each local model, determine the meta-gradient corresponding to each local model; The global model parameters are updated based on each meta-gradient, and the process jumps to the next step of training each local model based on the updated model parameters to obtain the local model parameters of each local model. This process continues until a set number of training iterations are reached, at which point the trained global model and each local model are obtained.

[0007] In one possible implementation, the step of training each local model based on the global model parameters to obtain the local model parameters for each local model includes: For each local model, obtain the support set corresponding to that local model; The global model parameters are used as the initial model parameters of the local model, and based on the initial model parameters, the loss function corresponding to the support set on the local model is determined. The initial model parameters are updated based on the loss function to obtain the local model parameters of the local model.

[0008] In one possible implementation, for each local model, the initial model parameters are updated based on the loss function to obtain the local model parameters, including: according to Determine the first Local model parameters for each local model; in, Indicates the first Local model parameters of a local model Represents global model parameters. Indicates the learning rate. Represents global model parameters The partial derivative, Indicates support set In the The corresponding loss function on each local model This indicates that the global model parameters are used as the first... When initializing the model parameters for the local model, the first... The model function corresponding to each local model.

[0009] In one possible implementation, determining the meta-gradient corresponding to each local model based on the parameters of each local model includes: For each local model, obtain the query set corresponding to that local model; Based on the local model parameters of the local model, determine the loss function corresponding to the query set on the local model; Determine the gradient of the global model parameters, and based on the gradient of the global model parameters and the loss function corresponding to the query set on the local model, determine the meta-gradient of the local model.

[0010] In one possible implementation, the meta-gradient corresponding to the local model is determined based on the gradient of the global model parameters and the loss function corresponding to the query set on the local model, including: according to Determine the first Meta gradients corresponding to each local model; in, Indicates the first The meta-gradient corresponding to each local model Represents global model parameters The partial derivative, Indicates support set In the The corresponding loss function on each local model Indicates the first The local model parameters of each local model are At that time, the first The model function corresponding to each local model.

[0011] In one possible implementation, updating the global model parameters based on each meta-gradient includes: The gradients of each element are summed with weights, and the global model parameters are updated based on the results of the weighted summation to obtain the updated global model parameters.

[0012] In one possible implementation, the global model parameters are updated based on the weighted summation result to obtain the updated global model parameters, including: according to Determine the updated global model parameters; in, This represents the updated global model parameters. This represents the global model parameters before the update. Represents the meta-learning rate. Indicates the first The weight coefficients of each local model. Indicates the first The meta-gradient corresponding to each local model.

[0013] In one possible implementation, the prediction model is used for: The historical industrial load time series is decomposed to determine the trend component characteristics and seasonal component characteristics in the historical industrial load time series; Determine the correlation of the seasonal component features under different time delays, and determine the top K time delays with the highest correlation as the optimal time delays; K is an integer greater than or equal to 1; Based on the optimal time delay, the seasonal component features are weighted and aggregated to obtain enhanced seasonal component features; The trend component features and the enhanced seasonal component features are encoded to obtain the load prediction results.

[0014] In one possible implementation, the weighted aggregation of the seasonal component features based on the optimal time delay to obtain enhanced seasonal component features includes: according to Determine the enhanced seasonal component characteristics; in, This indicates the enhanced seasonal component characteristics. Indicates the first The normalized correlation weights corresponding to the optimal time delay Indicates a time-delay shift operation. Indicates seasonal characteristics, Indicates the first The optimal time delay Secondly, embodiments of the present invention provide an industrial load forecasting device, comprising: The acquisition module is used to acquire the historical industrial load time series of the industrial entities to be measured in the target area; The prediction module is used to input the historical industrial load time series into the prediction model corresponding to the industrial entity to be measured, and to obtain the industrial load prediction result of the industrial entity to be measured. The prediction model is a local model trained in a federated learning framework. The federated learning framework includes a global model corresponding to the target region and local models corresponding to each industrial entity in the target region. The global model and each local model are trained using a meta-gradient update method to obtain the trained local model.

[0015] Thirdly, embodiments of the present invention provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect or any possible implementation thereof.

[0016] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the first aspect or any possible implementation thereof.

[0017] Fifthly, embodiments of the present invention provide a computer program product, including a computer program that, when executed by a processor, implements the method described in the first aspect or any possible implementation thereof.

[0018] In this embodiment of the invention, each industrial entity collaborates on modeling through a federated learning framework. This indirectly utilizes complementary information from multi-source industrial load data within the region, overcoming the limitations of insufficient data samples from a single entity. This allows for more accurate capture of dynamic fluctuations in regional industrial load, such as peak-to-valley changes and capacity adjustments, significantly improving the accuracy of short-term load forecasting. Furthermore, the global model and each local model are trained using a meta-gradient update approach, enabling the models to learn shared prior knowledge across tasks and possessing strong transfer capabilities. When facing new scenarios with small sample sizes, such as new industrial entities starting production or adjusting production conditions within the target region, the model only requires a small amount of local data to quickly complete personalized adaptation, effectively solving the cold start problem in new scenarios and broadening the applicability of the method. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the implementation of the industrial load forecasting method provided in this embodiment of the invention. Figure 2 This is a schematic diagram of the industrial load forecasting device provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0020] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0021] Traditional statistical models (such as ARIMA) and single data-driven models (such as LSTM and Transformer) struggle to accurately capture the dynamic evolution of regional industrial loads, resulting in prediction accuracy that falls short of practical application requirements. Furthermore, the scale of historical regional industrial load data used for modeling is often limited, leading to insufficient model training and inadequate generalization ability. This results in a significant decline in the performance of existing prediction models, making them unable to quickly adapt to dynamic changes in load characteristics.

[0022] To adapt to dynamic changes in load characteristics and improve the accuracy of industrial load forecasting, this embodiment employs a federated learning framework for collaborative modeling among industrial entities. This indirectly utilizes complementary information from multi-source industrial load data within the region, overcoming the limitations of insufficient data samples from a single entity. This allows for more accurate capture of dynamic fluctuations in regional industrial load, such as peak-valley changes and capacity adjustments, significantly improving short-term load forecasting accuracy. Furthermore, a meta-gradient update approach is used to train both the global and local models, enabling them to learn shared prior knowledge across tasks and possessing strong transfer capabilities. When facing new scenarios with small sample sizes, such as new industrial entities starting production or adjusting production conditions within the target region, the model can quickly complete personalized adaptation with only a small amount of local data, effectively solving the cold start problem in new scenarios and broadening the applicability of the method.

[0023] See Figure 1 The document illustrates a flowchart of the industrial load forecasting method provided in an embodiment of the present invention, which is described in detail below: Step 101: Obtain the historical industrial load time series of the industrial entities to be tested in the target area.

[0024] The target area may contain multiple industrial entities. For any industrial entity to be measured, this embodiment can obtain the historical industrial load time series of that industrial entity. The historical industrial load time series may contain the industrial load values ​​corresponding to each time point within the historical time period.

[0025] Step 102: Input the historical industrial load time series into the prediction model corresponding to the industrial entity to be measured to obtain the industrial load prediction result of the industrial entity to be measured.

[0026] The prediction model is a local model trained in the federated learning framework. The federated learning framework includes a global model corresponding to the target region and local models corresponding to each industrial entity in the target region. The global model and each local model are trained using a meta-gradient update method to obtain the trained local model.

[0027] In this embodiment, historical industrial load time series can be input into the trained local model to obtain industrial load prediction results.

[0028] Current regional industrial load forecasting methods often employ independent modeling for a single industrial entity, training and forecasting models based on local historical load data from that individual entity. This approach heavily relies on sufficient high-quality historical data; however, due to insufficient data accumulation and limited data coverage for a single entity, the sample size used for modeling is often inadequate, leading to insufficient model training. In scenarios with frequent load pattern switching and high data noise, overfitting is likely, resulting in weak generalization ability and an inability to accurately represent the overall regional industrial load characteristics.

[0029] This embodiment constructs a federated learning framework. The framework includes a local meta-learning module (i.e., the local model) and a federated aggregation module (i.e., the global model), employing a closed-loop collaborative optimization process of "local adaptation - global aggregation." Each industrial sector or regional industrial cluster (i.e., the industrial entity) can act as an independent client, retaining local historical industry-level load data, meteorological variables (temperature, humidity), timestamp features (time period, weekday, production cycle node), and other raw data. The server is responsible for receiving model updates from the clients, performing global aggregation and model iteration, and distributing the updated global model to each client. Only model parameters are transmitted between the client and the server, enabling cross-domain and cross-regional collaborative optimization of data and models, fully leveraging the complementary value of multi-domain data. Specifically, the client runs the local model, and the server runs the global model.

[0030] To address the overall forecasting needs of regional industries, the client segmentation follows the principle of "industrial relevance + load homogeneity," grouping industrial entities with similar production processes and load fluctuation patterns within the region into the same client (i.e., the same local model), thereby improving the training efficiency and adaptability of the local model. Simultaneously, a regional load feature fusion module is added to the server side to integrate the meta-gradient information of each client, extract the common features of regional industrial load, and ensure that the global model can accurately represent the dynamic evolution of the overall regional industrial load.

[0031] Each global collaborative iteration must go through a complete closed-loop process of "local training - global aggregation - model deployment," which is the core requirement for the deep integration of federated collaborative modeling and meta-learning. After multiple such closed-loop iterations, the shared prior knowledge of the global model is continuously optimized, and the local models of each client are also fine-tuned through multiple rounds of meta-learning, gradually completing personalized adaptation, and finally obtaining a trained local model for short-term load forecasting of regional industries.

[0032] Existing collaborative modeling solutions do not focus on the core needs of the overall load of regional industries, do not design extraction mechanisms for common characteristics of regional industries, and only focus on the load forecast of a single industrial entity. They cannot effectively integrate the load complementarity information of multiple industrial entities in the region, and it is difficult to accurately characterize the dynamic patterns of peak and valley changes and capacity adjustments of the overall industrial load in the region. They cannot provide comprehensive and reliable load data support for regional power grid dispatching.

[0033] This embodiment adopts a "local accurate prediction + global weighted aggregation" model. Based on the local model, it outputs the load prediction results of the corresponding industrial entities. Combining the capacity ratio and load contribution of each industrial entity in the target area, the load prediction results of each local model are weighted and fused to obtain the overall industrial load prediction results of the target area. This not only ensures the accuracy of the prediction of each industrial entity, but also achieves a comprehensive representation of the overall industrial load of the target area, solving the problem that existing methods cannot take into account both local and overall prediction accuracy.

[0034] Existing cross-entity collaborative forecasting methods mostly adopt a centralized architecture, training a unified forecasting model by aggregating load data from multiple industrial entities within a region. While this can expand the sample size and extract common data features to some extent, alleviating the problem of data scarcity for a single entity, it lacks an adaptation mechanism designed for the heterogeneity of load characteristics among different industrial entities within the region. Different entities have different production processes and equipment operating conditions, resulting in varying load fluctuation patterns. Directly applying a unified model can easily lead to negative transfer phenomena, making it difficult to take into account the individualized load characteristics of each entity, thus limiting the accuracy of overall regional industrial load forecasting.

[0035] Considering that most federated learning load prediction schemes in related technologies adopt a "unified global model" and do not design an adaptation mechanism for the heterogeneity of load characteristics of different industrial entities in the region, they fail to achieve the dual goals of "personalized adaptation + global collaboration". Moreover, conventional schemes only complete "local training - parameter aggregation" and do not form a closed loop of "local adaptation - global aggregation". They also have the limitation that "parameter aggregation ends" and are prone to negative transfer problems, and cannot take into account both the common load patterns across entities and the personalized fluctuation characteristics of individual entities.

[0036] The federated learning framework in this embodiment includes a global model corresponding to each region and a local model corresponding to each entity, and achieves a closed loop of "local training - global aggregation" through meta-gradient updates. This mechanism of "personalized adaptation + global collaboration" can extract the common patterns of regional industrial loads while retaining the individual fluctuation characteristics of each industrial entity, avoiding the negative transfer problem caused by the traditional unified global model ignoring load heterogeneity.

[0037] Furthermore, this embodiment treats the load forecasting task of each industrial entity as an independent task. Through task-level training, the model learns to share prior knowledge across tasks, optimizes global model parameters, and enables the model to quickly adapt to new industrial fields, new regions, or new operating conditions with only a small amount of local data. It can also alleviate the negative migration problem caused by load heterogeneity.

[0038] This embodiment adopts a federated learning paradigm, clearly defining each industrial sector and regional industrial cluster as an independent client. The server is responsible for global model aggregation and iteration, forming a closed-loop collaborative process of "local training - parameter upload - global aggregation - model distribution." This enables cross-domain and cross-regional model collaborative optimization, fully utilizing the complementary value of multi-domain data and alleviating the problem of data scarcity in a single domain. Unlike existing federated learning load prediction schemes, this framework adopts a "task-level partitioning to adapt to heterogeneity" approach. Instead of using a unified global model, it uses a meta-learning "task-level training" logic, allowing each client to conduct local training based on its own load characteristics. This achieves the dual goals of "personalized adaptation + global collaboration," ensuring that the model can capture the common patterns of the overall regional industrial load while adapting to the individual fluctuation characteristics of each industrial sector.

[0039] Existing solutions do not clearly define the relationship between meta-learning and federated learning, which can easily lead to the cognitive bias of "equating the meta-learning process with the federated learning process". Furthermore, they do not deeply integrate the task-level training and meta-gradient feedback mechanism of meta-learning with the collaborative aggregation architecture of federated learning, thus failing to fully leverage the advantages of both and making it difficult to achieve the dual goals of rapid adaptation to small samples and cross-client collaboration.

[0040] This embodiment explicitly defines meta-learning as the core module of local training within the federated learning framework, meaning that the local training process of each client adopts the meta-learning mechanism. Federated learning serves as the collaborative aggregation carrier of meta-learning results—clients obtain meta-gradients through meta-learning, then upload them to the server via the federated learning architecture. The server updates the global model using a federated aggregation algorithm, and then distributes the updated global model back to the clients for use in the next round of meta-learning. Together, they achieve the dual goals of "cross-client collaboration + rapid adaptation to small samples." Compared to existing fusion schemes, this invention replaces traditional parameter aggregation with meta-gradient aggregation, ensuring that the global model can effectively integrate the shared prior knowledge of each client while preserving the personalized load characteristics of each client, thus avoiding negative transfer problems.

[0041] In this embodiment, the local and global models can be Autoformer models. In some embodiments, the prediction model is used to: decompose the historical industrial load time series to determine the trend component features and seasonal component features in the historical industrial load time series; determine the correlation of the seasonal component features under different time delays, and determine the top K time delays with the highest correlation as the optimal time delays; K is an integer greater than or equal to 1; perform weighted aggregation on the seasonal component features based on the optimal time delays to obtain enhanced seasonal component features; and encode the trend component features and enhanced seasonal component features to obtain the load prediction result.

[0042] The prediction model is the trained local model, also known as the Autoformer model. This model can analyze historical industrial load time series data. Decomposed into trend component features and seasonal component characteristics The decomposition process is as follows: (1) Trend components can be extracted using moving average operations: (2) In the formula, Indicating trend component characteristics middle The eigenvalue corresponding to time step 1. Indicates the moving average window size. Representing historical industrial load time series The industrial load value corresponding to the time.

[0043] Seasonal component characteristics are obtained by subtracting trend component characteristics from historical industrial load time series: (3) In the formula, Indicating seasonal component characteristics middle The eigenvalue corresponding to the given time.

[0044] The Autoformer model incorporates a self-attention mechanism. This mechanism captures periodic dependencies by calculating the correlation of seasonal components at different time delays. The correlation calculation is as follows: (4) in, Indicates a time delay of Time-related correlation The sequence length representing the seasonal component characteristics, Indicating seasonal component characteristics middle The eigenvalue corresponding to the given time.

[0045] Self-attention mechanisms are selected based on the highest relevance. Using a time delay (i.e., the optimal time delay), the seasonal components are weighted and aggregated to obtain enhanced seasonal component features: (5) In the formula, This indicates an enhanced seasonal component characteristic. Indicates the first The normalized correlation weights corresponding to the optimal time delay Indicates a time-delay shift operation. Indicates seasonal characteristics, Indicates the first The optimal time delay.

[0046] The core applications of the autocorrelation attention mechanism are reflected in three aspects: First, it accurately captures the periodic dependence of regional industrial load, providing core feature support for meta-learning to extract shared prior knowledge across industrial fields and regions, thus solving the pain point of insufficient prior knowledge in meta-learning scenarios with small samples; second, it reduces modeling complexity by focusing on highly correlated time delay terms, replacing the full pairwise calculation of traditional self-attention, reducing model computational overhead, and adapting to the need for rapid updates of small samples in the meta-learning inner loop; third, it strengthens the representation of temporal features, working in conjunction with the sequence decomposition mechanism of Autoformer to accurately characterize the periodic fluctuation pattern of load, improving the model's ability to model the non-stationary and strongly nonlinear features of regional industrial load, ensuring that the meta-gradient can accurately capture shared information across tasks, and avoiding negative transfer during the meta-learning process.

[0047] The Autoformer model encodes the trend component features and the enhanced seasonal component features to output load forecast results.

[0048] Existing methods do not have a dedicated modeling mechanism for the temporal characteristics of regional industrial loads, cannot effectively decouple the trend and seasonal components in the load sequence, are difficult to accurately capture the periodic dependence characteristics of the load, and have high modeling complexity, which cannot provide strong support for cross-scenario transfer learning, further limiting the improvement of prediction accuracy and scenario adaptability.

[0049] This embodiment uses the Autoformer model as the basic prediction model. Its sequence decomposition mechanism decomposes the time series of regional industrial loads into trend and seasonal components. The trend component is extracted through moving average operations (filtering high-frequency noise and short-term disturbances), while the remaining seasonal component focuses on periodic fluctuation characteristics (such as industry production cycles and seasonal production differences), achieving decoupled modeling of complex time series structures. An autocorrelation attention mechanism is introduced to replace traditional self-attention. By calculating the sequence correlation under different time delays, the length of industry-level load cycles (such as monthly production cycles and quarterly capacity fluctuation cycles) is automatically identified. This reduces computational complexity while accurately capturing the periodic dependence of loads, reducing modeling complexity, strengthening the representation of time series characteristics, and improving the model's ability to model the complex fluctuation patterns of regional industrial loads, further enhancing prediction accuracy.

[0050] Based on Autoformer sequence decomposition and autocorrelation attention mechanism, this paper optimizes the time series feature extraction process to address the non-stationarity, strong nonlinearity, and periodicity of regional industrial loads. Through adaptive adjustment of the moving average window, it accurately extracts the trend components (reflecting regional industrial production scale adjustments and long-term operating condition changes) and seasonal components (reflecting short-term features such as production cycles and seasonal fluctuations) of the load sequence. Through parameter optimization of the autocorrelation attention mechanism, it automatically identifies the core cycles of regional industrial loads (such as monthly production cycles and quarterly capacity fluctuation cycles), focusing on highly correlated time-delay terms. This reduces modeling complexity while enhancing the ability to capture load peak-valley changes and abrupt fluctuations, providing more accurate feature support for cross-task transfer of meta-learning.

[0051] In some embodiments, the training process for each local model is as follows: The system receives the global model parameters from the global model and trains each local model based on these parameters to obtain the local model parameters. Based on these local model parameters, it determines the meta-gradients for each local model. The system updates the global model parameters based on these meta-gradients and then jumps to the next step of training each local model based on the global model parameters to obtain the local model parameters. This process continues until a set number of training iterations is reached, at which point the trained global model and each local model are obtained.

[0052] In this embodiment, each local model receives the global model parameters from the global model and uses these global model parameters as its own initial model parameters for model training, thereby obtaining the local model parameters.

[0053] Specifically, for each local model, the support set corresponding to that local model is obtained; the global model parameters are used as the initial model parameters of that local model, and based on the initial model parameters, the loss function corresponding to the support set on that local model is determined; the initial model parameters are updated based on the loss function to obtain the local model parameters of that local model.

[0054] This embodiment can treat the industrial load forecasting task of each industrial entity as an independent task. For example, the first... The task of forecasting the industrial load of the first industrial entity, that is, the first... A local model can be represented as For the first A local model with its local dataset It can be divided into support sets and query set The support set consists of a small sample (15%-20% of the local dataset) used for rapid model adaptation to local load characteristics; the query set, which can comprise 80%-85% of the local dataset, is used to validate the adaptation and solve for the meta-gradient. Initial model parameters (i.e., global model parameters) As initial parameters with strong transferability, they are first rapidly updated within the task on the support set to simulate the model's adaptation process to new industrial fields and new regional data, so as to update and determine the local model parameters.

[0055] In this embodiment, the first The local model parameters of a local model can be represented as: (6) in, Indicates the first Local model parameters of a local model This represents global model parameters, distributed from the server to each client, and possesses the characteristic of sharing prior knowledge across tasks. The learning rate controls the step size of gradient updates and determines the magnitude of parameter adjustments in a single update. Represents global model parameters The partial derivatives are used to calculate the loss function with respect to the global model parameters. The gradient direction; Indicates support set In the The loss function corresponding to the local model measures the loss function of the th local model. The prediction error of a local model; This indicates that the global model parameters are used as the first... When initializing the model parameters for the local model, the first... The model function corresponding to each local model.

[0056] here, For the first Local models The support set is a small amount of labeled data used to quickly fine-tune model parameters, simulating a few-shot learning scenario.

[0057] Subsequently, the model performance loss is calculated on the query set, and the meta-gradient is solved to capture shared information across industrial sectors and regional tasks.

[0058] In some embodiments, for each local model, the query set corresponding to the local model is obtained; based on the local model parameters of the local model, the loss function corresponding to the query set on the local model is determined; then, the gradient of the global model parameters is determined, and based on the gradient of the global model parameters and the loss function corresponding to the query set on the local model, the meta-gradient corresponding to the local model is determined.

[0059] No. The formula for calculating the meta-gradient corresponding to each local model can be expressed as: (7) in, Indicates the first The meta-gradient corresponding to each local model Represents global model parameters The partial derivatives are used to calculate the loss function with respect to the global model parameters. The gradient direction; Indicates support set In the The loss function corresponding to the local model measures the loss function of the th local model. The prediction error of a local model; Indicates the first The local model parameters of each local model are At that time, the first The model function corresponding to each local model.

[0060] For the first Local models The query set is used to evaluate the first The generalization performance of the local model, compared with the support set Mutually exclusive.

[0061] The global model can aggregate the meta-gradients corresponding to each local model, and then iteratively update the global model parameters. In some embodiments, the meta-gradients can be weighted and summed separately, and the global model parameters can be updated based on the weighted summation result to obtain the updated global model parameters.

[0062] The updated global model parameters can be represented as: (8) in, This represents the updated global model parameters. This represents the global model parameters before the update. This represents the meta-learning rate, used to control the step size for updating global model parameters; Indicates the first The weight coefficients of each local model. Indicates the first The meta-gradient corresponding to each local model.

[0063] To address the need for overall regional industrial load forecasting, this embodiment introduces weighting coefficients during the federated aggregation process. Here, different weighting coefficients can be assigned based on the capacity share and load contribution of each industrial sector and cluster within the regional industry. This ensures the global model can effectively capture the core characteristics of regional industrial load, improving the overall forecasting accuracy. Simultaneously, the closed-loop iteration process is optimized, shortening the iteration cycle of "local training - global aggregation - model deployment," improving collaborative modeling efficiency, and adapting to the rapidly changing needs of regional industrial load.

[0064] In existing methods, the temporal feature extraction and meta-learning transfer mechanism are independent of each other and do not achieve deep collaboration. As a result, the temporal features captured by the Autoformer model cannot effectively support the cross-task transfer of meta-learning, and the rapid adaptation capability of meta-learning cannot fully combine the temporal patterns of regional industrial load. This further reduces the prediction accuracy and adaptation efficiency of the model in new scenarios with small samples.

[0065] This embodiment relies on the temporal features (trend components, seasonal components, and periodic dependence features) of regional industrial load extracted by the Autoformer model to optimize the task-level training process of meta-learning. The correlation of temporal features is used as an important weight in the meta-gradient calculation, so that when meta-learning learns shared prior knowledge across tasks, it prioritizes capturing the common temporal patterns of regional industrial load, further improving the model's ability to represent the overall regional industrial load. At the same time, it strengthens the ability to quickly adapt to new scenarios with small samples, and realizes deep synergy between temporal modeling and cross-scenario transfer.

[0066] After receiving the meta-gradients uploaded by each local model, the global model can use the FedAvg algorithm to perform weighted aggregation and iteratively update the global meta-model parameters, and then distribute the updated global model to each local model. The local models can conduct a new round of meta-learning training based on local industrial load data, forming a closed-loop collaborative optimization of "local training - global aggregation - model distribution".

[0067] The specific implementation steps for iteratively updating the global meta-model parameters after weighted aggregation using the FedAvg algorithm are as follows: 1. Data reception: The global model receives the meta-gradients uploaded by each local model, and also receives the local data volume of each local model (i.e., the amount of data in the local dataset), which is used to calculate the weighting coefficients; 2. Calculate weight coefficients: Calculate the corresponding weight coefficients based on the amount of local data for each local model. The larger the amount of data, the higher the weight coefficient, ensuring that the local model with more data contributes more to the global model update and avoiding the impact of abnormal gradients from local models with less data on global performance. 3. Meta-gradient aggregation: Multiply the meta-gradients of each local model by their corresponding weight coefficients, then sum all the product results to obtain the aggregated total meta-gradient. ; 4. Global model parameter update: The global model parameters are iteratively updated using the aggregated total meta-gradient based on the above formula (8) and the meta-learning rate to obtain new global model parameters; 5. Model Deployment: The updated global model parameters are deployed to all local models, completing a global aggregation and meta-gradient update, providing initial parameters for the next round of local meta-learning.

[0068] Each global collaborative iteration must go through a complete closed-loop process of "local training - global aggregation - model deployment," which is the core requirement for the deep integration of federated collaborative modeling and meta-learning. After multiple such closed-loop iterations, the shared prior knowledge of the global model is continuously optimized, and each local model is also fine-tuned through multiple rounds of meta-learning, gradually completing personalized adaptation, and finally obtaining a well-trained local model for industrial load forecasting.

[0069] Compared to existing fusion schemes, this invention replaces traditional parameter aggregation with meta-gradient aggregation, ensuring that the global model can effectively integrate the shared prior knowledge of each client, while retaining the personalized load characteristics of each client and avoiding negative transfer problems.

[0070] This embodiment treats load forecasting tasks for each industrial sector and region as independent tasks, dividing the local dataset into a support set and a query set. Task-level training allows the model to learn shared prior knowledge across tasks, optimizing the initial parameters of the global model. This enables the model to quickly adapt to new industrial sectors, regions, or operating conditions with only a small amount of local data, mitigating the negative transfer problem caused by load heterogeneity. The support set is used for rapid model adaptation to local load characteristics, while the query set is used to verify the adaptation effect and solve for the meta-gradient, achieving a closed loop of "local adaptation - global aggregation," breaking the limitation of conventional federated learning where "parameter aggregation ends." Considering the characteristics of regional industrial loads, the meta-learning training process focuses on mining shared prior knowledge across industrial sectors, focusing on the common periodic characteristics and dynamic evolution patterns of regional industrial loads, further enhancing the model's cross-scenario adaptability.

[0071] The effectiveness of the industrial load forecasting method provided in this embodiment is demonstrated below with a specific example.

[0072] (1) Data source Industrial load data for a specific region was acquired, covering a period of 1.5-3 years, with a total installed industrial capacity ranging from 128-381MW. Simultaneously, industrial load data under new capacity deployment scenarios in the region (data period of 1 month) was selected to validate the model's rapid adaptation capability in new scenarios with a small sample size. The prediction resolution was 15 minutes, and the prediction target was the active power load of the region's industries at the next moment. Input features included historical load sequences. The dataset was divided into training, validation, and test sets at a ratio of 70% / 15% / 15%, and preprocessed with outlier removal, missing value imputation, and Min-Max normalization before being used for model training and validation.

[0073] (2) Experimental setup and evaluation indicators Model training was performed in a unified computing environment. The prediction model is based on an Autoformer-based temporal prediction network, constructing a collaborative prediction framework that integrates federated learning and meta-learning. The model hyperparameters are configured according to Table 1 to ensure fairness and reproducibility in comparisons between different models. Specific parameter settings are shown in Table 1. Table 1

[0074] To eliminate the impact of regional differences in industrial load on model evaluation results and ensure fairness in comparing different models, mean absolute percentage error (MAPE), normalized root mean square error (NRMSE), and normalized mean absolute error (NMAE) are used as model evaluation indicators. The smaller the values ​​of these three indicators, the higher the model's predictive accuracy. The definitions of the three indicators are as follows: Mean Absolute Percentage Error (MAPE): ; Normalized root mean square error (NRMSE): ; Normalized Mean Absolute Error (NMAE): ; In the formula, N represents the number of samples in the test set. for The actual value of industrial load at any given time. for Industrial load forecast at any given time This represents the total installed industrial capacity of the region.

[0075] (3) Analysis of experimental results The industrial load forecasting method provided in this embodiment is compared with the forecasting results of the traditional independent forecasting method and the traditional federated learning method (FL) under different base models (Autoformer, Transformer, TimeMixer). The comparison results are shown in Table 2.

[0076] Table 2. Industrial Load Forecasting Error (%)

[0077] As shown in Table 2, the federated collaborative method reduces the average MAPE by 1.99 percentage points compared to independent training; for the newly connected user F6, it reduces the MAPE by 2.08 percentage points compared to traditional federated learning and by 2.63 percentage points compared to local small sample modeling; and it reduces the average MAPE by 1.21 percentage points compared to the MP strategy. Moreover, the prediction curve is closer to the actual load, and it shows stronger stability and adaptability during peak-valley transitions and load fluctuations.

[0078] The industrial load forecasting method provided in this embodiment, verified by a real industrial load dataset from a certain region, demonstrates significant technical advantages and practical application value. Specific beneficial effects are as follows: (1) Significantly improved prediction accuracy: Compared with traditional independent prediction methods, the average MAPE of this method is reduced by 1.99 percentage points. It can accurately capture the dynamic fluctuation characteristics of industrial load peak and valley changes and capacity adjustment in a certain region, and provide reliable industrial load data support for regional power grid dispatch and new energy consumption.

[0079] (2) Strong adaptability to new scenarios: The meta-learning mechanism gives the model strong transfer capability. In small sample new scenarios such as the layout of new industrial capacity in a certain region, the MAPE of this method is reduced by 2.63 percentage points compared with local small sample modeling. It can quickly adapt to new industrial production conditions and new capacity layout in the region, effectively solve the cold start problem of new scenarios, and broaden the scope of application.

[0080] (3) The efficiency of collaborative prediction is greatly improved: the federal collaborative architecture realizes regional collaborative modeling only through the interaction of model parameters, avoiding massive original data transmission and reducing transmission costs; the meta-learning mechanism reduces the cost of model retraining in new scenarios and greatly improves the overall efficiency of regional industrial load collaborative prediction.

[0081] (4) Strong applicability to regional industrial scenarios: This method is designed for the industrial load characteristics of a certain region. The model process is simple and the deployment cost is low. There is no need to preprocess the load characteristics of each industrial entity in the region. It can directly adapt to the overall industrial load forecasting needs of the region, providing a practical and implementable technical path to help the power system achieve refined operation and efficient allocation of industrial energy.

[0082] This embodiment aims to improve the accuracy of industrial load forecasting in the target area and enhance the model's adaptability to various scenarios, integrating federated learning, meta-learning, and Autoformer time series modeling techniques. The core key points of the solution are as follows: (1) Regional Industrial Load Federated Collaborative Modeling Architecture: Construct a federated learning framework of "local training-parameter upload-global aggregation" to meet the prediction needs of industrial load in the target region. Allow each industrial entity in the region to conduct local training as an independent client and achieve regional collaborative modeling only through model parameter interaction. Fully explore the complementary value of regional multi-source industrial load data and solve the problem of data scarcity of a single entity.

[0083] (2) Meta-learning adaptation mechanism for regional industrial load: Based on the Autoformer model, a meta-learning strategy is introduced to treat the load prediction task of each industrial entity in the region as an independent sub-task. The global model initialization parameters are optimized through task-level training and meta-gradient aggregation, so that the model can quickly adapt to new scenarios such as new industrial production conditions and new capacity layout in a certain region with only a small amount of local data, thus solving the negative transfer problem caused by the heterogeneity of industrial load characteristics in the region.

[0084] (3) Optimization method for time series modeling of regional industrial load: Autoformer is used as the basic prediction model. The time series of regional industrial load is decoupled into trend components and seasonal components through the sequence decomposition mechanism to achieve hierarchical modeling of complex time series structure. The autocorrelation attention mechanism is used to accurately capture the periodic dependence characteristics of regional industrial load and improve the model's ability to model the non-stationary and strongly nonlinear characteristics of regional industrial load.

[0085] (4) Closed-loop optimization process of federation and meta-learning: Design a closed-loop optimization process of “local meta-learning training - meta-gradient upload - global aggregation update - model distribution iteration”, which deeply integrates the time series modeling capability of Autoformer, the rapid adaptation capability of meta-learning and the collaborative modeling capability of federated learning. Through continuous closed-loop iteration, the model performance is optimized, and the accuracy and stability of short-term load forecasting of regional industries are improved.

[0086] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0087] The following are device embodiments of the present invention. For details not described in detail, please refer to the corresponding method embodiments described above.

[0088] Figure 2 A schematic diagram of the industrial load forecasting device provided in an embodiment of the present invention is shown. For ease of explanation, only the parts related to the embodiment of the present invention are shown, and are described in detail below: like Figure 2 As shown, the industrial load forecasting device 2 includes an acquisition module 21 and a forecasting module 22.

[0089] Module 21 is used to acquire the historical industrial load time series of the industrial entities to be measured in the target area; Prediction module 22 is used to input the historical industrial load time series into the prediction model corresponding to the industrial entity to be measured, and obtain the industrial load prediction result of the industrial entity to be measured. The prediction model is a local model trained in the federated learning framework. The federated learning framework includes a global model corresponding to the target region and local models corresponding to each industrial entity in the target region. The global model and each local model are trained using a meta-gradient update method to obtain the trained local model.

[0090] In one possible implementation, the training process for each local model includes: Receive the global model parameters of the global model, and train each local model based on the global model parameters to obtain the local model parameters of each local model. Based on the parameters of each local model, determine the meta-gradient corresponding to each local model; The global model parameters are updated based on the gradients of each meta-model. Then, based on the updated model parameters, the process jumps to train each local model based on the global model parameters to obtain the local model parameters of each local model. This process continues until the set number of training iterations is reached, at which point the trained global model and each local model are obtained.

[0091] In one possible implementation, the prediction module 22 is further configured to: For each local model, obtain the support set corresponding to that local model; The global model parameters are used as the initial model parameters for the local model, and based on the initial model parameters, the loss function corresponding to the support set on the local model is determined. The initial model parameters are updated based on the loss function to obtain the local model parameters.

[0092] In one possible implementation, prediction module 22 is specifically used for: according to Determine the first Local model parameters for each local model; in, Indicates the first Local model parameters of a local model Represents global model parameters. Indicates the learning rate. Represents global model parameters The partial derivative, Indicates support set In the The corresponding loss function on each local model This indicates that the global model parameters are used as the first... When initializing the model parameters for the local model, the first... The model function corresponding to each local model.

[0093] In one possible implementation, prediction module 22 is specifically used for: For each local model, obtain the query set corresponding to that local model; Based on the local model parameters of the local model, determine the loss function corresponding to the query set on the local model; Determine the gradient of the global model parameters, and based on the gradient of the global model parameters and the loss function corresponding to the query set on the local model, determine the meta-gradient of the local model.

[0094] In one possible implementation, prediction module 22 is specifically used for: according to Determine the first Meta gradients corresponding to each local model; in, Indicates the first The meta-gradient corresponding to each local model Represents global model parameters The partial derivative, Indicates support set In the The corresponding loss function on each local model Indicates the first The local model parameters of each local model are At that time, the first The model function corresponding to each local model.

[0095] In one possible implementation, prediction module 22 is specifically used for: The gradients of each element are summed in weights, and the global model parameters are updated based on the results of the weighted summation to obtain the updated global model parameters.

[0096] In one possible implementation, prediction module 22 is specifically used for: according to Determine the updated global model parameters; in, This represents the updated global model parameters. This represents the global model parameters before the update. Represents the meta-learning rate. Indicates the first The weight coefficients of each local model. Indicates the first The meta-gradient corresponding to each local model.

[0097] In one possible implementation, the prediction model is used for: The historical industrial load time series is decomposed to determine the trend component characteristics and seasonal component characteristics in the historical industrial load time series; Determine the correlation of seasonal component characteristics under different time delays, and identify the top K time delays with the highest correlation as the optimal time delays; K is an integer greater than or equal to 1. Weighted aggregation of seasonal component features based on optimal time delay yields enhanced seasonal component features. The trend component features and enhanced seasonal component features are encoded to obtain the load forecast results.

[0098] In one possible implementation, seasonal component features are weighted and aggregated based on the optimal time delay to obtain enhanced seasonal component features, including: according to Identify and enhance seasonal component characteristics; in, This indicates an enhanced seasonal component characteristic. Indicates the first The normalized correlation weights corresponding to the optimal time delay Indicates a time-delay shift operation. Indicates seasonal characteristics, Indicates the first The optimal time delay This device embodiment can be used in the above method embodiment, and its technical effects and implementation principles are the same as those of the above method embodiment, so they will not be repeated here.

[0099] Figure 3 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. For example... Figure 3 As shown, the electronic device 3 of this embodiment includes a processor 30 and a memory 31. The memory 31 stores a computer program 32. When the processor 30 executes the computer program 32, it implements the steps in the various method embodiments described above. Alternatively, when the processor 30 executes the computer program 32, it implements the functions of each module / unit in the various device embodiments described above.

[0100] For example, computer program 32 may be divided into one or more modules / units, which are stored in memory 31 and executed by processor 30 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 32 in electronic device 3.

[0101] Electronic device 3 may include, but is not limited to, processor 30 and memory 31. Those skilled in the art will understand that...Figure 3 This is merely an example of electronic device 3 and does not constitute a limitation on electronic device 3. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device 3 may also include input / output devices, network access devices, buses, etc.

[0102] The processor 30 can 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 general-purpose processor can be a microprocessor or any conventional processor.

[0103] The memory 31 can be an internal storage unit of the electronic device 3, such as a hard disk or memory of the electronic device 3. The memory 31 can also be an external storage device of the electronic device 3, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the electronic device 3. Furthermore, the memory 31 can include both internal and external storage units of the electronic device 3. The memory 31 is used to store the computer program 32 and other programs and data required by the electronic device 3. The memory 31 can also be used to temporarily store data that has been output or will be output.

[0104] For the sake of simplicity and clarity, only the above-described functional modules / units are used as examples. In practical applications, the functions described above can be assigned to different functional modules / units as needed. These modules / units can be implemented in hardware, software, or a combination of both.

[0105] This invention also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the methods described in the above-described method embodiments.

[0106] This invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the methods described in the above-described method embodiments.

[0107] Computer programs include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. Computer-readable media can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0108] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0109] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. An industrial load forecasting method, characterized in that, include: Obtain the historical industrial load time series of the industrial entities to be measured in the target area; The historical industrial load time series is input into the prediction model corresponding to the industrial entity to be measured to obtain the industrial load prediction result of the industrial entity to be measured. The prediction model is a local model trained in a federated learning framework. The federated learning framework includes a global model corresponding to the target region and local models corresponding to each industrial entity in the target region. The global model and each local model are trained using a meta-gradient update method to obtain the trained local model.

2. The industrial load forecasting method according to claim 1, characterized in that, The training process for each local model includes: Receive the global model parameters of the global model, and train each local model based on the global model parameters to obtain the local model parameters of each local model; Based on the parameters of each local model, determine the meta-gradient corresponding to each local model; The global model parameters are updated based on each meta-gradient, and the process jumps to the step of training each local model based on the updated model parameters to obtain the local model parameters of each local model. This process continues until a set number of training iterations are reached, at which point the trained global model and each local model are obtained.

3. The industrial load forecasting method according to claim 2, characterized in that, The step of training each local model based on the global model parameters to obtain the local model parameters for each local model includes: For each local model, obtain the support set corresponding to that local model; The global model parameters are used as the initial model parameters of the local model, and based on the initial model parameters, the loss function corresponding to the support set on the local model is determined. The initial model parameters are updated based on the loss function to obtain the local model parameters of the local model.

4. The industrial load forecasting method according to claim 3, characterized in that, For each local model, the initial model parameters are updated based on the loss function to obtain the local model parameters, including: according to Determine the first Local model parameters for each local model; in, Indicates the first Local model parameters of a local model Represents global model parameters. Indicates the learning rate. Represents global model parameters The partial derivative, Indicates support set In the The corresponding loss function on each local model This indicates that the global model parameters are used as the first... When initializing the model parameters for the local model, the first... The model function corresponding to each local model.

5. The industrial load forecasting method according to any one of claims 2-4, characterized in that, The process of determining the meta-gradient corresponding to each local model based on the parameters of each local model includes: For each local model, obtain the query set corresponding to that local model; Based on the local model parameters of the local model, determine the loss function corresponding to the query set on the local model; Determine the gradient of the global model parameters, and based on the gradient of the global model parameters and the loss function corresponding to the query set on the local model, determine the meta-gradient of the local model.

6. The industrial load forecasting method according to claim 5, characterized in that, Based on the gradients of the global model parameters and the loss function corresponding to the query set on the local model, the meta-gradient corresponding to the local model is determined, including: according to Determine the first Meta gradients corresponding to each local model; in, Indicates the first The meta-gradient corresponding to each local model Represents global model parameters The partial derivative, Indicates support set In the The corresponding loss function on each local model Indicates the first The local model parameters of each local model are At that time, the first The model function corresponding to each local model.

7. The industrial load forecasting method according to any one of claims 2-4, characterized in that, The step of updating the global model parameters based on each meta-gradient includes: The gradients of each element are summed with weights, and the global model parameters are updated based on the results of the weighted summation to obtain the updated global model parameters.

8. The industrial load forecasting method according to claim 7, characterized in that, The global model parameters are updated based on the weighted summation result to obtain the updated global model parameters, including: according to Determine the updated global model parameters; in, This represents the updated global model parameters. This represents the global model parameters before the update. Represents the meta-learning rate. Indicates the first The weight coefficients of each local model. Indicates the first The meta-gradient corresponding to each local model.

9. The industrial load forecasting method according to any one of claims 1-4, characterized in that, The prediction model is used for: The historical industrial load time series is decomposed to determine the trend component characteristics and seasonal component characteristics in the historical industrial load time series; The correlation of the seasonal component features under different time delays is determined, and the top K time delays with the highest correlation are determined as the optimal time delays; K is an integer greater than or equal to 1. Based on the optimal time delay, the seasonal component features are weighted and aggregated to obtain enhanced seasonal component features; The trend component features and the enhanced seasonal component features are encoded to obtain the load prediction results.

10. The industrial load forecasting method according to claim 9, characterized in that, The step of weighted aggregation of the seasonal component features based on the optimal time delay to obtain enhanced seasonal component features includes: according to Determine the enhanced seasonal component characteristics; in, This indicates the enhanced seasonal component characteristics. Indicates the first The normalized correlation weights corresponding to the optimal time delay Indicates a time-delay shift operation. Indicates seasonal characteristics, Indicates the first The optimal time delay.

Images