Methods for determining load forecasting models, load data forecasting methods and equipment
By preprocessing the load data of power grid nodes and updating it with a graph neural network, combined with adaptive routing strategies and differential privacy protection, the adaptability and accuracy issues of cross-regional power grid load forecasting are solved, achieving efficient and secure load forecasting.
Patent Information
- Application Number
- CN202611318373.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-27
- Publication Date
- 2026-11-17
AI Technical Summary
Existing cross-regional power grid load forecasting technologies are poorly adapted to handling load heterogeneity and dynamic topology evolution of the power grid, making it difficult to meet the complex and ever-changing operating environment of new power systems. Furthermore, existing solutions struggle to balance data privacy protection with forecast accuracy.
By collecting load data from power grid nodes, performing data preprocessing to obtain standardized input features, updating the adjacency matrix of the graph neural network, and obtaining a load prediction model based on an adaptive routing strategy, local retraining and differential privacy protection are achieved.
It enhances the load forecasting model's adaptability to regional load heterogeneity and changes in power grid structure, improves forecast accuracy and real-time performance, reduces communication burden, and ensures data privacy and security.
Smart Images

Figure CN123121506A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of artificial intelligence technology, and in particular to a method for determining a load forecasting model, a load data forecasting method, and an apparatus. Background Technology
[0002] With the profound evolution of new power systems, the power grid structure is becoming increasingly complex, exhibiting characteristics such as a high proportion of renewable energy integration, widespread distribution of distributed resources, and dynamic topology reconfiguration. In cross-regional power grid collaborative forecasting scenarios, distributed collaboration among multiple levels of nodes, including provincial grids, urban distribution networks, and industrial zones, and utilizing historical load data (such as electricity consumption, ambient temperature, and time-series characteristics) for model optimization and load sensing, has become a key technological trend for improving the intelligent operation and real-time decision-making capabilities of the power grid.
[0003] Currently, existing technologies typically employ a centralized graph neural network architecture for power grid topology feature mining. The process involves a central server collecting operational data from all nodes across the network (such as bus voltage and branch power), constructing a global static adjacency matrix and a node feature matrix, and then inputting these matrices into a graph neural network model. Through multi-layered information transmission and aggregation calculations, the spatial dependencies of the power network topology are captured.
[0004] However, the above methods are poorly adapted to handling cross-regional load heterogeneity and dynamic topology evolution of the power grid, and are difficult to adapt to the complex and ever-changing operating environment of new power systems. Summary of the Invention
[0005] This application provides a method for determining load forecasting models, a method for forecasting load data, and equipment to enhance the model's adaptability to cross-regional load heterogeneity and dynamic topology evolution of the power grid, thereby effectively meeting the high-precision forecasting requirements of the complex and ever-changing operating environment of the new power system.
[0006] In a first aspect, this application provides a method for determining a load forecasting model, comprising:
[0007] Collect load data from power grid nodes;
[0008] The load data is preprocessed to obtain standardized input features;
[0009] Based on the standardized input features, the adjacency matrix of the graph neural network is updated to obtain the updated graph neural network.
[0010] Based on the updated graph neural network, a load prediction model is obtained through an adaptive routing strategy.
[0011] In one possible embodiment, after obtaining the load forecasting model, the method further includes:
[0012] The average error of the load forecasting model under the statistical time window;
[0013] When the average error exceeds the error threshold, a drift trigger signal is generated, wherein the error threshold is dynamically adjusted according to the fluctuation characteristics of the load data;
[0014] In response to the drift trigger signal, the load prediction model is retrained locally to obtain the retrained load prediction model;
[0015] By using version control, the load forecasting model is replaced with a retrained load forecasting model.
[0016] In one possible embodiment, in response to a drift trigger signal, the load prediction model is locally retrained to obtain a retrained load prediction model, including:
[0017] In response to the drift trigger signal, retrieve load data from the local cache within a set time period in the past;
[0018] Based on load data within a set time period in the past, the load prediction model is retrained locally to obtain a retrained load prediction model.
[0019] In one possible embodiment, the adjacency matrix of the graph neural network is updated based on the standardized input features to obtain the updated graph neural network, including:
[0020] Based on standardized input characteristics, determine the power difference, feature similarity, and geographical distance of the load data;
[0021] The adjacency weights of the graph neural network are calculated based on power difference, feature similarity, and geographical distance.
[0022] The adjacency matrix is updated using adjacency weights to obtain the updated graph neural network.
[0023] In one possible embodiment, a load prediction model is obtained based on the updated graph neural network and an adaptive routing strategy, including:
[0024] Based on standardized input characteristics, determine the load pattern of power grid nodes;
[0025] Based on the load pattern, an adaptive routing strategy is used to switch the computation path of the updated graph neural network to obtain the load prediction model.
[0026] In one possible embodiment, based on load patterns, an adaptive routing strategy is used to switch the computational path of the updated graph neural network to obtain a load prediction model, including:
[0027] Based on the load pattern, the computation path of the updated graph neural network is switched through an adaptive routing strategy to obtain the initial prediction model;
[0028] The model parameters of the initial forecast model are weighted and aggregated to generate a load forecast model.
[0029] In one possible embodiment, the model parameters of the initial prediction model are weighted and aggregated to generate a load prediction model, including:
[0030] The model parameters of the initial prediction model are weighted and aggregated to obtain a new initial prediction model;
[0031] By using a differential privacy method, noise is injected into a new initial prediction model to obtain a load prediction model.
[0032] In one possible embodiment, noise is injected into a new initial prediction model using a differential privacy method to obtain a load prediction model, including:
[0033] By using the differential privacy method, Gaussian noise is injected into the model parameters of the new initial prediction model according to a preset noise distribution;
[0034] By using homomorphic encryption technology, the model parameters after noise injection are encrypted to obtain the load prediction model.
[0035] Secondly, this application provides a load data forecasting method, including:
[0036] Obtain real-time load data for power grid nodes;
[0037] Real-time load data is input into the target prediction model to obtain load prediction data. The target prediction model includes the load prediction model provided above.
[0038] Thirdly, this application provides an electronic device, including: a memory and a processor;
[0039] The memory stores the instructions that the computer executes;
[0040] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.
[0041] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed, are used to implement the first aspect and / or various possible embodiments of the first aspect.
[0042] Fifthly, this application provides a computer program product, including a computer program that, when executed, implements the first aspect and / or various possible implementations of the first aspect.
[0043] This application provides a method for determining a load forecasting model, a method for forecasting load data, and an apparatus, which relate to the field of artificial intelligence technology. The method includes: collecting load data from power grid nodes; preprocessing the load data to obtain standardized input features; updating the adjacency matrix of a graph neural network based on the standardized input features to obtain an updated graph neural network; and obtaining a load forecasting model based on the updated graph neural network through an adaptive routing strategy. This application improves the consistency and usability of input data by collecting load data from power grid nodes and preprocessing it to obtain standardized input features. Updating the adjacency matrix of the graph neural network based on the standardized input features allows the graph neural network model to better characterize the dynamic relationships between power grid nodes. Furthermore, obtaining a load forecasting model based on the updated graph neural network through an adaptive routing strategy enhances the load forecasting model's adaptability to regional load heterogeneity and changes in power grid structure, reduces the communication burden caused by unnecessary parameter interactions, and further improves the accuracy and real-time performance of cross-regional power grid load forecasting. Attached Figure Description
[0044] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.
[0045] Figure 1 A flowchart illustrating the method for determining the load forecasting model provided in the embodiments of this application. Figure 1 ;
[0046] Figure 2 A flowchart illustrating the load data prediction method provided in this application embodiment;
[0047] Figure 3 A schematic diagram of the structure of the device for determining the load forecasting model provided in this application;
[0048] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.
[0049] The accompanying drawings have illustrated specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to specific embodiments. Detailed Implementation
[0050] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.
[0051] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with relevant laws, regulations and standards, and corresponding operation entry points are provided for users to choose to authorize or refuse.
[0052] Grid load forecasting, as a core component of power system operation analysis and intelligent dispatching technology, serves as a bridge connecting power production, transmission, and consumption. In the context of the construction of new power systems, with the integration of a high proportion of new energy sources (such as wind power and photovoltaics) and a high proportion of power electronic equipment (characterized by "dual high" characteristics), the randomness, volatility, and uncertainty of the power grid are increasing exponentially. Traditional single-point, isolated load forecasting models can no longer meet the requirements for the safe and stable operation of modern power grids. Therefore, cross-regional grid collaborative forecasting scenarios based on wide-area interconnection have emerged, becoming a key technological support for ensuring the safe and economical operation of large power grids.
[0053] In current cross-regional power grid load forecasting practices, to achieve efficient integration and utilization of information from multiple regions, the industry typically adopts two mainstream technical approaches: centralized modeling and federated learning collaborative training. These two models differ in their emphasis on data interaction methods, privacy protection levels, and system architecture.
[0054] Centralized Modeling: This model adopts a bottom-up data aggregation strategy. In practice, each regional power grid node uploads historical load data, meteorological information, and related operating parameters collected in its region to a central node on the dispatch side or in the cloud via a secure and reliable communication network. The central node utilizes this massive global data, combined with the unique physical characteristics and environmental variables of each region, to perform unified feature engineering and train a unified machine learning / deep learning model. The advantage of this method is that the central node can obtain a panoramic data view, and the model can globally capture the complex spatiotemporal correlations of cross-regional loads. However, this also places extremely high demands on the storage and computing performance of the central node. At the same time, the large-scale transmission and centralized storage of raw data objectively increases the security risks and privacy leakage hazards faced by the data in the transmission links and at the central node.
[0055] Federated Learning Collaborative Training: As a distributed artificial intelligence paradigm that balances data efficiency and privacy, federated learning fundamentally changes the logic of data interaction. In this model, following the core principle of "data remains still, model moves," each regional power grid node strictly retains the raw data required for model training locally, using only local computing power to iteratively train the initial model. After training, each node does not upload any raw load data, but only uploads encrypted or anonymized model gradients, weights, and other parameters to the central side. The central side uses a specific aggregation algorithm (such as FedAvg) to comprehensively update the collected parameters from multiple sources, generating a globally optimized model, which is then distributed to each regional node for the next iteration. This mechanism effectively breaks down "data silos" while completely avoiding direct sharing of raw load data between regions. It achieves a joint improvement in the load forecasting capabilities of multiple regional power grids while strictly meeting the requirements of data security, power grid protection, and commercial confidentiality protection.
[0056] In summary, existing cross-regional load forecasting typically employs either centralized modeling or federated learning collaborative training. Centralized modeling generally aggregates load data from various regions to a central location, then combines node features to build a unified model; federated learning, on the other hand, involves each region training its model locally and uploading the parameters, with the central location handling the aggregation and updates, thus avoiding direct sharing of raw data.
[0057] However, centralized modeling relies on the centralized transmission of data across regions, which makes it difficult to meet the requirements of power grid data security management and privacy protection. Although federated learning reduces the risk of raw data leakage, existing solutions often focus on time series fitting and do not make sufficient use of the correlation between power grid nodes, making it difficult to fully reflect the coupling characteristics between nodes.
[0058] Furthermore, when there are significant differences in regional load types, or when factors such as line switching or equipment additions or removals cause changes in the power grid structure, existing models typically lack the ability to dynamically adjust topological relationships, easily leading to a decrease in prediction accuracy. At the same time, multiple rounds of parameter interaction can also impose a high communication burden, affecting the real-time performance and engineering availability of collaborative prediction.
[0059] In view of this, how to improve the model's adaptability to node relationships, regional load differences, and changes in power grid structure while protecting the data privacy of cross-regional power grids has become an urgent technical problem to be solved.
[0060] To address the aforementioned issues, this application provides a method for determining a load forecasting model. This method involves collecting load data from power grid nodes, preprocessing the load data to obtain standardized input features, updating the adjacency matrix of a graph neural network based on the standardized input features, and then obtaining a load forecasting model based on the updated graph neural network using an adaptive routing strategy.
[0061] In the edge-cloud collaborative architecture for cross-regional power grid collaborative forecasting, each regional node can complete local data collection and processing, while the central side undertakes collaborative analysis and model application support. Adopting this technical approach enhances the ability to characterize dynamic relationships and heterogeneous load patterns in the power grid without relying on centralized sharing of raw data, thus balancing forecasting accuracy, adaptability, and practical deployment requirements.
[0062] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.
[0063] Figure 1 A flowchart illustrating the method for determining the load forecasting model provided in the embodiments of this application. Figure 1 ,like Figure 1 As shown, the method includes:
[0064] S101: Collect load data of power grid nodes.
[0065] In this step, the power grid node serves as the object of load data collection and the basic unit for distributed modeling. It can correspond to provincial grid nodes, municipal distribution network nodes, park nodes, or industrial load aggregation nodes in cross-regional power grid collaborative forecasting scenarios. Each power grid node stores data related to its own operating status locally. The load data serves as the raw input for subsequent model construction and can include historical active load values recorded according to a preset sampling period. It can also collect operating information associated with load changes. The operating information includes at least ambient temperature and date type, with the date type used to represent any of the time attributes such as weekdays, weekends, and holidays.
[0066] In practice, each power grid node acquires load time-series data through local measurement terminals, electricity consumption information collection systems, or dispatch automation interfaces, and aligns the timestamps according to a unified time base to form a set of data to be processed. The sampling period can be set to 15 minutes, 30 minutes, or 1 hour. When the original sampling intervals of nodes in different areas are inconsistent, they are first converted to a unified interval through local resampling before being output.
[0067] To ensure the effectiveness of subsequent modeling, missing record identification, duplicate record removal, and abnormal time label correction are performed simultaneously during the data collection phase. Missing records can be filled by interpolation between adjacent times or by backfilling the corresponding times of the same type of day. For duplicate records, the record with the higher quality identifier under the condition of consistent timestamps is retained. Load values that obviously exceed the physical range are marked as abnormal according to the local historical quantile threshold and temporarily stored for later use.
[0068] For example, each power grid node can directly retrieve the load sequence of the past N consecutive days from its local historical database, or it can continuously receive the latest sampled values during real-time operation and accumulate them into training samples using a sliding window method. N can be set to 7 days, 30 days, or 90 days depending on the prediction period. After collection, each node outputs a data segment that corresponds one-to-one with its local time series, retaining the node identifier, geographical location identifier, and sampling time information for subsequent processing.
[0069] Based on the above analysis, it can be seen that the load data collected from the power grid nodes constitutes the starting point for the entire load forecasting model determination process. The original data is kept in the local environment of the nodes, so that subsequent processing can continue without centrally sharing the original data.
[0070] S102: Perform data preprocessing on the load data to obtain standardized input features.
[0071] In this step, data preprocessing is used to convert the collected load data into a unified input representation suitable for graph neural network computation. The standardized input features are the result of this conversion, used to simultaneously characterize the node's local load level, environmental impact, and date attribute.
[0072] In practice, the load data within a preset time window is first read for each power grid node, and adaptive normalization is performed based on local historical statistics. Normalization can employ a minimum-maximum scaling method, mapping the load value x at a given moment to... ,in and These are the minimum and maximum load values within the historical window of this node, respectively. A smoothing constant is set to prevent the denominator from being zero.
[0073] For temperature data, it is first aligned with the load sequence according to the sampling time, and then converted into numerical features using a method that is the same scale as the load or zero mean and unit variance. For date types, it is discretely encoded according to weekdays, weekends, public holidays, and special adjusted workdays. The encoding form can be a one-heat vector or an ordered numerical label, as long as it can stably represent the date attribute. Subsequently, the normalized load values, temperature features, and date type codes are concatenated in chronological order to form the feature vector of each node at each sampling time. The feature vectors of multiple consecutive times are further stacked to form the node input matrix.
[0074] To ensure input stability, outliers marked with tags are truncated or replaced before splicing. Outlier load values can be corrected based on the average of adjacent time periods, missing temperature values can be filled with the most recent valid observation, and missing date attributes are directly restored based on the timestamp.
[0075] For example, when predicting load for the next hour, 24 past sampling points can be selected as the input window. Each node then forms a 24×d input matrix, where d is the concatenated feature dimension. In addition to serving as the attribute input for graph neural network nodes, the standardized input features can also be used to characterize inter-node relationships and support subsequent structural updates.
[0076] In one possible embodiment, to enhance the expression of differences in load types in different regions, slope features, periodic location features, or moving average features can be further extracted from the load sequence and concatenated with the aforementioned normalized load values to form an extended input vector, but its basic structure remains as standardized input features organized in a uniform format.
[0077] Based on the above analysis, it can be seen that by converting the raw load data into standardized input features, a common representation suitable for cross-regional joint calculation can be established, so that subsequent structural updates and model calculations are based on a consistent data interface, and the direct interference of regional dimensional differences on parameter learning can be reduced.
[0078] S103: Update the adjacency matrix of the graph neural network based on the standardized input features to obtain the updated graph neural network.
[0079] In this step, the adjacency matrix of the graph neural network is used to characterize the connection relationship and connection strength between power grid nodes. The purpose of updating the adjacency matrix is to make the network structure no longer fixed to static connections, but to adjust the relationship between nodes in combination with changes in node characteristics.
[0080] In practical implementation, the degree of association between nodes can be determined based on standardized input features, and the adjacency matrix can be updated accordingly to reflect changes in the association relationships between nodes in the current scenario. After obtaining the new adjacency matrix, it can be further processed by adaptive graph neural network calculations to form a structure matrix suitable for node information propagation.
[0081] After loading the updated adjacency matrix, the propagation range of node information and the aggregation objects change accordingly, resulting in an updated graph neural network. The adjacency matrix can be adjusted and updated according to changes in the operating state to keep the network structure consistent with the actual operating state. The standardized input features in this step not only serve as node attribute inputs but also as the basis for updating the adjacency matrix; therefore, the adjacency matrix can better reflect the dynamic relationships between nodes.
[0082] Based on the above analysis, it can be seen that the updated graph neural network can more accurately express the relationship between nodes across regions. When there are large differences in regional load types or changes in local structure, the updated graph neural network can still establish an effective information propagation path around the current relevant nodes.
[0083] Therefore, it can be deduced that this step directly addresses the technical problem in existing solutions that it is difficult to fully utilize node coupling relationships and lacks dynamic adjustment capabilities. By updating the adjacency matrix based on standardized input features, the graph structure is kept in sync with the actual operating state, thereby providing a network foundation consistent with the current scenario for subsequent model acquisition.
[0084] S104: Based on the updated graph neural network, a load prediction model is obtained through an adaptive routing strategy.
[0085] The updated graph neural network provides a dynamic topology foundation in this step, while the adaptive routing strategy provides a differentiated computation method for heterogeneous load patterns. The combination of the two enables the load forecasting model to no longer use a fixed single path to process all samples.
[0086] Based on the above analysis, this step, by switching computational paths, enables the same graph structure to invoke different feature extraction mechanisms under different operating scenarios. This allows the final load forecasting model to balance the ability to characterize cross-regional correlations with the ability to adapt to local anomalies. When changes in the power grid structure lead to a redistribution of node relationships, the updated graph neural network first completes the structural adjustment; when regional load conditions change, adaptive routing then completes the computational path adjustment. Therefore, the model is responsive to both structural and pattern changes.
[0087] This application embodiment collects load data from power grid nodes and preprocesses the load data to obtain standardized input features, thereby improving the consistency and usability of the input data. By updating the adjacency matrix of the graph neural network based on the standardized input features, the graph neural network model can better characterize the dynamic relationships between power grid nodes. Furthermore, based on the updated graph neural network, an adaptive routing strategy is used to obtain a load prediction model, which enhances the load prediction model's adaptability to regional load heterogeneity and changes in power grid structure, reduces the communication burden caused by unnecessary parameter interactions, and further improves the accuracy and real-time performance of cross-regional power grid load prediction.
[0088] In one possible implementation, after obtaining the load forecasting model, the method further includes: the average error of the load forecasting model under a statistical time window; generating a drift trigger signal when the average error is higher than an error threshold, wherein the error threshold is dynamically adjusted according to the fluctuation characteristics of the load data; responding to the drift trigger signal, retraining the load forecasting model locally to obtain a retrained load forecasting model; and replacing the load forecasting model with the retrained load forecasting model through version control.
[0089] In its specific implementation, this embodiment of the application locally stores historical load data and recently sampled data corresponding to the current region. At the end of the time window, it calculates the prediction error for each moment within the window and then averages this prediction error to obtain the average error. When the average error exceeds the dynamically adjusted error threshold, a drift trigger signal is activated. This embodiment of the application then calls local training resources to retrain the load prediction model locally. The local retraining process can use the original model's network structure and parameter initialization method, or it can update only some parameters to reduce computational overhead, outputting a retrained load prediction model. Subsequently, the version control module records the original model version and the retrained version. After verification, it switches the currently used version to the retrained load prediction model, retaining the old version for rollback if necessary.
[0090] This approach continuously monitors prediction bias during model operation and adaptively adjusts the error threshold based on load fluctuations, ensuring that drift judgments match current load changes. With local retraining and version control working together, model updates can be completed locally in a closed loop, and the model switching process is traceable and reversible, thus maintaining the load forecasting model's adaptability and operational stability to dynamic load scenarios.
[0091] In one possible implementation, in response to a drift trigger signal, the load prediction model is locally retrained to obtain a retrained load prediction model, including: in response to a drift trigger signal, retrieving load data from a local cache within a set past time period; and based on the load data within the set past time period, locally retraining the load prediction model to obtain a retrained load prediction model.
[0092] Upon receiving a drift trigger signal, data samples within the corresponding time span are retrieved from the local cache. These samples are then formatted into an input format consistent with the original load forecasting model before being input into the local training unit for local retraining. The local training unit can be deployed on regional edge nodes or local computing nodes. It uses the preserved model structure to incrementally update parameters, or refits the load distribution after reinitializing some parameters, thus forming a retrained load forecasting model. During local retraining, the training epochs, learning rate, and batch size can be adjusted based on the error changes corresponding to the drift trigger signal, allowing the model parameters to more quickly adapt to the current load change characteristics.
[0093] The aforementioned local retraining method updates the model without relying on external data transmission and directly applies the updated parameters to subsequent load forecasting tasks, enabling the load forecasting model to remain responsive to recent load distribution changes. Since the retraining data is taken from recent load records in the local cache, it can promptly incorporate drift-related, periodic load changes into the load forecasting model's adjustment range, thereby improving the adaptability and forecasting stability of the retrained load forecasting model to the current regional load characteristics.
[0094] Based on the aforementioned embodiments, further, the adjacency matrix of the graph neural network is updated according to the standardized input features to obtain the updated graph neural network, including: determining the power difference, feature similarity, and geographical distance of the load data according to the standardized input features; calculating the adjacency weights of the graph neural network based on the power difference, feature similarity, and geographical distance; and updating the adjacency matrix through the adjacency weights to obtain the updated graph neural network.
[0095] In implementation, the normalized load sequence, meteorological correlation, time period identifier, and node coordinate information of each node can be extracted first from the standardized input features. Then, the power difference between adjacent nodes is calculated based on the normalized load sequence, and a similarity measurement is performed on the multi-dimensional feature vectors within the same time window to obtain feature similarity. At the same time, the geographical distance is calculated based on the latitude and longitude of the nodes or the location of the power grid connection. Subsequently, the power difference, feature similarity, and geographical distance are input into a weight mapping function to obtain the adjacency weights of the corresponding node pairs. The weight mapping function can be implemented by linear weighting, normalization fusion, or threshold constraint. Based on this, the edge connection values in the original adjacency matrix can be increased, decreased, or set to zero to form an updated adjacency matrix.
[0096] Through the above update method, the graph neural network can simultaneously incorporate load variation differences, feature consistency, and spatial correlation into the node connection modeling, making the updated graph neural network more closely match the actual coupling relationship between cross-regional load data, and providing more accurate adjacency structure support for subsequent load prediction modeling.
[0097] Based on the aforementioned embodiments, further, according to the updated graph neural network, a load forecasting model is obtained through an adaptive routing strategy, including: determining the load pattern of the power grid node based on standardized input features; and based on the load pattern, switching the computation path of the updated graph neural network through an adaptive routing strategy to obtain the load forecasting model.
[0098] In practical implementation, the updated graph neural network can be pre-configured with multiple computational branches, each corresponding to different graph convolution parameter sets or message passing methods. The adaptive routing strategy selects the computational branch corresponding to the current node based on the load pattern. When the load pattern represents a peak state, the graph neural network switches to a computational path for strengthening spatial correlation modeling; when the load pattern represents a fluctuating state, it switches to a computational path for highlighting temporal change responses; and when the load pattern represents an abnormal state, it switches to a computational path for enhancing abnormal fluctuation identification. Routing control can be achieved through routing weights output by the gating unit, with each routing weight corresponding one-to-one with the load pattern, thereby determining the current path for inference and outputting the load prediction model.
[0099] In this operating mode, standardized input features first identify the node load status, and then the load pattern triggers a switching of the computation path. This allows the updated graph neural network to adopt matching parameter branches and information aggregation methods for different load states, thereby forming a load prediction model adapted to the current grid node load characteristics. Because path selection is linked to load patterns, the load prediction model's ability to represent node load differences, fluctuations, and abnormal states is enhanced, and the correspondence between the predicted output and the actual operating state is closer.
[0100] Based on the aforementioned embodiments, further, based on the load pattern, the computation path of the updated graph neural network is switched through an adaptive routing strategy to obtain a load prediction model, including: based on the load pattern, the computation path of the updated graph neural network is switched through an adaptive routing strategy to obtain an initial prediction model; the model parameters of the initial prediction model are weighted and aggregated to generate a load prediction model.
[0101] In its implementation, the updated graph neural network receives a load pattern determined based on standardized input features and switches between different computational branches accordingly. Different computational branches correspond to feature extraction paths under different regional load characteristics or different topological correlation strengths. The routing control unit outputs routing instructions based on the load pattern, causing the graph neural network to activate the corresponding sub-network structure under the current pattern, thus obtaining the initial prediction model. Subsequently, the model parameters participating in the output calculation in the initial prediction model are weighted and summarized. This can be done by linearly combining similar parameters from multiple sources according to their weights, or by assigning different weights to parameters from different layers within the same model before fusion, thereby forming a new parameter set and generating the load prediction model.
[0102] This approach first forms an initial prediction model based on the load pattern switching calculation path, and then performs weighted aggregation on its model parameters. This allows the model output to retain both pattern adaptability and overall consistency after parameter fusion, making the load prediction model more closely reflect the load variation patterns of current power grid nodes. As a result, the model exhibits more stable prediction performance when facing regional differences and changes in operating conditions, and can maintain good collaborative modeling effects.
[0103] In one possible implementation, the model parameters of the initial forecasting model are weighted and aggregated to generate a load forecasting model, including: weighting and aggregating the model parameters of the initial forecasting model to obtain a new initial forecasting model; and injecting noise into the new initial forecasting model using a differential privacy method to obtain the load forecasting model.
[0104] In practical applications, the weighting coefficients can be calculated by the central side based on the data integrity of the participating nodes and the current training quality, or they can be generated by preset weighting rules. This application embodiment does not limit this.
[0105] Differential privacy methods are used to perturb the aggregated parameters. The perturbation can generate Gaussian noise based on a preset noise distribution and inject it into the corresponding parameters of the new initial prediction model according to a set privacy budget. This allows the parameter values to maintain predictive availability while possessing irreversible inference characteristics. Furthermore, the model parameters after injecting noise can be used as the final parameter output of the load prediction model and used for subsequent load prediction inference.
[0106] During the process, the central unit first performs weighted fusion of the initial prediction model parameters from each participant to obtain an updated unified parameter set. Then, based on differential privacy constraints, random perturbations are superimposed on this unified parameter set to form a load prediction model that can be directly deployed. This approach introduces a privacy protection mechanism after model parameter fusion, enabling the model to inherit multi-source training information while reducing the risk of inferring regional load information from parameters.
[0107] Through the above methods, the load forecasting model further obtains differential privacy protection after completing cross-regional parameter aggregation, which can enhance privacy security while maintaining the model's expressive power, and make the aggregated parameters more suitable for application in cross-regional collaborative forecasting scenarios.
[0108] In one possible implementation, noise is injected into a new initial prediction model using a differential privacy method to obtain a load prediction model. This includes: injecting Gaussian noise into the model parameters of the new initial prediction model according to a preset noise distribution using a differential privacy method; and encrypting the noise-injected model parameters using homomorphic encryption technology to obtain the load prediction model.
[0109] During noise injection, the parameters of each layer of the new initial prediction model are traversed item by item. Then, Gaussian noise consistent with the parameter dimensions is generated according to a preset noise distribution, and this noise is superimposed on the corresponding parameters to form the injected noise model parameters. Subsequently, the injected noise model parameters are encrypted using homomorphic encryption technology, ensuring that the encrypted parameters can still support subsequent aggregation, transmission, or retrieval processing in their ciphertext state. The homomorphic encryption algorithm used can be an algorithm system supporting additive homomorphism or finite multiplicative homomorphism. Its key is generated and distributed centrally, and the ciphertext parameters are stored and exchanged without exposing the plaintext content.
[0110] Through the above processing, the model parameters are injected with differential privacy noise and then encapsulated in ciphertext. This maintains strong privacy protection during parameter interaction, model deployment, and cross-domain transmission, while preserving the availability of model parameters. This allows the load forecasting model to continue to be used for forecasting tasks under a protected state. Since Gaussian noise and encryption processing are applied to the parameter perturbation and information encapsulation stages, respectively, sensitive information in the model training results can be protected twice, reducing the impact of parameter leakage on the data privacy of power grid nodes.
[0111] Taking a provincial power company as an example, the power grid nodes in its subordinate cities A, B, and C collaboratively predict the load demand for the next 24 hours. City A is mainly industrial load, city B is mainly residential load, and city C is a mixed load area. For data security, the raw load data of cities A, B, and C cannot be directly shared, but the models are collaboratively trained through a federated learning framework. At the same time, city A's topology expands due to the addition of a new substation, city B experiences partial line disconnection due to aging lines, and city C experiences abnormal load fluctuations due to extreme weather such as typhoons.
[0112] In one possible implementation, power grid nodes in cities A, B, and C first collect local load data and perform data preprocessing. This preprocessing includes cleaning outliers, filling in missing values, aligning to a uniform time granularity, and standardizing the processed features. The industrial load data range for city A is [500, 2000], the residential load data range for city B is [100, 800], and the mixed load data range for city C is [200, 1500]. Each node adaptively normalizes the load data based on local historical extreme values, using the following formula: ,in To prevent the division of minute values with a denominator of zero, the normalized load values, temperature, and date types are concatenated into a multidimensional feature vector. Temperature is normalized to [0,1], and date type uses One-Hot encoding.
[0113] In this example, the industrial load feature vector of City A includes One-Hot encoding for "weekdays" and "high-temperature weather," while the residential load feature vector of City B includes One-Hot encoding for "weekends" and "low-temperature weather." The mixed load feature vector of City C, combined with the load, temperature, and date type at the corresponding sampling time, constitutes the standardized input features. Based on the standardized input features, the power difference, feature similarity, and geographical distance of the load data are determined. The power difference reflects the difference in load power variation between different grid nodes, the feature similarity reflects the proximity of different nodes in terms of load curves, periodicity, or statistical attributes, and the geographical distance reflects the spatial distribution relationship of nodes. The adjacency weights of the graph neural network are calculated by combining the power difference, feature similarity, and geographical distance. ,in, Indicates power difference, Indicates feature similarity. This represents geographical distance. Next, the adjacency matrix is updated using adjacency weights, resulting in an updated graph neural network.
[0114] Based on the updated graph neural network, the load pattern of the power grid nodes is determined according to the standardized input features. The load pattern is used to characterize the load change status, peak-valley characteristics, or time-series operation characteristics of the nodes in the current time period. The standardized feature vector X of each node is input into the adaptive graph neural network. The multi-mode routing mechanism dynamically switches between computation paths such as GCN, STConv, and LSTM according to the real-time load pattern, and executes the adaptive routing strategy in combination with the load pattern. The computation path of the updated graph neural network is switched according to different load patterns, so that the input data is processed in the corresponding network branch, message passing channel, or aggregation path to obtain the initial prediction model. Specifically, the softmax function is used to score the standardized feature vector under different computation paths, and the computation path with the highest score is selected. The adaptive routing strategy is executed in combination with the load pattern, and the computation path of the updated graph neural network is switched according to different load patterns, so that the input data is processed in the corresponding network branch, message passing channel, or aggregation path to obtain the initial prediction model.
[0115] When industrial load peaks occur in City A, the STConv module is prioritized to handle spatial dependencies. When residential load fluctuations occur in City B, the computation path corresponding to the residential load pattern is switched. During periods of abnormal load caused by typhoons in City C, the LSTM module is used to capture time series fluctuations. When a new substation is added in City A, the weights of the corresponding nodes in the adjacency matrix are dynamically adjusted to 0 to avoid erroneous connections interfering with the model, and the connections between the new nodes and other nodes are expanded in the graph structure. When some lines are disconnected in City B due to aging lines, the corresponding connection weights are adjusted synchronously. When City C is affected by typhoons, the node feature representations are dynamically updated according to changes in abnormal load patterns.
[0116] Subsequently, the model parameters of the initial prediction model are weighted and aggregated. Aggregation is first performed according to the weights of different computation paths, the contribution of different nodes, or the update results of different parameters, resulting in a new initial prediction model. The server then adjusts the data quality based on cities A, B, and C. With model accuracy Calculate dynamic weights and aggregate gradients City A's gradient weight ratio increases due to its high data quality and excellent model accuracy; City B's weight automatically decreases due to data anomalies and lower data quality caused by equipment failure; City C's parameter updates participate in aggregation based on their corresponding quality and accuracy.
[0117] After aggregation, Gaussian noise is injected into the model parameters of the new initial prediction model using a differential privacy method based on a preset noise distribution. Homomorphic encryption is then used to encrypt the noise-injected model parameters. The client uploads gradients using a formula... Where Noise corresponds to differential privacy injection noise, and Enc represents homomorphic encryption (PHE) processing, based on which the load prediction model is obtained.
[0118] In one possible implementation, after the load forecasting model is put into operation, the server and client continuously calculate the average error of the load forecasting model within a time window. The time window is updated by sliding according to a continuous sampling period, and a 24-hour error window is maintained. The average error is calculated based on the deviation between the predicted results and the actual load data, while also incorporating historical baseline errors. The comparison is performed. The error threshold is dynamically adjusted based on the fluctuation characteristics of the load data, and can be adjusted according to the load fluctuation rate. With average load according to Calculations were performed; the error threshold was increased when load fluctuations increased and decreased when load fluctuations decreased. The load fluctuation rate was affected by the typhoon in City C. During a sudden increase, the threshold The threshold is automatically increased; when the load in City B is stable, the threshold is decreased; when a new substation is added in City A, causing a topology change, the error window may suddenly increase. When the average error exceeds the current error threshold, a drift trigger signal is generated.
[0119] In response to the drift trigger signal, clients in cities A, B, and C retrieve load data from their local cache within a specified past time period (24 hours of new data). Based on this past load data, they retrain the load prediction model locally. During retraining, the aforementioned normalization and feature concatenation logic is reused. Outlier cleaning, missing value completion, unified time granularity alignment, and standardized input features are used again for model training. After initializing the global model, supervised training is performed for 10 epochs, with the learning rate set to 0.1 times that of the global training model.
[0120] After a new substation was added in City A, local retraining adjusted the adjacency matrix weights using 24-hour new data. In City C, after a typhoon caused load anomalies, retraining optimized the LSTM module parameters. City B simultaneously participated in retraining based on new data after local load stabilized or equipment failures were repaired. The latest 2-hour data after the retraining window was used as the validation set to evaluate the MAE or RMSE of the retrained load forecasting model. Retraining was only confirmed as complete after passing validation; otherwise, the training window needed to be expanded. After retraining, version control replaced the currently running load forecasting model with the retrained model, and the retrained model replaced the current local model. Simultaneously, the baseline value of the error monitoring window was updated.
[0121] Furthermore, embodiments of this application also provide a method for predicting load data. Figure 2 This is a flowchart illustrating the load data forecasting method provided in an embodiment of this application. Figure 2 As shown, the method includes: acquiring real-time load data of power grid nodes; inputting the real-time load data into a target prediction model to obtain load prediction data, wherein the target prediction model is the load prediction model provided above.
[0122] In this embodiment, by acquiring real-time load data of power grid nodes, the load changes at the current moment can be mapped to the model input in a timely manner, making the prediction process directly oriented towards the online operating status. After inputting the real-time load data into the target prediction model, the model's ability to model node relationships, regional differences, and changes in power grid structure can be utilized to predict the current load evolution trend, thereby outputting load prediction data that better reflects the actual operating characteristics of cross-regional power grids. Since the target prediction model can complete collaborative modeling while protecting the privacy of the original data, this prediction method balances data security and prediction accuracy, thereby improving the real-time performance and applicability of short-term load forecasting, scheduling decisions, and supply-demand balance analysis.
[0123] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.
[0124] Figure 3 A schematic diagram of the structure of the device for determining the load forecasting model provided in this application is shown below. Figure 3 As shown, the load forecasting model determination device 300 provided in this embodiment includes:
[0125] The data acquisition module 301 is used to acquire load data of power grid nodes;
[0126] Module 302 is used to preprocess the load data to obtain standardized input features.
[0127] Module 302 is obtained, which is used to update the adjacency matrix of the graph neural network based on the standardized input features, and obtain the updated graph neural network.
[0128] Module 302 is also used to obtain a load prediction model based on the updated graph neural network and through an adaptive routing strategy.
[0129] In one possible embodiment, a retraining module is also included, which is specifically used for:
[0130] The average error of the load forecasting model under the statistical time window;
[0131] When the average error exceeds the error threshold, a drift trigger signal is generated, wherein the error threshold is dynamically adjusted according to the fluctuation characteristics of the load data;
[0132] In response to the drift trigger signal, the load prediction model is retrained locally to obtain the retrained load prediction model;
[0133] By using version control, the load forecasting model is replaced with a retrained load forecasting model.
[0134] In one possible embodiment, the retraining module is specifically used for:
[0135] In response to the drift trigger signal, retrieve load data from the local cache within a set time period in the past;
[0136] Based on load data within a set time period in the past, the load prediction model is retrained locally to obtain a retrained load prediction model.
[0137] In one possible embodiment, module 302 is specifically used for:
[0138] Based on standardized input characteristics, determine the power difference, feature similarity, and geographical distance of the load data;
[0139] The adjacency weights of the graph neural network are calculated based on power difference, feature similarity, and geographical distance.
[0140] The adjacency matrix is updated using adjacency weights to obtain the updated graph neural network.
[0141] In one possible embodiment, module 302 is specifically used for:
[0142] Based on standardized input characteristics, determine the load pattern of power grid nodes;
[0143] Based on the load pattern, an adaptive routing strategy is used to switch the computation path of the updated graph neural network to obtain the load prediction model.
[0144] In one possible embodiment, module 302 is specifically used for:
[0145] Based on the load pattern, the computation path of the updated graph neural network is switched through an adaptive routing strategy to obtain the initial prediction model;
[0146] The model parameters of the initial forecast model are weighted and aggregated to generate a load forecast model.
[0147] In one possible embodiment, module 302 is specifically used for:
[0148] The model parameters of the initial prediction model are weighted and aggregated to obtain a new initial prediction model;
[0149] By using a differential privacy method, noise is injected into a new initial prediction model to obtain a load prediction model.
[0150] In one possible embodiment, module 302 is specifically used for:
[0151] By using the differential privacy method, Gaussian noise is injected into the model parameters of the new initial prediction model according to a preset noise distribution;
[0152] By using homomorphic encryption technology, the model parameters after noise injection are encrypted to obtain the load prediction model.
[0153] The load forecasting model determination device provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.
[0154] It should be noted that the division of the various modules in the above device is merely a logical functional division. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. Furthermore, these modules can be implemented entirely in software via processing element calls; they can be fully implemented in hardware; or some modules can be implemented by processing element calls to software, while others are implemented in hardware. For example, a processing module can be a separate processing element, or it can be integrated into a chip within the device. Alternatively, it can be stored as program code in the device's memory, and its functions can be called and executed by a processing element. The implementation of other modules is similar. Moreover, these modules can be fully or partially integrated together, or they can be implemented independently. The processing element here can be an integrated circuit with signal processing capabilities. During implementation, each step of the above method or each of the above modules can be completed through integrated logic circuits in the hardware of the processor element or through software instructions.
[0155] For example, these modules can be one or more integrated circuits configured to implement the above methods, such as one or more Application Specific Integrated Circuits (ASICs), one or more Digital Signal Processors (DSPs), or one or more Field Programmable Gate Arrays (FPGAs). As another example, when a module is implemented using processing element scheduler code, the processing element can be a general-purpose processor, such as a Central Processing Unit (CPU) or other processor capable of calling program code. Furthermore, these modules can be integrated together as a System-On-a-Chip (SOC).
[0156] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 4 As shown, the electronic device 400 provided in this application embodiment may include: a processor 401, and a memory 402 communicatively connected to the processor, wherein:
[0157] The memory stores the instructions that the computer executes;
[0158] The processor executes computer execution instructions stored in memory to implement the method described in the foregoing method embodiments.
[0159] It should be understood that processor 401 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the application can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor. Memory 402 may include high-speed random access memory (RAM), and may also include non-volatile memory (NVM), such as at least one disk storage device, or a USB flash drive, external hard drive, read-only memory, disk, or optical disc, etc.
[0160] Optionally, the electronic device 400 may also include a communication interface 403. In specific implementations, if the communication interface 403, memory 402, and processor 401 are implemented independently, they can be interconnected via a bus to complete communication. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc., but this does not imply that there is only one bus or one type of bus.
[0161] Optionally, in a specific implementation, if the communication interface 403, memory 402 and processor 401 are integrated on a single chip, then the communication interface 403, memory 402 and processor 401 can communicate through an internal interface.
[0162] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed, are used to implement the methods described in any of the foregoing embodiments.
[0163] It is understood that the computer-readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read Only Memory (PROM), Read Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.
[0164] An exemplary computer-readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the computer-readable storage medium. Of course, the computer-readable storage medium can also be a component of the processor. The processor and the computer-readable storage medium can reside in an ASIC. Alternatively, the processor and the computer-readable storage medium can exist as discrete components in an electronic device.
[0165] The integrated modules implemented as software functional modules described above can be stored in a computer-readable storage medium. These software functional modules, stored in a computer-readable storage medium, include several instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods described in the various embodiments of this application.
[0166] This application also provides a computer program product, including a computer program that, when executed, implements the method described in any of the foregoing embodiments.
[0167] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.
[0168] It should be further noted that although the steps in the flowchart are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.
[0169] In the above embodiments, the descriptions of each embodiment have their own emphasis. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.
[0170] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.
[0171] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.
Claims
1. A method for determining a load forecasting model, characterized in that, include: Collect load data from power grid nodes; The load data is preprocessed to obtain standardized input features; Based on the standardized input features, the adjacency matrix of the graph neural network is updated to obtain the updated graph neural network. Based on the updated graph neural network, a load prediction model is obtained through an adaptive routing strategy.
2. The method according to claim 1, characterized in that, After obtaining the load forecasting model, the following is also included: The average error of the load forecasting model under the statistical time window; When the average error exceeds an error threshold, a drift trigger signal is generated, wherein the error threshold is dynamically adjusted based on the fluctuation characteristics of the load data; In response to the drift trigger signal, the load prediction model is retrained locally to obtain a retrained load prediction model; The load prediction model was replaced with a retrained load prediction model through version control.
3. The method according to claim 2, characterized in that, In response to the drift trigger signal, the load prediction model is locally retrained to obtain a retrained load prediction model, including: In response to the drift trigger signal, load data within a set time period in the past is retrieved from the local cache; Based on load data within a set time period in the past, the load prediction model is retrained locally to obtain a retrained load prediction model.
4. The method according to any one of claims 1 to 3, characterized in that, The step of updating the adjacency matrix of the graph neural network based on the standardized input features to obtain the updated graph neural network includes: Based on the standardized input features, the power difference, feature similarity, and geographical distance of the load data are determined; The adjacency weights of the graph neural network are calculated based on the power difference, the feature similarity, and the geographical distance. The adjacency matrix is updated using the adjacency weights to obtain the updated graph neural network.
5. The method according to any one of claims 1 to 3, characterized in that, The process of obtaining a load prediction model based on the updated graph neural network and an adaptive routing strategy includes: Based on the standardized input characteristics, the load pattern of the power grid node is determined; Based on the load pattern, the load prediction model is obtained by switching the computation path of the updated graph neural network through the adaptive routing strategy.
6. The method according to claim 5, characterized in that, The process of obtaining the load prediction model by switching the computation path of the updated graph neural network based on the load pattern and through the adaptive routing strategy includes: Based on the load pattern, the computation path of the updated graph neural network is switched using the adaptive routing strategy to obtain the initial prediction model. The load forecasting model is generated by weighted aggregation of the model parameters of the initial forecasting model.
7. The method according to claim 6, characterized in that, The step of weighted aggregation of the model parameters of the initial prediction model to generate the load prediction model includes: The model parameters of the initial prediction model are weighted and aggregated to obtain a new initial prediction model; The load prediction model is obtained by injecting noise into the new initial prediction model using a differential privacy method.
8. The method according to claim 7, characterized in that, The step of injecting noise into the new initial prediction model using a differential privacy method to obtain the load prediction model includes: The differential privacy method injects Gaussian noise into the model parameters of the new initial prediction model according to a preset noise distribution. The load prediction model is obtained by encrypting the model parameters after injecting noise using homomorphic encryption technology.
9. A method for predicting load data, characterized in that, include: Obtain real-time load data for power grid nodes; The real-time load data is input into the target prediction model to obtain load prediction data, wherein the target prediction model is the load prediction model as described in any one of claims 1 to 8.
10. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-8 or 9.
