Load prediction method, device and storage medium

Abstract

The application provides a load prediction method and device and a storage medium, and relates to the technical field of power grids. The method comprises the following steps: acquiring a node topology graph and historical load data corresponding to a power grid, wherein the node topology graph represents the correlation between electrical nodes in the power grid; based on a graph attention network, extracting load space features of the power grid according to the node topology graph; based on a long short-term memory network model and a Transformer model, performing hybrid time sequence extraction on the historical load data of the power grid to obtain load time sequence features of the power grid; and predicting the load of the power grid according to the load space features and the load time sequence features. The application can improve the load prediction accuracy of the power grid.

Description

Technical Field

[0001] This application relates to the field of power grid technology, and in particular to a load forecasting method, device and storage medium. Background Technology

[0002] Grid load forecasting is a core component of the safe operation of smart grids. Its main challenge lies in handling the complex correlations between multi-source heterogeneous data, such as historical load, weather, holiday, and user behavior data. Traditional methods for obtaining correlations between multi-source heterogeneous data mainly fall into two categories: the first is statistical models, such as Autoregressive Integrated Moving Average (ARIMA) and exponential smoothing models. This approach relies on linear assumptions and struggles to capture nonlinear dynamics. The second category is shallow machine learning models, such as Support Vector Machine (SVM) and backpropagation neural networks. This approach requires manual feature design, has limited ability to fit high-dimensional data, and cannot effectively model spatiotemporal coupling relationships.

[0003] With the widespread adoption of smart meters, power grid loads exhibit characteristics such as high dimensionality, strong nonlinearity, and spatiotemporal correlation. For example, a 1°C increase in ambient temperature can lead to a significant surge in air conditioning load, while regional power grid topology (such as node electrical distances) amplifies the spatial transmission effects of load fluctuations. Limitations of traditional methods in feature engineering and dynamic adaptability result in poor power grid load forecasting performance. Summary of the Invention

[0004] This application provides a load forecasting method, device, and storage medium to improve the accuracy of power grid load forecasting.

[0005] In a first aspect, embodiments of this application provide a load forecasting method, including:

[0006] Obtain the node topology diagram and historical load data corresponding to the power grid. The node topology diagram represents the relationship between electrical nodes in the power grid.

[0007] Based on graph attention networks, load spatial characteristics of the power grid are extracted according to the node topology graph;

[0008] Based on the Long Short-Term Memory network model and the Transformer model, a hybrid time-series extraction is performed on the historical load data of the power grid to obtain the load time-series characteristics of the power grid.

[0009] The load of the power grid is predicted based on the spatial and temporal characteristics of the load.

[0010] In one possible implementation, based on a graph attention network, the load spatial characteristics of the power grid are extracted according to the node topology graph, including:

[0011] The strength of the association between nodes and their neighboring nodes in the node topology graph;

[0012] Normalize the association strength to obtain the normalized attention weight of the node;

[0013] Based on the normalized attention weights, a multi-head attention weight mechanism is used to enhance the feature extraction of nodes and obtain the node features corresponding to the nodes.

[0014] Based on graph attention networks, hierarchical spatial feature extraction is performed on the node topology graph according to node characteristics to obtain load spatial features.

[0015] In one possible implementation, based on a graph attention network, hierarchical spatial feature extraction is performed on the node topology graph according to node characteristics to obtain load spatial features, including:

[0016] Based on graph attention networks, local topological features of the node topology graph are extracted within a first radius according to node characteristics.

[0017] Based on graph attention networks, regional topological features of the node topological graph are extracted within a second radius range according to local topological features, wherein the second radius range is larger than the first radius range;

[0018] Based on graph attention networks, global average pooling is performed on regional topological features to obtain load space features.

[0019] In one possible implementation, a hybrid time-series extraction is performed on historical load data of the power grid based on a long short-term memory network model and a Transformer model to obtain the load time-series characteristics of the power grid, including:

[0020] Based on historical load data, the first temporal features of the historical load data are extracted using a long short-term memory network model.

[0021] Based on historical load data, the second time-series feature of the historical load data is extracted using the Transformer model, which employs trigonometric function position encoding.

[0022] The first time-series feature and the second time-series feature are concatenated to obtain the combined feature;

[0023] The combined features are fused using the first fully connected layer to obtain the initial fused features. The first fully connected layer uses the ReLU activation function.

[0024] The initial fusion features are fused using a second fully connected layer to obtain load time series features. The second fully connected layer uses the LeakyReLU activation function.

[0025] In one possible implementation, it also includes:

[0026] Based on reinforcement learning mechanisms, optimize the long short-term memory network model and the Transformer model;

[0027] The state space of the reinforcement learning mechanism includes model hyperparameters, prediction error distribution, and load fluctuation index; the action space includes fine-tuning operations, number of units in the long short-term memory network model, structure switching, and parameter reset.

[0028] The reward function is ,in, and The preset weighting coefficients, The mean absolute percentage error, For training time, As a stability reward, The number of days of continuous operation without reconfiguration.

[0029] In one possible implementation, the Long Short-Term Memory network model and the Transformer model are optimized based on a reinforcement learning mechanism, including:

[0030] Based on the proximal policy optimization algorithm and online learning mechanism, the optimized parameters of the Long Short-Term Memory Network Model and the Transformer Model are determined. The online learning mechanism includes fault injection, online training, and setting safety constraints.

[0031] The Long Short-Term Memory (LSTM) network model and the Transformer model are optimized based on the optimization parameters.

[0032] In one possible implementation, load forecasting of the power grid is performed based on load spatial characteristics and load temporal characteristics, including:

[0033] The first comprehensive feature is obtained by concatenating the load spatial characteristics and the load temporal characteristics;

[0034] The weight matrix is ​​used to transform the comprehensive features to obtain the second comprehensive feature;

[0035] Based on the second comprehensive feature, the ReLU activation function is used to generate the fusion feature;

[0036] Based on the multilayer perceptron (MLP), the load of the power grid is predicted according to the fusion characteristics and the meteorological correction factor, where the meteorological correction factor represents the impact of meteorological factors on the load.

[0037] In one possible implementation, it also includes:

[0038] Based on the load forecast results, load data mining applications are carried out, including abnormal load identification, attribution of influencing factors, and scenario simulation.

[0039] Secondly, embodiments of this application provide a load forecasting device, comprising:

[0040] The acquisition module is used to acquire the node topology diagram and historical load data corresponding to the power grid. The node topology diagram represents the relationship between electrical nodes in the power grid.

[0041] The spatial feature extraction module is used to extract the load spatial features of the power grid based on the node topology graph using a graph attention network.

[0042] The time-series feature extraction module is used to perform hybrid time-series extraction on historical load data of the power grid based on the long short-term memory network model and the Transformer model to obtain the load time-series features of the power grid.

[0043] The load forecasting module is used to forecast the load of the power grid based on the spatial and temporal characteristics of the load.

[0044] Thirdly, embodiments of this application provide a load forecasting device, including: a memory and a processor;

[0045] The memory stores instructions that the computer executes;

[0046] The processor executes computer execution instructions stored in memory, causing the processor to perform the methods described in the various possible implementations of the first aspect above.

[0047] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed, are used to implement the methods described in the various possible implementations of the first aspect above.

[0048] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed, implements the methods described in the various possible implementations of the first aspect above.

[0049] The load forecasting method, device, storage medium, and program products provided in this application acquire the node topology map and historical load data corresponding to the power grid. The node topology map represents the correlation between electrical nodes in the power grid. Based on a graph attention network, the load spatial characteristics of the power grid are extracted according to the node topology map. Based on a long short-term memory network model and a Transformer model, a hybrid time-series extraction is performed on the historical load data of the power grid to obtain the load time-series characteristics of the power grid. Based on the load spatial characteristics and load time-series characteristics, the load of the power grid is predicted. This application extracts the load spatial characteristics of the power grid based on a graph attention network, fully exploring the spatial correlation between nodes in the power grid and the nonlinear correlation between loads. By performing a hybrid time-series extraction using a long short-term memory network model and a Transformer model, the load time-series characteristics are obtained, achieving high-dimensional mining of historical load data. Based on the load spatial characteristics and load time-series characteristics, the load of the power grid is predicted, improving the accuracy of load forecasting. Attached Figure Description

[0050] 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.

[0051] Figure 1 A schematic flowchart illustrating the load forecasting method provided in this application embodiment;

[0052] Figure 2 This is a schematic diagram of the load forecasting device provided in the embodiments of this application;

[0053] Figure 3 This is a schematic diagram of the load forecasting device provided in an embodiment of this application.

[0054] The accompanying drawings illustrate 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 concepts of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0055] 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.

[0056] To facilitate understanding of the embodiments of this application, the following points are explained first:

[0057] First, in the embodiments of this application, the use of prefixes such as "first" and "second" is merely for the purpose of distinguishing and describing different things belonging to the same name category, and does not constrain the order, size, or quantity of things. For example, "first communication device" and "second communication device" are simply different devices, and do not limit the number of devices or their priority; similarly, "first information" and "second information" are simply different pieces of information, and there is no temporal sequence, size, or priority relationship between them.

[0058] Second, in the embodiments of this application, "send" and "receive" indicate the direction of signal transmission. For example, "send information to A" can be understood as the destination of the information being A, which may include sending directly through the air interface or sending indirectly through the air interface by other units or modules. "Receive information from B" can be understood as the sender of the information being B, which may include receiving directly from B through the air interface or receiving indirectly from B through the air interface by other units or modules. "Send" can also be understood as the "output" of the chip interface, and "receive" can also be understood as the "input" of the chip interface.

[0059] In other words, sending and receiving can be done between devices, such as between a first communication device and a second communication device; or it can be done within a device, such as between components, modules, chips, software modules, or hardware modules within a device via a bus, wiring, or interface.

[0060] It is understandable that information may undergo necessary processing, such as encoding and modulation, before being sent from the source to the destination. Similarly, the destination, upon receiving information from the source, can also perform corresponding processing, such as decoding and demodulation, to interpret the valid information from the source. Similar expressions in this application can be understood in a similar way and will not be elaborated further.

[0061] Third, in the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates an "or" relationship between the preceding and following related objects, but it does not exclude the possibility of indicating an "and" relationship. The specific meaning can be understood in conjunction with the context. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can represent: a, b, c; a and b; a and c; b and c; or a and b and c. Here, a, b, and c can be single or multiple.

[0062] Fourth, in the embodiments of this application, "instruction" can include direct instruction and indirect instruction, as well as explicit instruction and implicit instruction. The information indicated by a certain piece of information (hereinafter referred to as instruction information) is called the information to be instructed. In the specific implementation process, there are many ways to indicate the information to be instructed, such as, but not limited to, directly indicating the information to be instructed, such as the information to be instructed itself or its index. It can also indirectly indicate the information to be instructed by indicating other information, where there is a correlation between the other information and the information to be instructed; or it can only indicate a part of the information to be instructed, while the other parts of the information to be indicated are known or pre-agreed upon. For example, the instruction of specific information can be achieved by using a pre-agreed (e.g., protocol predefined) arrangement order of various pieces of information, thereby reducing the instruction overhead to a certain extent. This application does not limit the specific method of instruction.

[0063] It is understandable that, for the sending end of the indication information, the indication information can be used to indicate the information to be indicated, and for the receiving end of the indication information, the indication information can be used to determine the information to be indicated.

[0064] Fifth, in the embodiments of this application, descriptions such as "when," "under the circumstances," "if," and "if" all refer to the fact that the device (e.g., the first communication device or the second communication device) will make corresponding processing under certain objective circumstances. They are not time limits, nor do they require the device (e.g., the first communication device or the second communication device) to have a judgment action when implementing it, nor do they mean that there are other limitations.

[0065] Sixth, the pre-definition in the embodiments of this application can be understood as: definition, pre-definition, storage, pre-storage, pre-negotiation, pre-configuration, solidification, or pre-firing.

[0066] Seventh, the storage involved in the embodiments of this application can refer to storage in one or more memories. These one or more memories can be a separate setting or integrated into an encoder, decoder, processor, or communication device. Alternatively, some of these memories can be set separately, while others can be integrated into the decoder, processor, or communication device. The type of memory can be any form of storage medium, and this application is not limited to this.

[0067] Traditional load forecasting technology has the following bottlenecks:

[0068] 1. Limitations of Statistical Models: Time series analysis, such as ARIMA models, seasonality decomposition models, and regression analysis, has limited capabilities in handling nonlinear relationships. For example, when a sudden temperature change triggers a surge in air conditioning load, which is superimposed on holiday factors, traditional linear models struggle to capture the cross-coupling effects. For every 1°C increase in temperature, summer air conditioning load may increase by 3%-5%, and under extreme weather conditions, this increase exhibits a nonlinear jump, which traditional models cannot accurately describe.

[0069] 2. Machine Learning Bottlenecks: While algorithms such as Support Vector Machines and shallow neural networks can handle some nonlinear problems, they heavily rely on manual feature engineering. When predicting load in commercial areas, features such as "weekday identifiers" and "temperature thresholds" need to be manually constructed. Furthermore, when users' electricity consumption behavior undergoes structural changes, the model needs to redesign features to maintain accuracy, resulting in poor adaptability to dynamic patterns.

[0070] 3. Lack of spatiotemporal correlation: Most existing solutions ignore the spatial correlation inherent in the power grid topology, such as the load transmission effect between adjacent substations and the electrical coupling relationship between regions. Key information is not effectively utilized. Studies have shown that the load correlation of substations within a 10km distance can reach over 0.85, but traditional methods can only process single-point time-series data.

[0071] To address the aforementioned technical bottlenecks, this application proposes the following technical concept:

[0072] First, this invention aims to overcome the bottlenecks of manual feature engineering. Traditional methods heavily rely on expert experience to design features, which is not only inefficient but also makes it difficult to comprehensively mine the potential information in multi-source data. This invention leverages the powerful automatic learning capabilities of deep networks to perform deep processing on multi-source data, including temporal, spatial, and environmental data, automatically extracting high-dimensional nonlinear features. This eliminates excessive reliance on human experience and improves the comprehensiveness and accuracy of feature extraction.

[0073] Secondly, the dynamic adaptability of the model is enhanced. Power grid load is affected by various unforeseen factors, such as heavy rain that may lead to concentrated use of electrical appliances by residents, causing sudden load fluctuations. This invention introduces a reinforcement learning mechanism to construct a decision-making environment that can perceive load changes in real time. This allows the model to automatically adjust and optimize its parameters in real time based on current load fluctuations. In this way, the model can quickly adapt to dynamic load changes, significantly improving the accuracy and stability of predictions and effectively responding to various sudden load events.

[0074] Furthermore, this invention delves into spatial correlations. In power grids, the spatial relationships between nodes play a crucial role in load propagation and distribution. Traditional methods often use fixed formulas to calculate the correlation weights between nodes, such as considering only electrical distance, which fails to accurately reflect the complex and ever-changing influence relationships between nodes in actual operation. This invention utilizes a Graph Attention Network (GAT) and designs an adaptive attention mechanism to dynamically quantify the influence weights between nodes based on historical data and real-time information. This approach breaks the limitations of fixed formulas, enabling more accurate capture of the dynamic correlations between nodes in the power grid topology and significantly improving the accuracy of power grid topology modeling.

[0075] Finally, an end-to-end spatiotemporal collaborative framework is constructed. This invention uses GAT for spatial feature modeling, LSTM-Transformer for temporal feature fusion, and reinforcement learning for dynamic optimization of model parameters. By organically combining these three technologies, a complete collaborative system is formed. This framework can perform in-depth analysis and mining of load data at multiple scales, achieving accurate load forecasting. Simultaneously, it can also perform deep mining of multi-source data, providing richer and more valuable information for power grid operation decisions.

[0076] 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.

[0077] Figure 1 This is a flowchart illustrating the load forecasting method provided in the embodiments of this application, as shown below. Figure 1 As shown, the method includes:

[0078] S101. Obtain the node topology diagram and historical load data corresponding to the power grid. The node topology diagram represents the relationship between electrical nodes in the power grid.

[0079] Specifically, define the node set. For power substations / distribution transformers, edge set E is constructed according to the following rules: based on physical connectivity and the topology of the power grid geographic information system, the degree of connectivity between nodes is calculated. Directly connected nodes When the degree of connection Edges are established when the connection level exceeds a preset threshold; the electrical coupling between nodes is calculated, specifically represented by impedance, and the impedance matrix Z between nodes is calculated. Establish edges at time, Use the impedance reference value; calculate the Pearson correlation coefficient for historical loads. ,when Add virtual edges as needed. Edge weights. , , is an adjustable proportional coefficient. This is the normalized impedance matrix.

[0080] In one implementation, historical load data is obtained by: performing wavelet denoising on the raw load data to retain key fluctuation features; using a periodic encoder to process date types, with weekdays / weekends using one-hot encoding, and constructing influence factors for holidays (e.g., 0.8 for Spring Festival and 0.6 for National Day), and using sin / cos cyclic encoding for the 24-hour period; extracting higher-order statistics such as load skewness, kurtosis, and coefficient of variation within a sliding window (e.g., 3 hours) to capture abrupt changes in fluctuation patterns.

[0081] Meteorological data are represented using multidimensional tensors: ,in For temperature, For humidity, For wind speed, For precipitation; special events construct impact propagation models, such as sports events, to estimate regional increments by multiplying the number of participants by the per capita load factor (e.g., 1.2kW).

[0082] S102. Based on graph attention network, extract the load spatial characteristics of the power grid according to the node topology graph.

[0083] Specifically, graph attention networks, through their unique adaptive attention mechanism, overcome the limitations of traditional graph neural networks that use fixed weights to aggregate neighboring nodes. They can automatically learn and assign differentiated attention weights based on the degree of influence of different nodes on the target load node, prioritizing the aggregation of node features that play a key role in load changes. Furthermore, by combining a hierarchical aggregation strategy, multi-layer graph attention network structures can be constructed according to different geographical or electrical distance scales, capturing the spatial correlation patterns of loads from local node interactions to regional power grid linkages layer by layer. Ultimately, this achieves accurate extraction of the spatial distribution characteristics of power grid loads, load coupling characteristics between nodes, and regional load evolution characteristics, providing high-value spatial dimension support for subsequent load forecasting.

[0084] S103. Based on the Long Short-Term Memory Network Model and the Transformer Model, a hybrid time-series extraction is performed on the historical load data of the power grid to obtain the load time-series characteristics of the power grid.

[0085] A hybrid time-series feature extraction framework based on a Long Short-Term Memory (LSTM) network model and a Transformer model is used to deeply mine multi-scale time-series patterns in historical power grid load data to obtain comprehensive and accurate power grid load time-series features. The LTM network model, with its unique gating unit mechanism, effectively captures long-range dependencies in load data, solving the gradient vanishing problem of traditional recurrent neural networks and accurately fitting the correlation patterns between distant time points in historical load sequences, such as the load fluctuation correlation between weekdays and holidays, and seasonal load change trends across months. The Transformer model, relying on its multi-head self-attention mechanism, overcomes the limitations of LTM networks in parallel computing, enabling simultaneous correlation analysis of load data at multiple time steps. It flexibly captures dependencies at any time interval in the load sequence, and further enhances the ability to learn the periodic features of load data by combining trigonometric function time position encoding. The outputs of the LTM network model and the Transformer model are fused to form power grid load time-series features that combine long-range dependencies and global correlations, providing high-quality time-series support for subsequent load forecasting.

[0086] S104. Based on the load spatial characteristics and load temporal characteristics, predict the load of the power grid.

[0087] Specifically, a multi-source feature fusion prediction model is constructed based on load spatial features extracted from the power grid topology using graph attention networks and load temporal features mined by a hybrid long short-term memory network model and a Transformer model. First, spatial and temporal features are deeply fused through feature concatenation and weight learning to generate a fused feature vector that simultaneously represents the spatiotemporal coupling characteristics of the power grid load. Then, external influencing factors such as meteorological correction factors and holiday offsets are incorporated to correct and optimize the fused features. Finally, nonlinear fitting models such as multilayer perceptrons are used to map and learn the fused features, outputting the predicted power grid load for future periods. This achieves accurate and efficient prediction of power grid load, providing reliable data support for power grid safety scheduling, resource allocation, and risk management.

[0088] The load forecasting method provided in this application embodiment is based on graph attention network to extract the load spatial characteristics of the power grid, fully explore the spatial correlation between nodes in the power grid and the nonlinear correlation between loads, and obtain load time series characteristics through hybrid time series extraction using long short-term memory network model and Transformer model, realizing high-dimensional mining of historical load data. Based on the load spatial characteristics and load time series characteristics, the load of the power grid is predicted, improving the accuracy of load forecasting.

[0089] In one possible implementation, based on a graph attention network, the load spatial characteristics of the power grid are extracted according to the node topology graph, including:

[0090] The strength of the association between nodes and their neighboring nodes in the node topology graph;

[0091] Normalize the association strength to obtain the normalized attention weight of the node;

[0092] Based on the normalized attention weights, a multi-head attention weight mechanism is used to enhance the feature extraction of nodes and obtain the node features corresponding to the nodes.

[0093] Based on graph attention networks, hierarchical spatial feature extraction is performed on the node topology graph according to node characteristics to obtain load spatial features.

[0094] Specifically, for each node in the node topology graph The following formula is used to calculate the relationship with neighboring nodes. Association strength:

[0095]

[0096] in For trainable weight matrix, This is the attention vector.

[0097] Furthermore, the attention weights are normalized to obtain normalized attention weights:

[0098]

[0099] The node feature update formula is: And introduce residual connections A multi-head attention mechanism is employed; for example, an 8-head attention mechanism is used. Enhanced feature extraction is achieved to obtain the node features corresponding to each node. Furthermore, a graph attention network is employed to perform hierarchical spatial feature extraction on the node topology graph based on node features, thereby obtaining the load spatial features.

[0100] The load forecasting method provided in this application accurately quantifies the relative importance of each node in the power grid by quantifying and normalizing the association strength between nodes and their neighbors in the node topology graph, generating comparable normalized attention weights. Based on this, a multi-head attention weighting mechanism is introduced to enhance feature extraction, which can uncover hidden associations between nodes from multiple perspectives, effectively avoiding the one-sidedness of single-head attention and obtaining more comprehensive and robust node features. Finally, hierarchical spatial feature extraction is implemented based on a graph attention network, which can capture the load spatial patterns from local node interactions to regional power grid linkages layer by layer. The output load spatial features are both accurate and hierarchical, providing high-quality spatial dimension support for subsequent load forecasting.

[0101] In one possible implementation, based on a graph attention network, hierarchical spatial feature extraction is performed on the node topology graph according to node characteristics to obtain load spatial features, including:

[0102] Based on graph attention networks, local topological features of the node topology graph are extracted within a first radius according to node characteristics.

[0103] Based on graph attention networks, regional topological features of the node topological graph are extracted within a second radius range according to local topological features, wherein the second radius range is larger than the first radius range;

[0104] Based on graph attention networks, global average pooling is performed on regional topological features to obtain load space features.

[0105] For example, a hierarchical spatial feature extraction is achieved through a three-layer structure. The first radius is 5km, and the first-layer graph attention network extracts local topological features. The second radius is 20km, and the second-layer graph attention network captures region-level associations. The third-layer graph attention network performs global average pooling to generate graph-level feature vectors. .

[0106] The load forecasting method provided in this application first captures local topological relationships between nodes within a small first radius, then integrates local features within a larger second radius to uncover topological linkage patterns at the regional scale, and finally uses global average pooling to aggregate the dispersed regional features into unified load spatial features. This achieves hierarchical spatial feature mining from local to regional to global, and accurately quantifies the association weights of nodes within different radii using the adaptive attention mechanism of graph attention networks. The output load spatial features not only fully preserve the interaction details of local nodes in the power grid and the coupling characteristics of regional loads, but also have good global representativeness, providing spatial dimension support that is more consistent with the actual physical topology of the power grid for subsequent load forecasting.

[0107] In one possible implementation, a hybrid time-series extraction is performed on historical load data of the power grid based on a long short-term memory network model and a Transformer model to obtain the load time-series characteristics of the power grid, including:

[0108] Based on historical load data, the first temporal features of the historical load data are extracted using a long short-term memory network model.

[0109] Based on historical load data, the second time-series feature of the historical load data is extracted using the Transformer model, which employs trigonometric function position encoding.

[0110] The first time-series feature and the second time-series feature are concatenated to obtain the combined feature;

[0111] The combined features are fused using the first fully connected layer to obtain the initial fused features. The first fully connected layer uses the ReLU activation function.

[0112] The initial fusion features are fused using a second fully connected layer to obtain load time series features. The second fully connected layer uses the LeakyReLU activation function.

[0113] Specifically, a two-layer 128-neuron Long Short-Term Memory (LSTM) network model is used, inputting minute-level workload data from the past 24 hours. In practical applications, each LSTM unit updates its state using a cell state update formula. Calculate the forget gate using sigmoid and tanh activation functions Input gate and candidate cell status .

[0114] Specifically, the Gate of Oblivion Determines the cell state at the previous moment Information that needs to be retained; input gate and Together, they determine the information that needs to be updated at the current moment. Taking a weekday scenario as an example, this branch can effectively capture the load drop caused by reduced commercial electricity consumption during the lunch break (11:30-13:00) and the short-term load fluctuations caused by the surge in residential electricity consumption after the evening commute (17:30-19:00).

[0115] The Transformer model is configured with eight attention heads, processing daily load data from the past four weeks. Trigonometric function formulas are used. , Perform time location encoding, where This represents location information, where i is the dimension index. At the model level, this encoding method can learn the periodicity of any time interval. In holiday forecasting scenarios, the Transformer model demonstrates a powerful trend-capturing ability. For example, around the Spring Festival holiday, by using an attention mechanism to weighted calculate historical data from the same period, it can accurately capture the trend of a 32% decrease in weekly load, and in error analysis, the average absolute error for this period is controlled within 5.2%.

[0116] The 128-dimensional feature vector output by the Long Short-Term Memory (LSTM) network model and the 512-dimensional feature vector output by the Transformer model are concatenated along the channel dimension to form a 640-dimensional combined feature. To avoid overfitting and enhance feature representation, two fully connected layers with 256 and 128 neurons respectively are used. The first fully connected layer uses the ReLU activation function, and the second layer uses the LeakyReLU activation function. Finally, a 64-dimensional comprehensive temporal feature is generated. This feature retains the high-frequency fluctuation information of minute-level data and integrates the long-term trend characteristics of daily data, which can effectively improve the model's adaptability to complex load patterns in actual forecasting tasks.

[0117] The load forecasting method provided in this application utilizes LSTM to capture long-range temporal dependencies in load data, combines a Transformer with trigonometric function position encoding to accurately identify the periodicity and global correlation characteristics of the load, and then integrates two types of complementary temporal features through feature concatenation. Subsequently, a first fully connected layer activated by ReLU is used to achieve initial nonlinear fusion, and a second fully connected layer activated by LeakyReLU is used to further enhance the nonlinear expressive ability of the features and solve the problem of neuron inactivation in ReLU. The final generated load temporal features have both long-range dependency capture ability and periodic pattern mining ability, which not only effectively improves the richness and robustness of features, but also provides high-quality temporal dimension support for subsequent spatiotemporal feature fusion and load forecasting.

[0118] In one possible implementation, it also includes:

[0119] Based on reinforcement learning mechanisms, optimize the long short-term memory network model and the Transformer model;

[0120] The state space of the reinforcement learning mechanism includes model hyperparameters, prediction error distribution, and load fluctuation index; the action space includes fine-tuning operations, number of units in the long short-term memory network model, structure switching, and parameter reset.

[0121] The reward function is ,in, and The preset weighting coefficients, The mean absolute percentage error, For training time, As a stability reward, The number of days of continuous operation without reconfiguration.

[0122] Specifically, the state space S includes: model hyperparameters, such as learning rate and dropout rate; prediction error distribution, such as sliding window statistics of mean absolute error (MAE) and mean absolute percentage error (MAPE); and load fluctuation indicators, such as the load variation coefficient of the most recent 24 hours.

[0123] Action space A includes: fine-tuning operations, such as adjusting the number of GAT heads (4 / 8 / 12); the number of LSTM units, such as 64 / 128 / 256; structure switching, such as enabling / disabling Transformer branches; and parameter resetting, such as reinitializing the weights of fully connected layers.

[0124] The load forecasting method provided in this application overcomes the limitations of traditional gradient descent algorithms that rely on fixed loss functions by dynamically interacting between the agent in the reinforcement learning mechanism and the model training process, thus achieving adaptive optimization of key hyperparameters and training strategies. Reinforcement learning can adjust the optimization direction in real time based on the model's ability to extract time-series features. This not only enhances the LSTM's ability to capture long-term time-series dependencies and alleviates its gradient vanishing problem, but also improves the Transformer's multi-head attention mechanism's accuracy in mining load cycle patterns. At the same time, the design of the reward function effectively balances the model's fitting ability and generalization ability, avoiding overfitting or underfitting. Finally, the optimized LSTM and Transformer models can generate more targeted and robust time-series features, providing more efficient model support for subsequent spatiotemporal feature fusion and accurate power grid load forecasting, significantly improving the stability and practicality of the entire forecasting framework.

[0125] In one possible implementation, the Long Short-Term Memory network model and the Transformer model are optimized based on a reinforcement learning mechanism, including:

[0126] Based on the proximal policy optimization algorithm and online learning mechanism, the optimized parameters of the Long Short-Term Memory Network Model and the Transformer Model are determined. The online learning mechanism includes fault injection, online training, and setting safety constraints.

[0127] The Long Short-Term Memory (LSTM) network model and the Transformer model are optimized based on the optimization parameters.

[0128] Specifically, a near-end policy optimization algorithm and a policy network are employed. A 3-layer MLP (256-128-64 cells) is used; the value function V(s) employs a two-stream structure to evaluate state value; the dominance function is implemented through... Calculation, where .

[0129] During the exploration phase, 20 fault conditions were introduced into the simulation environment, such as a sudden drop in new energy output and transformer overload; the offline-trained policy network was migrated to the online system; and safety constraints were set to ensure that the action space was limited and optimized, with a single adjustment of MAPE fluctuation of <0.5%.

[0130] The load forecasting method provided in this application leverages the near-end constraint characteristics of the near-end strategy optimization algorithm to avoid training crashes caused by excessive model parameter updates, ensuring the stability and efficiency of parameter optimization. Furthermore, it simulates complex operating conditions of the real power grid through fault injection in an online learning mechanism, achieves dynamic iteration of model parameters through online training, and limits the reasonable range of parameter adjustments to safety constraints, making the optimization process more aligned with actual engineering needs. The final optimized LSTM and Transformer models not only accurately capture the long-term temporal dependencies and periodic correlations of power grid loads but also possess stronger anti-interference capabilities and robustness, effectively improving the quality and reliability of subsequent load temporal feature extraction, and providing more stable and accurate model support for power grid load forecasting.

[0131] In one possible implementation, load forecasting of the power grid is performed based on load spatial characteristics and load temporal characteristics, including:

[0132] The first comprehensive feature is obtained by concatenating the load spatial characteristics and the load temporal characteristics;

[0133] The weight matrix is ​​used to transform the comprehensive features to obtain the second comprehensive feature;

[0134] Based on the second comprehensive feature, the ReLU activation function is used to generate the fusion feature;

[0135] Based on the multilayer perceptron (MLP), the load of the power grid is predicted according to the fusion characteristics and the meteorological correction factor, where the meteorological correction factor represents the impact of meteorological factors on the load.

[0136] Specifically, the feature fusion layer fuses the spatial features and temporal features of the load to obtain fused features. Specifically, ,in It is the load space feature extracted from the power grid node topology graph by the graph attention network; The first comprehensive feature is formed by concatenating the load time series features extracted by the LSTM and Transformer hybrid model. . It is a fusion weight matrix. These are bias terms, and both are parameters that the model needs to learn. Their function is to perform a linear transformation on the concatenated high-dimensional features to obtain a second comprehensive feature, which adapts to the dimensionality requirements of subsequent prediction tasks. ReLU is an activation function that introduces a non-linear transformation, enabling the model to capture the complex relationships between spatiotemporal features.

[0137] Based on fusion features Furthermore, at the forecast output layer, the load forecast result is obtained using the following formula. :

[0138]

[0139] in, It uses a multilayer perceptron to perform a nonlinear mapping on the fused features and outputs a base load forecast value. It is the Hadamard product (element-by-element multiplication), and its function is to use a weather correction factor. Adjust the base forecast value proportionally. These are quantified meteorological impact parameters. For example, when meteorological factors cause an increase in load, If the result is positive, the predicted value is increased accordingly; otherwise, it is decreased, thereby improving the prediction accuracy. It is a holiday correction item used to compensate for the load difference between weekdays and holidays. For example, industrial load decreases and residential load increases during holidays. It is an offset learned from historical holiday load data.

[0140] The load forecasting method provided in this application integrates the spatial topological features and multi-dimensional temporal features of the power grid load through a feature fusion layer. It uses the ReLU activation function to mine the complex nonlinear correlations between features and generates fused features with both spatiotemporal coupling characteristics. Then, by adjusting the meteorological correction factor ratio and compensating for holiday deviations in the forecast output layer, the accuracy of the basic forecast values ​​and the actual operating conditions are greatly improved. At the same time, by relying on multi-task learning to jointly optimize load forecasting and key site classification, feature sharing and mutual promotion between tasks are achieved. This not only enhances the accuracy and robustness of load forecasting, but also outputs the key site identification results simultaneously, providing dual data support for the precise scheduling, risk management and resource allocation of the power grid.

[0141] In one possible implementation, it also includes:

[0142] Based on the load forecast results, load data mining applications are carried out, including abnormal load identification, attribution of influencing factors, and scenario simulation.

[0143] Specifically, construct the prediction error distribution. An alarm is triggered in time, and the abnormal source node is located by combining the GAT attention weight.

[0144] Through feature importance The impact of various factors was quantified, such as the Spring Festival causing a 42% decrease in industrial load and a 28% increase in residential load.

[0145] The scenario simulation includes temperature rise scenarios (+1℃~+5℃), new energy penetration rate (20%~50%), etc., and outputs load curves and key indicators. For example, a 3℃ temperature rise in summer leads to a 12.8% increase in peak load, and when the new energy penetration rate is 40%, the load fluctuation variance increases by 150%.

[0146] The system deployment architecture is as follows: sensor data → Kafka streaming platform → real-time feature extraction and computation → inference model service, which in turn realizes the functions of prediction API (scheduling system), attribution report generation, and scene simulation engine. Finally, the results are cached through Redis and displayed in a web visualization.

[0147] The load forecasting method provided in this application relies on the deviation between the forecast value and the actual load to pinpoint the time and regional range of abnormal load fluctuations, quickly identifying potential risks in power grid operation. By comparing the difference between the forecast value and the actual value and combining external factors such as spatiotemporal characteristics, weather, and holidays, the core driving factors affecting load fluctuations can be accurately traced, providing data support for the formulation of load control strategies. At the same time, based on the forecast results, diverse load scenario simulations can be constructed to predict the power grid load change trends under different operating conditions in advance, helping power grid dispatching departments to optimize resource allocation and formulate emergency plans. Ultimately, it achieves closed-loop management from load forecasting to actual application, significantly improving the safety, economy, and intelligence level of power grid operation.

[0148] This solution overcomes the bottlenecks of traditional methods in feature engineering, spatiotemporal correlation modeling, and dynamic adaptability through a three-level collaborative architecture: GAT spatial modeling, LSTM-Transformer temporal fusion, and reinforcement learning dynamic optimization. Experiments show that its prediction accuracy (MAE ≤ 4.2%) and anomaly detection efficiency (dynamic threshold) are significantly superior to existing technologies. In the future, it can be combined with federated learning technology to achieve cross-regional load collaborative prediction while ensuring user privacy, further supporting the intelligent scheduling and safe operation of new power systems.

[0149] Figure 2 This is a schematic diagram of the load forecasting device provided in an embodiment of this application. Figure 2 As shown, the load forecasting device 20 provided in this embodiment includes:

[0150] The acquisition module 201 is used to acquire the node topology diagram and historical load data corresponding to the power grid. The node topology diagram represents the relationship between electrical nodes in the power grid.

[0151] The spatial feature extraction module 202 is used to extract the load spatial features of the power grid based on the node topology graph using a graph attention network.

[0152] The time-series feature extraction module 203 is used to perform hybrid time-series extraction on historical load data of the power grid based on the long short-term memory network model and the Transformer model to obtain the load time-series features of the power grid.

[0153] The load forecasting module 204 is used to forecast the load of the power grid based on the load spatial characteristics and load temporal characteristics.

[0154] In one possible implementation, the spatial feature extraction module 202 is specifically used for:

[0155] The strength of the association between nodes and their neighboring nodes in the node topology graph;

[0156] Normalize the association strength to obtain the normalized attention weight of the node;

[0157] Based on the normalized attention weights, a multi-head attention weight mechanism is used to enhance the feature extraction of nodes and obtain the node features corresponding to the nodes.

[0158] Based on graph attention networks, hierarchical spatial feature extraction is performed on the node topology graph according to node characteristics to obtain load spatial features.

[0159] In one possible implementation, the spatial feature extraction module 202 is further configured to:

[0160] Based on graph attention networks, local topological features of the node topology graph are extracted within a first radius according to node characteristics.

[0161] Based on graph attention networks, regional topological features of the node topological graph are extracted within a second radius range according to local topological features, wherein the second radius range is larger than the first radius range;

[0162] Based on graph attention networks, global average pooling is performed on regional topological features to obtain load space features.

[0163] In one possible implementation, the time-series feature extraction module 203 is further configured to:

[0164] Based on historical load data, the first temporal features of the historical load data are extracted using a long short-term memory network model.

[0165] Based on historical load data, the second time-series feature of the historical load data is extracted using the Transformer model, which employs trigonometric function position encoding.

[0166] The first time-series feature and the second time-series feature are concatenated to obtain the combined feature;

[0167] The combined features are fused using the first fully connected layer to obtain the initial fused features. The first fully connected layer uses the ReLU activation function.

[0168] The initial fusion features are fused using a second fully connected layer to obtain load time series features. The second fully connected layer uses the LeakyReLU activation function.

[0169] In one possible implementation, a model optimization module is also included, for:

[0170] Based on reinforcement learning mechanisms, optimize the long short-term memory network model and the Transformer model;

[0171] The state space of the reinforcement learning mechanism includes model hyperparameters, prediction error distribution, and load fluctuation index; the action space includes fine-tuning operations, number of units in the long short-term memory network model, structure switching, and parameter reset.

[0172] The reward function is ,in, and The preset weighting coefficients, The mean absolute percentage error, For training time, As a stability reward, The number of days of continuous operation without reconfiguration.

[0173] In one possible implementation, the model optimization module is also used for:

[0174] Based on the proximal policy optimization algorithm and online learning mechanism, the optimized parameters of the Long Short-Term Memory Network Model and the Transformer Model are determined. The online learning mechanism includes fault injection, online training, and setting safety constraints.

[0175] The Long Short-Term Memory (LSTM) network model and the Transformer model are optimized based on the optimization parameters.

[0176] In one possible implementation, the load forecasting module 204 is specifically used for:

[0177] The first comprehensive feature is obtained by concatenating the load spatial characteristics and the load temporal characteristics;

[0178] The weight matrix is ​​used to transform the comprehensive features to obtain the second comprehensive feature;

[0179] Based on the second comprehensive feature, the ReLU activation function is used to generate the fusion feature;

[0180] Based on the multilayer perceptron (MLP), the load of the power grid is predicted according to the fusion characteristics and the meteorological correction factor, where the meteorological correction factor represents the impact of meteorological factors on the load.

[0181] In one possible implementation, the load forecasting module 204 is further configured to:

[0182] Based on the load forecast results, load data mining applications are carried out, including abnormal load identification, attribution of influencing factors, and scenario simulation.

[0183] The load forecasting 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.

[0184] Figure 3 This is a schematic diagram of the load forecasting device provided in an embodiment of this application. Figure 3 As shown, the load forecasting device 30 provided in this embodiment includes at least one processor 301 and a memory 302. Optionally, the load forecasting device 30 further includes a communication interface 303. The processor 301, memory 302, and communication interface 303 are connected via a communication bus 304.

[0185] In a specific implementation, at least one processor 301 executes computer execution instructions stored in memory 302, causing at least one processor 301 to perform the above-described method.

[0186] The specific implementation process of processor 301 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0187] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0188] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0189] 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. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0190] This application also provides a computer program product, including a computer program that, when executed, implements the above-described method.

[0191] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed, implement the above-described method.

[0192] The aforementioned 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.

[0193] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an application-specific integrated circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0194] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0195] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0196] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0197] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0198] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0199] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A load forecasting method, characterized in that, include: Obtain the node topology diagram and historical load data corresponding to the power grid, wherein the node topology diagram represents the relationship between electrical nodes in the power grid; Based on graph attention networks, the load spatial characteristics of the power grid are extracted according to the node topology graph; Based on the Long Short-Term Memory network model and the Transformer model, a hybrid time-series extraction is performed on the historical load data of the power grid to obtain the load time-series characteristics of the power grid. The load of the power grid is predicted based on the load spatial characteristics and the load temporal characteristics.

2. The load forecasting method according to claim 1, characterized in that, The graph attention-based network extracts the load spatial features of the power grid based on the node topology graph, including: Regarding the association strength between nodes and their neighboring nodes in the node topology graph; The association strength is normalized to obtain the normalized attention weight of the node; Based on the normalized attention weights, a multi-head attention weight mechanism is used to enhance the feature extraction of nodes and obtain the node features corresponding to the nodes. Based on graph attention networks, hierarchical spatial feature extraction is performed on the node topology graph according to the node features to obtain the load spatial features.

3. The load forecasting method according to claim 2, characterized in that, The graph attention-based network extracts hierarchical spatial features from the node topology graph based on the node features to obtain the load spatial features, including: Based on the graph attention network, local topological features of the node topology graph are extracted within a first radius according to the node features; Based on the graph attention network, according to the local topological features, the regional topological features of the node topological graph are extracted within a second radius range, wherein the second radius range is larger than the first radius range; Based on the graph attention network, global average pooling is performed on the topological features of the region to obtain the load space features.

4. The load forecasting method according to any one of claims 1 to 3, characterized in that, The method based on the Long Short-Term Memory (LSTM) network model and the Transformer model performs hybrid time-series extraction on the historical load data of the power grid to obtain the load time-series characteristics of the power grid, including: Based on the historical load data, and using a long short-term memory network model, the first temporal feature of the historical load data is extracted. Based on the historical load data, a second time-series feature of the historical load data is extracted using a Transformer model, wherein the Transformer model employs trigonometric function position encoding. The first temporal feature and the second temporal feature are concatenated to obtain the combined feature; The combined features are fused using a first fully connected layer to obtain initial fused features. The first fully connected layer uses the ReLU activation function. The initial fusion features are fused using a second fully connected layer to obtain the load time series features. The second fully connected layer uses the LeakyReLU activation function.

5. The load forecasting method according to any one of claims 1 to 3, characterized in that, Also includes: Based on the reinforcement learning mechanism, the long short-term memory network model and the Transformer model are optimized. The state space of the reinforcement learning mechanism includes model hyperparameters, prediction error distribution, and load fluctuation index; the action space includes fine-tuning operations, number of units in the long short-term memory network model, structure switching, and parameter reset. The reward function is ,in, and The preset weighting coefficients, The mean absolute percentage error, For training time, As a stability reward, The number of days of continuous operation without reconfiguration.

6. The load forecasting method according to claim 5, characterized in that, The optimization of the Long Short-Term Memory network model and the Transformer model based on the reinforcement learning mechanism includes: Based on the proximal policy optimization algorithm and the online learning mechanism, the optimized parameters of the Long Short-Term Memory Network Model and the Transformer Model are determined. The online learning mechanism includes fault injection, online training, and setting security constraints. The long short-term memory network model and the Transformer model are optimized based on the optimization parameters.

7. The load forecasting method according to any one of claims 1 to 3, characterized in that, The method of predicting the load of the power grid based on the load spatial characteristics and the load temporal characteristics includes: The load spatial characteristics and the load temporal characteristics are concatenated to obtain the first comprehensive characteristic; The combined features are transformed using a weight matrix to obtain the second combined features; Based on the second comprehensive feature, the ReLU activation function is used to generate the fused feature; Based on a multilayer perceptron (MLP), the load of the power grid is predicted according to the fused features and the meteorological correction factor, wherein the meteorological correction factor represents the impact of meteorological factors on the load.

8. The load forecasting method according to any one of claims 1 to 3, characterized in that, Also includes: Based on the load forecast results, load data mining applications are performed, including abnormal load identification, attribution of influencing factors, and scenario simulation.

9. A load forecasting 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 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed, are used to implement the method as described in any one of claims 1 to 8.

Images