A seasonal residual plot state space power load forecasting method and system

Abstract

The application discloses a kind of seasonal residual graph state space power load prediction method and system, comprising: obtaining multiple power load node prediction data, the prediction data includes historical load sequence, node covariant sequence, node graph structure information, edge attribute information, edge type information, domain identification information and event marker information;The prediction data is preprocessed;According to the historical load sequence after preprocessing, construct seasonal anchor point, and generate seasonal reference item of each load node in the prediction interval according to the seasonal anchor point;The historical load sequence after preprocessing is input into the residual prediction model trained in advance with the seasonal reference item and processed, to obtain the generation multi-node power load prediction result.The method of the application can effectively reduce the difficulty of model training, alleviate the error accumulation problem under long prediction step, filter invalid noise neighbors in topology, suppress error propagation between nodes, and improve the explainability and engineering practicability of the prediction result.

Description

Technical Field

[0001] This invention relates to a seasonal residual diagram state-space power load prediction method and system, belonging to the field of power system load prediction technology. Background Technology

[0002] Currently, power load forecasting is widely used in scenarios such as unit combination, reserve capacity arrangement, demand response, distribution network planning, congestion assessment, power market operation, and regional energy consumption management. With the popularization of smart meters, distribution automation terminals, industrial equipment monitoring systems, and regional load acquisition platforms, a large amount of multi-node load data with time correlation, seasonal periodicity, and cross-node coupling relationships has been generated in the power system.

[0003] Existing short-term load forecasting techniques include statistical regression, weather-sensitive models, recurrent neural networks, temporal convolutional networks, Transformer-type models, Patch-type models, graph neural networks, and state-space models. Current deep time-series forecasting methods typically learn the complete load curve directly, without explicitly distinguishing between the seasonal inertia and short-term disturbance residuals represented by historical load data. This leads to model capacity being used to repeatedly learn date, period, and weekday patterns that can be directly provided by historical data, increasing training difficulty and causing error accumulation over long forecast steps. Existing graph spatiotemporal forecasting models typically use graph convolution, graph attention, or diffusion convolution as the main forecasting operators. The influence of graph channels on node states lacks normalization constraints and upper bounds on gain, making it easy to introduce unreliable neighbor information in scenarios with statistical correlation graphs, equipment similarity graphs, or noisy topology graphs, thus causing cross-node error propagation. While existing neural network prediction models can output point prediction results, they typically cannot simultaneously explain how much of the predicted value comes from normal seasonal load inertia, how much from short-term residual dynamics, and how much from cross-node coupling correction. This makes it difficult for power grid operators to determine whether the prediction results follow normal demand patterns or are affected by abnormal events, weather disturbances, and changes in neighboring nodes. Electricity load data exhibits conceptual drift characteristics; for example, holidays, extreme weather, changes in industrial production plans, changes in residential electricity consumption behavior, and equipment failures can all alter load distribution. Existing offline training models lack lightweight online drift detection and adaptive fine-tuning mechanisms, resulting in the model's inability to promptly update seasonal anchor points, residual branch weights, and graph correction strength after changes in load statistical characteristics. Summary of the Invention

[0004] The technical problem that this invention aims to solve is: the lack of separation of time-series components, the susceptibility of graph structure to propagation errors, and the inexplicable causes of predictions in existing power load forecasting.

[0005] To solve the above-mentioned technical problems, the present invention is implemented using the following technical solution.

[0006] On one hand, the present invention provides a method for predicting seasonal residual diagram state-space power load, comprising:

[0007] Acquire prediction data for multiple power load nodes, including historical load sequences, node covariate sequences, node graph structure information, edge attribute information, edge type information, domain identifier information, and event marker information;

[0008] The predicted data is preprocessed;

[0009] Based on the preprocessed historical load sequence, seasonal anchor points are constructed, and seasonal reference terms for each load node within the prediction interval are generated based on the seasonal anchor points.

[0010] The preprocessed historical load sequence and the seasonal benchmark term are input into a pre-trained residual prediction model for processing to obtain multi-node power load prediction results.

[0011] The processing performed by the residual prediction model includes:

[0012] The residual input is constructed based on the difference between the preprocessed historical load sequence and the seasonal benchmark term. The residual input is then subjected to multi-scale selective state space scanning to extract the residual fluctuation features at each time scale. The residual fluctuation features at each time scale are then mapped to the corresponding scale decoding head to obtain residual prediction components at multiple time scales.

[0013] Learnable hybrid weights are obtained through Softmax normalization. The residual prediction components of the multiple time scales are fused based on the learnable hybrid weights to obtain the fused residual term.

[0014] Based on the preprocessed historical load sequence, node covariate sequence, node graph structure information, and edge attribute information, normalized edge gating weights are calculated, the neighbor node states of each node are obtained, and the neighbor node states are aggregated based on the normalized edge gating weights to obtain graph correction terms.

[0015] The seasonal baseline term, the fused residual term, and the graph correction term are added together to generate a multi-node power load forecast. The corresponding operational profile is obtained based on the seasonal baseline term, the fused residual term, the graph correction term, the historical load sequence, the node covariate sequence, the domain identifier information, and the event marker information.

[0016] By constructing seasonal anchor points to generate seasonal baseline terms, and using the difference between historical loads and seasonal baseline terms as residual inputs, the residual prediction model only needs to learn the residual dynamics that deviate from seasonal inertia, reducing the difficulty of model learning and reducing error accumulation under long prediction steps. By aggregating the states of neighboring nodes with normalized edge-gated weights to generate graph correction terms, the cross-node coupling is ensured to be a bounded perturbation.

[0017] The preprocessing includes:

[0018] The historical load sequence is processed by imputing missing values, truncating outliers, normalizing by time and node, encoding holidays, encoding weekdays, and normalizing by weather features;

[0019] The edges in the node graph structure information are labeled with types, and the discrete edge type labels are mapped to vectors through an embedding layer;

[0020] The node covariate sequence is subjected to holiday coding, weekday coding, and weather feature normalization.

[0021] The edge attribute information is then subjected to numerical normalization and post-processing.

[0022] The edge type information is discretely labeled and mapped to an edge type embedding vector;

[0023] The domain identification information is encoded to generate a domain embedding vector;

[0024] The event tagging information is binarized or hierarchically encoded to generate event covariates.

[0025] The seasonal anchor points include at least one of the following: daily cycle anchor points, weekly cycle anchor points, weekday condition matching anchor points, holiday condition matching anchor points, or weighted fusion anchor points of multiple historical load values ​​for the same period.

[0026] By constructing various anchor points such as daily cycles, weekly cycles, and weekday / holiday condition matching, the inherent strong seasonality of power load is explicitly introduced into the prediction framework, so that the residual prediction model does not need to relearn the cycle pattern from the data.

[0027] The residual prediction model adopts a multi-scale topological conditional selective state-space model, the architecture of which includes:

[0028] The input layer is used to receive the preprocessed multi-node historical load sequence, node graph structure, node covariates, and edge attributes.

[0029] The multi-scale branching layer includes at least two of the fast-scale branching, daily-scale branching, and weekly-scale branching, and each branch contains an L-layer topological condition-selective state space layer.

[0030] The graph message passing layer generates bounded graph correction terms based on the node graph structure information by normalizing edge gating weights and injects them into the state updates of each branch.

[0031] The multi-scale fusion layer is used to fuse the outputs of each branch with learnable Softmax weights and then map them to residual prediction components by the decoder head.

[0032] The training of the residual prediction model includes: training using Huber loss, and incorporating one or more of the following: aggregated balance error term, residual scaling constraint term, and graph Laplace energy constraint term, as shown in the following formula:

[0033] ,

[0034] in, For the total loss, For Huber's losses, To predict load values, This represents the actual load value. To aggregate and balance error weights, This is the aggregation equilibrium error term. As the residual proportion constraint weight, For residual proportion constraint term, The Laplace energy constraint weights are used for the graph. For the graph Laplace energy constraint term;

[0035] The formula for the Graph Laplace energy constraint is as follows:

[0036] ,

[0037] in, To transpose the predicted load value, Let be the graph Laplace matrix constructed using normalized edge weights.

[0038] By extracting residual fluctuation features at different time scales such as fast, daily, and weekly through multi-scale selective state-space scanning, it can simultaneously capture short-term disturbances, daily periodic deviations, and weekly periodic changes, thereby improving prediction accuracy and robustness. Huber loss is insensitive to outliers. Aggregate balancing error terms ensure that the contributions of each branch are balanced. Residual proportion constraint terms prevent residual terms from dominating predictions. Graph Laplace energy constraint terms encourage prediction results to be smooth on the graph, avoiding overfitting to noisy topologies.

[0039] The multi-scale selective state-space scanning includes: setting a set of time scales. The set of time scales includes at least two of the following: fast scale, daily scale, and weekly scale;

[0040] The preprocessed historical load sequence and node covariate sequence together constitute the historical input tensor for each scale. The historical input tensor is converted into a scale-dependent sequence, and node-level hidden states are maintained.

[0041] Any scale Next node The hidden state satisfies:

[0042] ,

[0043] in, For scale Next node At any moment The hidden state, For input-related attenuation gating, For scale Next node In the hidden state of the previous moment, It is the identity matrix. For the input gate vector, For element-wise multiplication, For candidate residual states, To illustrate message gain, For image message;

[0044] The input-related attenuation gating satisfies:

[0045] ,

[0046] in, It is a natural exponential function. The positive step size parameter is generated from the input features. This is the softplus activation function. For scale State transition parameters under the following conditions;

[0047] The graph message gain satisfies:

[0048] ,

[0049] in, For the Sigmoid function, For learnable parameters, This is the preset maximum graph gain.

[0050] The formula for obtaining the learnable hybrid weights through Softmax normalization is as follows:

[0051] ,

[0052] in, For scale Mixed weights, It is a natural exponential function. For scale Learnable weight parameters, For scale Learnable weight parameters, It is a set of time scales.

[0053] The step of calculating normalized edge gating weights based on the preprocessed historical load sequence, the node covariate sequence, the node graph structure information, and the edge attribute information includes:

[0054] After performing the preprocessing on the node graph structure information, a discrete edge type label set is obtained, and the discrete edge type labels are mapped to edge type embedding codes.

[0055] After performing the preprocessing on the historical load sequence, the node features of each node at each time step are obtained, wherein the node features of the current node are used as the target node features, and the node features of the neighboring nodes are used as the neighboring node features.

[0056] The event tagging information is preprocessed to obtain event covariates, and the event tags include at least one of extreme weather tags, holiday tags, and equipment shutdown tags.

[0057] The normalized edge gating weights are calculated based on the target node features, neighbor node features, edge attributes, edge type embeddings, and event covariates to determine the unnormalized edge gating values. The gating values ​​of incoming edges to the same destination node are then normalized, as shown in the following formula:

[0058] ,

[0059] in, To normalize the edge gate weights, For unnormalized edge gating values, To prevent the stability constant from being divided by zero, Indexing neighboring nodes, For nodes The neighborhood group, Neighboring nodes To the node The unnormalized edge gating value.

[0060] By normalizing the destination node, the graph message is made into a weighted average of the neighbor states, avoiding different degrees of graph injection to different nodes due to differences in the number of edges. Event covariates are introduced in the edge gating weight calculation, enabling the graph structure to perceive the current system state and dynamically adjust the impact of neighbors in the event of abnormal events.

[0061] The operational profile includes one or more of the following: time branch mixed weight, residual magnitude ratio, average side gating strength, graph gain, seasonal source share, residual source share, and graph source share;

[0062] The time-branch hybrid weight is obtained by taking the mean of the historical load sequence and the node covariate sequence in the time dimension and the node dimension to obtain a global feature vector; the event tag information is preprocessed to obtain event covariates, which are used as event summaries; the domain identifier information is preprocessed to obtain domain embedding vectors, which are used as domain embeddings; the global feature vector is concatenated with the event summary and the domain embeddings and mapped, and then normalized by Softmax to obtain the time-branch hybrid weight.

[0063] The domain embedding is a vector corresponding to the domain identification information obtained by querying the learnable embedding table built into the residual prediction model based on the preprocessed domain identification information. Different data domains correspond to different domain embedding vectors. The domain identification information is used to identify the type of data source. The learnable embedding table is randomly generated when the residual prediction model is initialized and is jointly optimized with other network parameters through backpropagation during the training process of the residual prediction model.

[0064] The formula for the residual amplitude ratio is as follows:

[0065] ,

[0066] in, The residual amplitude ratio, To merge residual terms, To prevent the stability constant from being divided by zero, For seasonal baseline items;

[0067] The formula for the average side-gating strength is as follows:

[0068] ,

[0069] in, For average edge gating strength, For the set of side-gated samples in the test batch, The reciprocal of the number of samples in the test batch. To normalize the edge gate weights;

[0070] The seasonal source share Residual source share and the share of the source of the image The formula is as follows:

[0071] ,

[0072] ,

[0073] ,

[0074] in, For seasonal sources of quantity, For seasonal baselines, For residual source components, For time scale sets, For scale Mixed weights, For scale The residual prediction components below, For the source of the image, For the graph correction item, The normalization constant is To prevent the stability constant from being divided by zero, For seasonal source share, For the residual source share, For the source share, and .

[0075] The operational profile includes time branch mixing weights, residual magnitude ratio, average edge gating strength, graph gain, and seasonality, residual, and graph source share, comprehensively explaining the prediction results from four dimensions: time frequency, residual strength, graph activation level, and source composition.

[0076] The operation profile is output in the form of JSON, CSV, database fields, or scheduling system interface messages;

[0077] The formula for the multi-node power load prediction result is as follows:

[0078] ,

[0079] in, For nodes In the prediction step The predicted load value, For seasonal baselines, For scale Mixed weights, For scale The residual prediction components below, For graph correction terms.

[0080] In a second aspect, the present invention provides a computer system including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the seasonal residual diagram state-space power load forecasting method.

[0081] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0082] This invention separates the seasonal baseline term from the load residual by constructing a seasonal anchor point, enabling the residual prediction model to learn only short-term disturbance biases, effectively reducing the difficulty of model training and alleviating the error accumulation problem under long prediction steps. It introduces normalized edge-gated weights to aggregate neighboring node information, setting constraint boundaries for the influence of graph channel nodes, filtering out invalid noisy neighbors in the topology, and suppressing error propagation across nodes. The final prediction result is composed of the superposition of the seasonal baseline term, multi-scale fused residual term, and graph correction term, and generates a corresponding operational profile, clearly distinguishing the contribution ratios of seasonal inertia, short-term fluctuation residuals, and node coupling correction. This facilitates power grid personnel in identifying the causes of load anomalies, accurately judging the impact of abnormal events, weather disturbances, and changes in neighboring node electricity consumption, and improving the interpretability and engineering applicability of the prediction results. Attached Figure Description

[0083] Figure 1 This is a flowchart illustrating the seasonal residual diagram state-space power load prediction method shown in Embodiment 1 of the present invention.

[0084] Figure 2 This is a schematic diagram of the multi-scale selective state-space residual scanning module structure shown in Embodiment 1 of the present invention;

[0085] Figure 3 This is a flowchart of the normalized edge-gated graph correction module as shown in Embodiment 1 of the present invention;

[0086] Figure 4 This is a heatmap showing the average ranking of MAE, RMSE, and WAPE indicators on three publicly available load datasets as shown in Embodiment 3 of the present invention.

[0087] Figure 5 This is a bar chart comparing the scale normalization point error shown in Embodiment 3 of the present invention;

[0088] Figure 6 The ablation experiment thermogram shown in Example 3 of this invention;

[0089] Figure 7 This is an interpretable operational statistics chart of the present invention on UCI test segmentation, as shown in Embodiment 3 of the present invention;

[0090] Figure 8 This is a sensitivity analysis diagram of the non-seasonal module shown in Embodiment 3 of the present invention. Detailed Implementation

[0091] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0092] The term "and / or" simply 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. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0093] Example 1

[0094] This embodiment introduces a seasonal residual diagram state-space power load forecasting method, including:

[0095] Acquire prediction data for multiple power load nodes. The prediction data includes historical load sequences, node covariate sequences, node graph structure information, edge attribute information, edge type information, domain identifier information, and event marker information.

[0096] Preprocess the forecast data;

[0097] Seasonal anchor points are constructed based on the preprocessed historical load sequence, and seasonal reference terms for each load node within the prediction interval are generated based on the seasonal anchor points.

[0098] The preprocessed historical load sequence and seasonal baseline terms are input into a pre-trained residual prediction model for processing to obtain multi-node power load prediction results.

[0099] The processing performed by the residual prediction model includes:

[0100] The residual input is constructed based on the difference between the preprocessed historical load sequence and the seasonal baseline term. Multi-scale selective state space scanning is performed on the residual input to extract the residual fluctuation characteristics at each time scale. The residual fluctuation characteristics at each time scale are mapped by the corresponding scale decoding head to obtain the residual prediction components at multiple time scales.

[0101] Learnable hybrid weights are obtained through Softmax normalization. The residual prediction components of multiple time scales are fused based on the learnable hybrid weights to obtain the fused residual term.

[0102] Based on the preprocessed historical load sequence, node covariate sequence, node graph structure information and edge attribute information, normalized edge gating weights are calculated, the neighbor node states of each node are obtained, and the neighbor node states are aggregated based on the normalized edge gating weights to obtain the graph correction term.

[0103] The seasonal baseline term, the fused residual term, and the graph correction term are added together to generate a multi-node power load forecast. The corresponding operational profile is obtained based on the seasonal baseline term, the fused residual term, the graph correction term, the historical load sequence, the node covariate sequence, the domain identifier information, and the event marker information.

[0104] Specifically, the seasonal anchor point includes at least one of the following: daily cycle anchor point, weekly cycle anchor point, weekday condition matching anchor point, holiday condition matching anchor point, or weighted fusion anchor point of multiple historical load values ​​for the same period; the formula for the seasonal benchmark term is as follows:

[0105] ,

[0106] in, For nodes In the prediction step Seasonal benchmark, For nodes At a historical moment The load value, This represents the seasonal cycle offset.

[0107] Preprocessing includes:

[0108] The historical load series is processed by imputing missing values, truncating outliers, normalizing by time nodes, encoding holidays, encoding weekdays, and normalizing by weather characteristics;

[0109] Type labeling is performed on the edges in the node graph structure information, and the discrete edge type labels are mapped to vectors through the embedding layer;

[0110] The node covariate sequence is processed by holiday coding, weekday coding, and weather feature normalization.

[0111] After normalizing the edge attribute information, perform post-processing.

[0112] Discrete label encoding is performed on edge type information, and it is mapped to edge type embedding vector;

[0113] The domain identification information is encoded to generate a domain embedding vector;

[0114] Binarize or hierarchically encode the event tagging information to generate event covariates.

[0115] The residual prediction model employs a multi-scale topological condition-selective state-space model, the architecture of which includes:

[0116] The input layer is used to receive the preprocessed multi-node historical load sequence, node graph structure, node covariates, and edge attributes.

[0117] The multi-scale branching layer includes at least two of the fast-scale branching, daily-scale branching, and weekly-scale branching, and each branch contains an L-layer topological condition-selective state space layer.

[0118] The graph message passing layer generates bounded graph correction terms based on node graph structure information through normalized edge gating weights and injects them into the state updates of each branch.

[0119] The multi-scale fusion layer is used to fuse the outputs of each branch with learnable Softmax weights and then map them to residual prediction components by the decoder head.

[0120] The training of the residual prediction model includes: using Huber loss and incorporating one or more of the following: aggregated balancing error term, residual scaling constraint term, and graphical Laplace energy constraint term, as shown in the following formula:

[0121] ,

[0122] in, For the total loss, For Huber's losses, To predict load values, This represents the actual load value. To aggregate and balance error weights, This is the aggregation equilibrium error term. As the residual proportion constraint weight, For residual proportion constraint term, The Laplace energy constraint weights are used for the graph. For the graph Laplace energy constraint term;

[0123] The formula for the Graph Laplace energy constraint is as follows:

[0124] ,

[0125] in, To transpose the predicted load value, Let be the graph Laplace matrix constructed using normalized edge weights.

[0126] Specifically, each branch contains an L-layer topological conditional selective state space layer, where L is either 1 or 2. State recursive updates are achieved through input-related decay gating, input gating, bounded injection of candidate states and graph messages.

[0127] like Figure 2 As shown, multi-scale selective state-space scanning includes: setting a set of time scales. The timescale set includes at least two of the following: rapid scale, daily scale, and weekly scale;

[0128] The preprocessed historical load sequence and the nodal covariate sequence together constitute the historical input tensor for each scale. The historical input tensor is converted into a scale-dependent sequence, and node-level hidden states are maintained.

[0129] Any scale Next node The hidden state satisfies:

[0130] ,

[0131] in, For scale Next node At any moment The hidden state, For input-related attenuation gating, For scale Next node In the hidden state of the previous moment, It is the identity matrix. For the input gate vector, For element-wise multiplication, For candidate residual states, To illustrate message gain, For image message;

[0132] Input-related attenuation gating satisfies:

[0133] ,

[0134] in, It is a natural exponential function. The positive step size parameter is generated from the input features. This is the softplus activation function. For scale State transition parameters under the following conditions;

[0135] Image message gain satisfies:

[0136] ,

[0137] in, For the Sigmoid function, For learnable parameters, This is the preset maximum graph gain.

[0138] The formula for obtaining learnable hybrid weights through Softmax normalization is as follows:

[0139] ,

[0140] in, For scale Mixed weights, It is a natural exponential function. For scale Learnable weight parameters, For scale Learnable weight parameters, It is a set of time scales.

[0141] Specifically, preset maximum graph gain The value ranges from 0.01 to 0.30 to prevent graph messages from becoming unbounded hidden channels.

[0142] like Figure 3 As shown, the steps for calculating normalized edge gating weights based on the preprocessed historical load sequence, node covariate sequence, node graph structure information, and edge attribute information include:

[0143] After preprocessing the node graph structure information, a set of discrete edge type labels is obtained, and the discrete edge type labels are mapped to edge type embedding codes.

[0144] After preprocessing the historical load sequence, the node features of each node at each time step are obtained, where the node features of the current node are used as the target node features, and the node features of the neighboring nodes are used as the neighboring node features.

[0145] After preprocessing, the event tagging information is used to obtain event covariates. The event tags include at least one of the following: extreme weather tags, holiday tags, and equipment shutdown tags.

[0146] The normalized edge gating weights calculate the unnormalized edge gating values ​​based on the target node features, neighbor node features, edge attributes, edge type embeddings, and event covariates, and normalize the incoming edge gating values ​​for the same destination node, as shown in the following formula:

[0147] ,

[0148] in, To normalize the edge gate weights, For unnormalized edge gating values, To prevent the stability constant from being divided by zero, Indexing neighboring nodes, For nodes The neighborhood group, Neighboring nodes To the node The unnormalized edge gating value.

[0149] By normalizing the destination node, the graph message is made into a weighted average of the neighbor states, avoiding different degrees of graph injection to different nodes due to differences in the number of edges. Event covariates are introduced in the edge gating weight calculation, enabling the graph structure to perceive the current system state and dynamically adjust the impact of neighbors in the event of abnormal events.

[0150] The operational profile includes one or more of the following: time branch mixed weight, residual magnitude ratio, average edge gating strength, graph gain, seasonal source share, residual source share, and graph source share;

[0151] Temporal branch hybrid weights: The global feature vector is obtained by averaging the historical load sequence and the node covariate sequence in the time and node dimensions; the event tag information is preprocessed to obtain event covariates, which are used as event summaries; the domain identifier information is preprocessed to obtain domain embedding vectors, which are used as domain embeddings; the global feature vector is concatenated with the event summary and the domain embeddings and mapped, and then normalized by Softmax to obtain the temporal branch hybrid weights.

[0152] Domain embedding is a vector corresponding to the domain identifier information obtained by querying the learnable embedding table built into the residual prediction model based on the preprocessed domain identifier information. Different data domains correspond to different domain embedding vectors. The domain identifier information is used to identify the type of data source. The learnable embedding table is randomly generated when the residual prediction model is initialized and is jointly optimized with other network parameters through backpropagation during the training process of the residual prediction model.

[0153] The formula for the residual magnitude ratio is as follows:

[0154] ,

[0155] in, The residual amplitude ratio, To merge residual terms, To prevent the stability constant from being divided by zero, For seasonal baselines;

[0156] The formula for the average side-gating strength is as follows:

[0157] ,

[0158] in, For average edge gating strength, For the set of side-gated samples in the test batch, The reciprocal of the number of samples in the test batch. To normalize the edge gate weights;

[0159] Seasonal source share Residual source share and the share of the source of the image The formula is as follows:

[0160] ,

[0161] ,

[0162] ,

[0163] in, For seasonal sources of quantity, For seasonal baseline items, For residual source components, For time scale sets, For scale Mixed weights, For scale The residual prediction components below, For the source of the image, For the graph correction item, The normalization constant is To prevent the stability constant from being divided by zero, For seasonal source share, For the residual source share, For the source share, and .

[0164] The operational profile includes time branch mixing weights, residual magnitude ratio, average edge gating strength, graph gain, and seasonality, residual, and graph source share, comprehensively explaining the prediction results from four dimensions: time frequency, residual strength, graph activation level, and source composition.

[0165] The operation profile is output in the form of JSON, CSV, database fields, or scheduling system interface messages;

[0166] The formula for multi-node power load forecasting is as follows:

[0167] ,

[0168] in, For nodes In the prediction step The predicted load value, For seasonal baselines, For scale Mixed weights, For scale The residual prediction components below, For graph correction terms.

[0169] In this embodiment, such as Figure 1 As shown, the steps of the state-space power load forecasting method based on seasonal residual diagrams are as follows:

[0170] Step S1: Obtain historical load, meteorological features, calendar features, event features, and node graph structure data for multiple nodes. Preprocessing includes outlier truncation, linear interpolation or forward padding, timestamp alignment, node-based normalization or maximum-minimum normalization, holiday one-heat encoding, weekday type encoding, and temperature normalization. The processed data yields the input tensor.

[0171] Step S2: If a power grid physical topology exists, construct an edge set based on feeder connections, line impedance, electrical distance, transformer affiliation, or voltage level; if a physical topology is missing, construct a statistical correlation graph based on historical load correlation coefficients, dynamic time warping distance, or similarity in electricity consumption behavior. Preferably, retain a Top-level graph for each node. A neighbor, The value ranges from 3 to 20.

[0172] Step S3: For the prediction step The seasonal baseline is obtained by extracting historical load data from the same period. Preferably, when the forecast date is a working day, the same time location of the most recent working day is selected; when the forecast date is a public holiday, the same time location of a historical public holiday or an adjacent weekend is selected.

[0173] Step S4: Input the preprocessed historical load series and seasonal baseline term into the residual prediction model for processing;

[0174] Step S4.1: Construct residual input based on the difference between the preprocessed historical load sequence and the seasonal baseline term, and send the residual input to the three branches: fast, daily, and weekly. The fast branch uses the original sampling step size, the daily branch uses the daily cycle alignment window, and the weekly branch uses the weekly cycle alignment window.

[0175] Step S4.2: According to the formula: and Normalized neighbor messages are generated and graph correction terms are obtained through linear mapping.

[0176] Step S5: Output the prediction and operational profile, based on the formula: Generate the predicted load and output the operational profile, which can be output in JSON, CSV or scheduling system interface field format.

[0177] Example 2

[0178] Based on the same inventive concept as Embodiment 1, this embodiment introduces a computer system including a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the steps of the seasonal residual map state-space power load prediction method.

[0179] Example 3

[0180] This embodiment describes the experimental verification of a seasonal residual diagram state-space power load prediction method based on three publicly available power load datasets.

[0181] The three datasets are the UCI power load dataset, the EWELD industrial load dataset, and the Guangxi Zhuang Autonomous Region residential load dataset. They differ significantly in terms of temporal regularity, node semantics, and graph construction methods, which can comprehensively test the generalization ability of the method of this invention in different scenarios.

[0182] The UCI power load dataset, as a high-dimensional multi-node benchmark dataset, has a graph structure derived from historical load correlation coefficients. It is used to verify the ability of the proposed method to predict loads using statistical coupling relationships in the absence of physical power grid topology information. The EWELD dataset contains equipment and industrial-grade load data, characterized by strong local irregularities and dense similarity relationships. It is used to verify the residual modeling capability of the proposed method under heterogeneous load patterns. The Guangxi Zhuang Autonomous Region resident dataset represents regional residential electricity demand and includes event covariates and calendar covariates. Its load series exhibits a significant seasonal component dominance, used to verify the predictive performance of the proposed method in strongly seasonal scenarios.

[0183] The evaluation metrics are set, including mean absolute error (MAE), root mean square error (RMSE), weighted absolute percentage error (WAPE), and coefficient of determination. MAE is used to measure absolute point prediction error, RMSE is used to amplify the effect of larger deviations, and WAPE is used as a scale-normalized engineering error measure. This is used to measure the model's ability to explain load variability. Preferably, WAPE is used as one of the main evaluation metrics because it has higher numerical stability than MAPE under small load denominator conditions. All reported values ​​are derived from the original output file generated by the experimental script and have not been manually adjusted.

[0184] Baseline models are established, including PlainMamba, Transformer, PatchTST, GRU, LSTM, TCN, STGCN, linear regression, and seasonal naive models. These baseline models cover mainstream deep learning architectures such as recurrent neural networks, temporal convolutional networks, attention mechanism networks, block processing networks, state-space models, and graph neural networks.

[0185] Test results on the UCI dataset show that the optimal metrics for this method are MAE=0.0214, RMSE=0.0990, and WAPE=7.7758. =0.9853, achieving a steady and slight performance improvement compared to PlainMamba, and showing a significant decrease in MAE and WAPE compared to Transformer and PatchTST. On the EWELD dataset, our method achieves MAE=0.1251 and WAPE=29.4435, outperforming recurrent networks, TCN, STGCN, seasonal naive models, and PlainMamba overall. The best accuracy was achieved in linear regression on this dataset, indicating that traditional linear models are competitive when the load coupling is approximately linear. The Guangxi Zhuang Autonomous Region resident dataset exhibits strong seasonal inertia; our method, PlainMamba, and seasonal naive models show similar performance (MAE approximately 0.1645), while the other comparative models show a significant disadvantage in prediction accuracy.

[0186] Figure 4 The multi-model average ranking heatmap shows that this method ranks best or second best across the three datasets, with robust cross-scenario prediction performance. PlainMamba follows closely behind, while Transformer and PatchTST rank lower, and linear regression and seasonal naive statistical baseline rank in the middle. Figure 5 The WAPE histograms for each model are visually apparent. Our method achieves WAPEs of 7.78 and 11.46 for the UCI and Guangxi Zhuang Autonomous Region datasets, respectively, which are generally better than Transformer and PlainMamba. The linear regression model on the EWELD dataset has the lowest WAPE, consistent with the aforementioned analysis.

[0187] Ablation test results as follows Figure 6 As shown, after removing the seasonal anchor module, the MAE of the three datasets increased significantly, with UCI increasing by 132.3%, EWELD by 83.2%, and the Guangxi Zhuang Autonomous Region dataset by 154.2%, proving that the seasonal anchor is the core performance module of this method. Removing graph correction and topology gating only brought a small error fluctuation of -1.4% to 2.5%, indicating that the graph-related module is a bounded constraint correction term and not an accuracy-driven module. Figure 7 To analyze the interpretability of the results of this method on the UCI dataset: the model has the highest daily-scale mixing weight, and the residual and graph correction magnitudes are relatively small. The prediction results can be quantitatively decomposed into seasonal inertia accounting for 82%, residual fluctuation accounting for 10%, and nodal graph coupling correction accounting for 8%, thus making the prediction causes transparent and traceable. Figure 8 Sensitivity analysis of the non-seasonal module further verifies that after removing graph correction, topology gating, and single-scale simplification, the model error fluctuation range is only -2.18% to 2.78%, which is much smaller than the performance degradation caused by removing seasonal anchors. This again confirms that the graph module only plays a limited constraint correction role.

[0188] In summary, this invention is adaptable to prediction scenarios with different topology types and load characteristics, and has advantages in overall prediction accuracy. The architecture design of separating the seasonal benchmark and residuals by seasonal anchor points is the core key to improving prediction accuracy. The normalized side gating graph module can constrain cross-node error propagation and achieve controllable correction. At the same time, the model has the ability to interpret component decomposition, making it more practical for engineering applications.

[0189] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0190] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0191] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0192] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0193] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A seasonal residual diagram state-space power load forecasting method, characterized in that, include: Acquire prediction data for multiple power load nodes, including historical load sequences, node covariate sequences, node graph structure information, edge attribute information, edge type information, domain identifier information, and event marker information; The predicted data is preprocessed; Based on the preprocessed historical load sequence, seasonal anchor points are constructed, and seasonal reference terms for each load node within the prediction interval are generated based on the seasonal anchor points. The preprocessed historical load sequence and the seasonal benchmark term are input into a pre-trained residual prediction model for processing to obtain multi-node power load prediction results.

2. The seasonal residual diagram state-space power load forecasting method according to claim 1, characterized in that, The residual prediction model performs the following processing: The residual input is constructed based on the difference between the preprocessed historical load sequence and the seasonal benchmark term. The residual input is then subjected to multi-scale selective state space scanning to extract the residual fluctuation features at each time scale. The residual fluctuation features at each time scale are then mapped to the corresponding scale decoding head to obtain residual prediction components at multiple time scales. Learnable hybrid weights are obtained through Softmax normalization. The residual prediction components of the multiple time scales are fused based on the learnable hybrid weights to obtain the fused residual term. Based on the preprocessed historical load sequence, node covariate sequence, node graph structure information, and edge attribute information, normalized edge gating weights are calculated, the neighbor node states of each node are obtained, and the neighbor node states are aggregated based on the normalized edge gating weights to obtain graph correction terms. The seasonal baseline term, the fused residual term, and the graph correction term are added together to generate a multi-node power load forecast value; the corresponding operational profile is obtained based on the seasonal baseline term, the fused residual term, the graph correction term, the historical load sequence, the node covariate sequence, the domain identifier information, and the event marker information.

3. The seasonal residual diagram state-space power load forecasting method according to claim 1, characterized in that, The preprocessing includes: The historical load sequence is processed by imputing missing values, truncating outliers, normalizing by time and node, encoding holidays, encoding weekdays, and normalizing by weather features; The edges in the node graph structure information are labeled with types, and the discrete edge type labels are mapped to vectors through an embedding layer; The node covariate sequence is subjected to holiday coding, weekday coding, and weather feature normalization. The edge attribute information is then subjected to numerical normalization and post-processing. The edge type information is discretely labeled and mapped to an edge type embedding vector; The domain identification information is encoded to generate a domain embedding vector; The event tagging information is binarized or hierarchically encoded to generate event covariates.

4. The seasonal residual diagram state-space power load forecasting method according to claim 1, characterized in that, The residual prediction model employs a multi-scale topological conditional selective state-space model, including: The input layer is used to receive the preprocessed multi-node historical load sequence, node graph structure, node covariates, and edge attributes. The multi-scale branching layer includes at least two of the fast-scale branching, daily-scale branching, and weekly-scale branching, and each branch contains an L-layer topological condition-selective state space layer. The graph message passing layer generates bounded graph correction terms based on the node graph structure information by normalizing edge gating weights and injects them into the state updates of each branch. The multi-scale fusion layer is used to fuse the outputs of each branch with learnable Softmax weights and then map them to residual prediction components by the decoder head. The training of the residual prediction model includes: training using Huber loss, and incorporating one or more of the following: aggregated balance error term, residual scaling constraint term, and graph Laplace energy constraint term, as shown in the following formula: ， in, For the total loss, For Huber's losses, To predict load values, This represents the actual load value. To aggregate and balance error weights, This is the aggregation equilibrium error term. As the residual proportion constraint weight, For residual proportion constraint term, The Laplace energy constraint weights are used for the graph. For the graph Laplace energy constraint term; The formula for the Graph Laplace energy constraint is as follows: ， in, To transpose the predicted load value, Let be the graph Laplace matrix constructed using normalized edge weights.

5. The seasonal residual diagram state-space power load forecasting method according to claim 1, characterized in that, The multi-scale selective state-space scanning includes: setting a set of time scales. The set of time scales includes at least two of the following: fast scale, daily scale, and weekly scale; The preprocessed historical load sequence and node covariate sequence together constitute the historical input tensor for each scale. The historical input tensor is converted into a scale-dependent sequence, and node-level hidden states are maintained. Any scale Next node The hidden state satisfies: ， in, For scale Next node At any moment The hidden state, For input-related attenuation gating, For scale Next node In the hidden state of the previous moment, It is the identity matrix. For the input gate vector, For element-wise multiplication, For candidate residual states, To illustrate message gain, For image message; The input-related attenuation gating satisfies: ， in, It is a natural exponential function. The positive step size parameter is generated from the input features. This is the softplus activation function. For scale State transition parameters under the following conditions; The graph message gain satisfies: ， in, For the Sigmoid function, For learnable parameters, This is the preset maximum graph gain.

6. The seasonal residual diagram state-space power load forecasting method according to claim 1, characterized in that, The formula for obtaining the learnable hybrid weights through Softmax normalization is as follows: ， in, For scale Mixed weights, It is a natural exponential function. For scale Learnable weight parameters, For scale Learnable weight parameters, It is a set of time scales.

7. The seasonal residual diagram state-space power load forecasting method according to claim 2, characterized in that, The step of calculating normalized edge gating weights based on the preprocessed historical load sequence, the node covariate sequence, the node graph structure information, and the edge attribute information includes: After performing the preprocessing on the node graph structure information, a discrete edge type label set is obtained, and the discrete edge type labels are mapped to edge type embedding codes. After performing the preprocessing on the historical load sequence, the node features of each node at each time step are obtained, wherein the node features of the current node are used as the target node features, and the node features of the neighboring nodes are used as the neighboring node features. The event tagging information is preprocessed to obtain event covariates, and the event tags include at least one of extreme weather tags, holiday tags, and equipment shutdown tags. The normalized edge gating weights are calculated based on the target node features, neighbor node features, edge attributes, edge type embeddings, and event covariates to determine the unnormalized edge gating values. The gating values ​​of incoming edges to the same destination node are then normalized, as shown in the following formula: ， in, To normalize the edge gate weights, For unnormalized edge gating values, To prevent the stability constant from being divided by zero, Indexing neighboring nodes, For nodes The neighborhood group, For neighboring nodes To the node The unnormalized edge gating value.

8. The seasonal residual diagram state-space power load forecasting method according to claim 2, characterized in that, The operational profile includes one or more of the following: time branch mixed weight, residual magnitude ratio, average side gating strength, graph gain, seasonal source share, residual source share, and graph source share; The global feature vector is obtained by taking the mean values ​​of the historical load sequence and the node covariate sequence in the time dimension and node dimension, respectively. The event tagging information is preprocessed to obtain event covariates, which are used as event summaries. The domain identification information is preprocessed to obtain domain embedding vectors, which are used as domain embeddings. The global feature vector is concatenated with the event summaries and domain embeddings and mapped. Then, the time branch hybrid weights are obtained by Softmax normalization. The domain embedding is a vector corresponding to the domain identification information obtained by querying the learnable embedding table built into the residual prediction model based on the preprocessed domain identification information. Different data domains correspond to different domain embedding vectors. The domain identification information is used to identify the type of data source. The learnable embedding table is randomly generated when the residual prediction model is initialized and is jointly optimized with other network parameters through backpropagation during the training process of the residual prediction model. The formula for the residual amplitude ratio is as follows: ， in, For residual amplitude ratio, To merge residual terms, To prevent the stability constant from being divided by zero, For seasonal baseline items; The formula for the average side-gating strength is as follows: ， in, For average edge gating strength, For the set of side-gated samples in the test batch, The reciprocal of the number of samples in the test batch. To normalize the edge gate weights; The seasonal source share Residual source share and the share of the source of the image The formula is as follows: ， ， ， in, For seasonal sources of quantity, For seasonal baselines, For residual source components, For time scale sets, For scale Mixed weights, For scale The residual prediction components below, For the source of the image, For the graph correction item, The normalization constant is To prevent the stability constant from being divided by zero, For seasonal source share, For the residual source share, For the source share, and .

9. The seasonal residual diagram state-space power load forecasting method according to claim 1, characterized in that, The operation profile is output in the form of JSON, CSV, database fields, or scheduling system interface messages; The formula for the multi-node power load prediction result is as follows: ， in, For nodes In the prediction step The predicted load value, For seasonal baselines, For scale Mixed weights, For scale The residual prediction components below, For graph correction terms.

10. A computer system comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the seasonal residual diagram state-space power load forecasting method as described in any one of claims 1-9.

Images