Electric quantity prediction method and system based on improved multi-scale LSTM model
By constructing a multi-scale LSTM model with a hierarchical tree structure and incorporating a KAN network, the problem of insufficient multi-source feature recognition capability of long short-term memory networks in power prediction is solved, and efficient and accurate power prediction is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-08
- Publication Date
- 2026-04-14
AI Technical Summary
Existing long short-term memory network models lack the ability to identify multi-source features in power prediction, resulting in low prediction accuracy and an inability to balance low data dependence and efficient computation.
A hierarchical tree structure is constructed to extract multi-dimensional features. A multi-scale LSTM model with KAN network is used for power prediction. Power is accumulated from bottom to top through the hierarchical order of the tree structure, which reduces data dependence and computational complexity.
It improves the accuracy of power generation forecasting, reduces reliance on data and computational complexity, adapts to low-data scenarios, and enhances the model's adaptability.
Smart Images
Figure CN121863352A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power prediction technology, and in particular to a power prediction method and system based on an improved multi-scale LSTM model. Background Technology
[0002] As an important component of smart grid operation and energy internet construction, electricity demand forecasting is not only crucial for supporting the power system to achieve planned scheduling, safe and stable operation, and improved economic efficiency, but also helps to smooth peak and valley loads, avoid redundant investment in power facilities, and enable dynamic matching between electricity production and electricity demand.
[0003] Traditional power generation forecasting is typically based on statistical models, such as the autoregressive integral moving average model, which predicts future values by linearly combining historical power generation data. With the introduction of machine learning techniques, deep learning models such as Long Short-Term Memory networks and Transformer architectures, relying on their multi-layer neural network structures, have the advantage of effectively representing the spatiotemporal nonlinear dynamic characteristics in power data, and have gradually occupied the mainstream position in the field of power generation forecasting.
[0004] However, electricity consumption characteristics are complexly influenced by the coupling of multiple dynamic factors such as climate conditions, economic activities, and energy policies, making it difficult for existing Long Short-Term Memory (LSTM) network models to efficiently and accurately infer the evolution of electricity consumption patterns. Furthermore, LSM networks themselves have significant data dependencies; in scenarios with insufficient training data, their generalization ability and predictive stability will be severely limited. In addition, due to their large number of parameters and complex structure, these models often suffer from high computational overhead and inference latency. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of existing technologies that use long short-term memory networks for power prediction, such as lack of multi-source feature recognition capabilities, low prediction accuracy, and inability to simultaneously meet the requirements of low data dependence and efficient computation. This invention provides a power prediction method and system based on an improved multi-scale LSTM model. By constructing a hierarchical tree structure and extracting multi-dimensional features, the complex coupling relationship of multi-source features is characterized, thereby improving the accuracy of subsequent predictions. Furthermore, a multi-scale LSTM model incorporating KAN networks is used to predict power consumption at the lowest level nodes, and the data dependence and computational complexity of power prediction are reduced by summing based on hierarchical relationships.
[0006] The objective of this invention is achieved through the following technical solution: A power prediction method based on an improved multi-scale LSTM model includes: Construct a corresponding hierarchical tree structure for electricity consumption based on the electricity consumption characteristics and basic distribution information of the target area; Based on historical data, multi-dimensional features of the target area are extracted to construct an electricity consumption feature set; Using the electricity consumption feature set as input, the predicted electricity consumption value of each leaf node in the hierarchical tree structure of electricity consumption is obtained by using a multi-scale LSTM model that integrates KAN network. Based on the predicted electricity consumption values of each leaf node, the predicted electricity consumption of each level node is accumulated from bottom to top according to the hierarchical order of the electricity consumption hierarchy tree structure to obtain the overall electricity consumption prediction result of the target area.
[0007] Furthermore, the step of using the electricity consumption feature set as input and predicting the predicted electricity consumption value of each leaf node in the electricity consumption hierarchical tree structure through a multi-scale LSTM model fused with a KAN network includes: The electricity consumption feature set is input into a multi-scale LSTM model. The electricity consumption feature set is split into preset time scales, and key time-series features at each time scale are extracted and fused. Using the fusion result as input to the KAN network, the predicted power consumption value of each leaf node in the power consumption hierarchical tree structure is obtained through the KAN network.
[0008] Furthermore, the step of splitting the electricity consumption feature set by a preset time scale, extracting key time-series features at each time scale, and fusing them includes: The electricity consumption feature set is divided into several sub-time series based on a preset time scale; The multi-scale LSTM model extracts features from the sub-time series corresponding to each time scale through a corresponding gating mechanism, thereby obtaining the key time series features at the corresponding time scale. By using a fully connected layer, key temporal features across all time scales are fused to obtain fused temporal features.
[0009] Furthermore, the step of using the fusion result as input to the KAN network and predicting the predicted electricity consumption value of each leaf node in the electricity consumption hierarchical tree structure through the KAN network includes: Construct a two-layer KAN network, each layer of which is composed of a univariate spline function matrix; The fused temporal features of each leaf node are input into the first layer of the KAN network. The fused temporal features are then subjected to nonlinear transformation based on the corresponding univariate spline function matrix. The fused temporal features after nonlinear transformation are accumulated and summed to obtain the intermediate feature vector. The intermediate feature vector is input into the second-layer KAN network, and a nonlinear transformation is performed through the corresponding univariate spline function matrix. The intermediate feature vectors after the nonlinear transformation are accumulated and summed to obtain the predicted power value of the corresponding leaf node.
[0010] Furthermore, the construction of the corresponding hierarchical tree structure of electricity consumption based on the electricity consumption characteristics and basic distribution information of the target area includes: Define a tree structure, with the total electricity consumption of the target area as the root node, and initialize the node attributes; Based on administrative regions, regional sub-nodes are divided to form a first-level sub-node hierarchy; Under each regional-level sub-node, charging station sub-nodes are divided according to the physical location of the charging station, forming a two-level sub-node hierarchy; Under each charging station site sub-node, charging equipment sub-nodes are divided according to the equipment identifier, and the charging equipment sub-nodes are the leaf nodes of the tree structure. Based on the root node, regional child nodes, charging station site child nodes, and leaf nodes, a hierarchical tree structure of electricity consumption is obtained by combining the corresponding hierarchical distribution, and corresponding social factors and electricity consumption behavior characteristics are embedded in each node.
[0011] Furthermore, the predicted electricity consumption based on each leaf node is accumulated from bottom to top according to the hierarchical order of the electricity consumption hierarchy tree structure to obtain the overall electricity consumption prediction result for the target area, including: Iterate through each charging station's sub-nodes, filter out the corresponding associated leaf nodes through the edge set, sum the predicted power values of the associated leaf nodes, and obtain the predicted power value of each charging station's sub-node. Iterate through each administrative region sub-node, filter out the corresponding associated charging station sub-nodes through the edge set, sum the power prediction values of the associated charging station sub-nodes, and obtain the power prediction value of each administrative region sub-node. By filtering out the associated administrative region child nodes of the root node through the edge set, the power prediction values of the associated administrative region child nodes are summed to obtain the overall power consumption prediction result of the target area.
[0012] Furthermore, the step of extracting multi-dimensional features of the target area based on historical data to construct an electricity consumption feature set includes: Based on preset dimensions, corresponding feature data are extracted, and the correlation coefficient between feature data of each dimension and electricity is calculated; Based on the correlation coefficient calculation results, key features are selected by combining the preset coefficient threshold, and an electricity consumption feature set is constructed based on the selected key features.
[0013] Furthermore, the multi-dimensional features include at least time features, weather features, and derived features.
[0014] A power prediction system based on an improved multi-scale LSTM model, used to perform any of the above-mentioned power prediction systems, includes: The hierarchical processing module is used to construct a corresponding hierarchical tree structure of electricity consumption based on the electricity consumption characteristics and basic distribution information of the target area. The data processing module is used to extract multi-dimensional features of the target area based on historical data and construct an electricity consumption feature set; The prediction module takes the electricity consumption feature set as input, uses a multi-scale LSTM model that integrates the KAN network to predict the electricity consumption of each leaf node in the electricity consumption hierarchical tree structure, and accumulates the predicted electricity consumption of each level node from bottom to top according to the hierarchical order of the electricity consumption hierarchical tree structure to obtain the overall electricity consumption prediction result of the target area.
[0015] The data processing module further includes: The feature filtering unit is used to calculate the correlation coefficient between feature data of each dimension and electricity, and to filter key features based on the correlation coefficient calculation results and preset coefficient thresholds.
[0016] The beneficial effects of this invention are: A hierarchical tree structure for electricity consumption is constructed based on administrative regions, charging station locations, and charging equipment levels, embedding social factors and electricity consumption behavior characteristics. Simultaneously, multi-dimensional features such as time, weather, and derivatives are extracted from historical data. Key features are selected by calculating correlation coefficients to construct an electricity consumption feature set, integrating the electricity distribution of the target area with multiple influencing factors. This accurately depicts the complex coupling relationship between dynamic factors such as climate, economy, and policy, improving the accuracy of subsequent electricity consumption forecasts. Furthermore, a multi-scale LSTM model incorporating a KAN network is employed. The electricity consumption feature set is split into preset time scales, and time-series features at each scale are extracted using a gating mechanism and fused through a fully connected layer. Then, a nonlinear transformation is performed using a KAN network composed of two layers of univariate spline function matrices to achieve electricity consumption prediction at leaf nodes. The final electricity consumption prediction result is obtained by summing the hierarchical order of the constructed tree structure from bottom to top. This approach enhances the model's adaptability to data through multi-scale time-series feature capture and nonlinear mapping, reducing dependence on massive training data. It also avoids redundant calculations caused by complex architectures, reducing the computational complexity of electricity consumption prediction and balancing accurate prediction with efficient computation. Attached Figure Description
[0017] Figure 1 This is a schematic diagram of a process of the present invention; Figure 2 This is a schematic diagram of a hierarchical tree structure for electricity consumption according to an embodiment of the present invention. Detailed Implementation
[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0019] Example: A power prediction method based on an improved multi-scale LSTM model, such as Figure 1 As shown, it includes: Construct a corresponding hierarchical tree structure for electricity consumption based on the electricity consumption characteristics and basic distribution information of the target area; Based on historical data, multi-dimensional features of the target area are extracted to construct an electricity consumption feature set; Using the electricity consumption feature set as input, the predicted electricity consumption value of each leaf node in the hierarchical tree structure of electricity consumption is obtained by using a multi-scale LSTM model that integrates KAN network. Based on the predicted electricity consumption values of each leaf node, the predicted electricity consumption of each level node is accumulated from bottom to top according to the hierarchical order of the electricity consumption hierarchy tree structure to obtain the overall electricity consumption prediction result of the target area.
[0020] Because the electricity consumption characteristics and basic distribution information of the target area naturally possess hierarchical association attributes, and Long Short-Term Memory (LSTM) networks rely on the self-attention mechanism of flat feature sequences for modeling, they cannot identify hierarchical dependencies and semantic relationships between features. This easily leads to confusion of multi-source features and dilution of key features, resulting in weak multi-source feature recognition capabilities. In contrast, the hierarchical tree structure of electricity consumption, through hierarchical mapping of root nodes, intermediate nodes, and leaf nodes, can transform scattered, unstructured multi-source features into a structured hierarchical feature system. This allows the prediction model to accurately capture the correlation logic of features at different levels, improving the prediction model's ability to recognize multi-source features.
[0021] Furthermore, Long Short-Term Memory (LSTM) networks require massive amounts of data for training to achieve convergence of their attention mechanisms. However, in real-world scenarios, there are often issues with limited or missing data samples in specific power consumption units, such as remote transformer areas or niche industry users. The hierarchical association of tree structures enables data reuse and error compensation. When leaf node data is insufficient, historical data from parent nodes can be reused to assist in modeling. The hierarchical structure allows the prediction model to accurately fit a single pattern based on a small amount of data at the fine-grained unit level, without relying on massive amounts of data to cover all mixed features. This effectively reduces the model's dependence on massive amounts of data and meets the needs of low-data scenarios.
[0022] Therefore, in order to improve the prediction accuracy of the subsequent power consumption prediction using the Long Short-Term Memory Network and achieve low data dependence, a corresponding power consumption hierarchical tree structure is first established for the target area, and subsequent power consumption prediction is carried out based on the constructed power consumption hierarchical tree structure.
[0023] The construction of the corresponding hierarchical tree structure for electricity consumption based on the electricity consumption characteristics and basic distribution information of the target area includes: Define a tree structure, with the total electricity consumption of the target area as the root node, and initialize the node attributes; Based on administrative regions, regional sub-nodes are divided to form a first-level sub-node hierarchy; Under each regional-level sub-node, charging station sub-nodes are divided according to the physical location of the charging station, forming a two-level sub-node hierarchy; Under each charging station site sub-node, charging equipment sub-nodes are divided according to the equipment identifier, and the charging equipment sub-nodes are the leaf nodes of the tree structure. Based on the root node, regional child nodes, charging station site child nodes, and leaf nodes, a hierarchical tree structure of electricity consumption is obtained by combining the corresponding hierarchical distribution, and corresponding social factors and electricity consumption behavior characteristics are embedded in each node.
[0024] The electricity consumption characteristics of the target area include information such as time dimension characteristics, user attribute characteristics, load fluctuation characteristics, and external correlation characteristics. The basic distribution information includes geographical and administrative distribution information, power grid topology distribution information, and user and facility distribution information.
[0025] Set the tree structure as ,in, Represents the set of nodes of a tree. This represents the set of edges connecting parent and child nodes. Using the total electricity consumption of the target area as the root node, initialize node attributes including name, type, and a list of child nodes. Nodes at all levels use a unified attribute structure.
[0026] Based on basic distribution information, the target area is divided into several independent regional units according to administrative jurisdiction boundaries. Each sub-node corresponds to the cluster electricity consumption of all charging stations within a certain administrative region, forming a first-level sub-node hierarchy.
[0027] Under each regional-level sub-node, the charging station is further subdivided into secondary sub-nodes according to its physical location. That is, based on the specific geographical coordinates of the charging station, a single regional-level sub-node is divided into several sub-nodes corresponding to a single charging station, and each sub-node represents the total electricity consumption of a single charging station.
[0028] Under each site-level sub-node, charging equipment is broken down into leaf nodes based on its unique identifier, such as device ID, charging pile number, port code, etc., which are the smallest power consumption units in the tree structure. Each leaf node corresponds to the independent power consumption of a single charging device.
[0029] Finally, based on the hierarchical relationships of the root node, region-level child nodes, site-level child nodes, and leaf nodes, a complete hierarchical tree structure for electricity consumption is formed. The constructed hierarchical tree structure for electricity consumption is as follows: Figure 2 As shown, based on the functional positioning and characteristics of each level, corresponding social factors and electricity consumption behavior characteristics are embedded.
[0030] Specifically, the root node embeds macro-social factors such as regional economic growth rate and annual extreme weather frequency, as well as electricity consumption behavior characteristics such as peak-valley difference and load factor of the region. Regional-level sub-nodes embed meso-social factors such as monthly population flow data and the prosperity of leading industries, as well as characteristics such as differences in charging volume between weekdays and weekends and load fluctuations caused by seasonal changes. Site-level sub-nodes embed micro-social factors such as the average daily travel distance of surrounding users and site operation activities, as well as characteristics such as the daily charging duration distribution and equipment idle rate of the site. Leaf nodes embed factors such as the travel habits of users served by the equipment and equipment maintenance cycles, as well as characteristics such as the hourly charging volume fluctuation rate and the distribution of single charging duration.
[0031] By establishing a hierarchical structure and embedding features, the problem of multi-source feature confusion caused by flat feature inputs in long short-term memory networks can be effectively solved, ensuring prediction accuracy.
[0032] Based on this, historical data of the target area is obtained, and through multi-dimensional feature extraction and feature set construction, the scattered and disordered raw data is transformed into structured and interpretable feature vectors. The extracted feature vectors are then used for subsequent model training and power prediction.
[0033] The step of extracting multi-dimensional features of the target area based on historical data and constructing an electricity consumption feature set includes: Based on preset dimensions, corresponding feature data are extracted, and the correlation coefficient between feature data of each dimension and electricity is calculated; Based on the correlation coefficient calculation results, key features are selected by combining the preset coefficient threshold, and an electricity consumption feature set is constructed based on the selected key features.
[0034] The multi-dimensional features include at least time features, weather features, and derived features.
[0035] The time features encompass periodic time attributes such as weekday type and holiday identifiers. Weekday type distinguishes between weekdays and weekends, while holiday identifiers clearly identify statutory holidays and adjusted workdays. Weather features include meteorological data such as daily average temperature and extreme temperatures, with extreme temperatures covering both maximum and minimum temperatures. Derived features supplement information such as the heat accumulation effect in reference electricity consumption behavior, including the average temperature of the previous 3 days and the average temperature of the previous 5 days. The initial feature set is formed by combining these multi-dimensional features to cover key factors affecting electricity consumption.
[0036] Then, based on the Pearson correlation coefficient, the correlation coefficient between the feature data of each dimension and the electricity consumption is calculated, and its expression is: ; in, For the first The feature values of each feature This is the battery level value. and They represent and The first in One element, The total number of feature dimensions. and They represent and The mean. The range is [-1, 1], where 0 indicates no correlation. This indicates a completely positive or negative correlation.
[0037] Key features are selected based on preset coefficient thresholds to construct an electricity consumption feature set. The preset coefficient thresholds can be set according to actual needs.
[0038] After the relevant basic data has been processed and integrated, subsequent model training and prediction can be carried out.
[0039] As the smallest granular unit of the hierarchical tree structure of electricity consumption, the leaf nodes exhibit not only temporal fluctuations at multiple time scales such as hourly and daily, but also a large number of nonlinear correlations. Traditional single-scale LSTM can only capture the temporal patterns of a single time dimension, and its ability to fit nonlinear features is limited. It is prone to prediction bias in leaf nodes due to the omission of fine-grained temporal features and the inability to adapt to nonlinear correlations, which in turn amplifies the overall error through hierarchical accumulation.
[0040] Therefore, this embodiment further employs a multi-scale LSTM model that integrates a KAN network to perform power consumption prediction. The multi-scale LSTM can capture power consumption features at multiple time scales, such as hourly and daily, in parallel, making up for the shortcomings of traditional LSTM single-scale modeling. Meanwhile, the KAN network can approximate complex nonlinear mappings through spline functions, enhancing the fitting ability of the nonlinear relationship between equipment attributes, user behavior, and power consumption. It can accurately capture random and dynamic power consumption patterns and ensure the accuracy of leaf node prediction.
[0041] Furthermore, considering that the prediction error of electricity consumption will be amplified as the prediction granularity is coarsened, if the overall electricity consumption of the region is predicted directly, the local special patterns are easily masked by the averaging effect. The prediction logic of fine-to-coarse construction of tree structure corresponds to the leaf node as the smallest monitorable electricity consumption unit, and its electricity consumption characteristics are more singular and the patterns are more stable. Therefore, a multi-scale LSTM model that integrates KAN network is further used to accurately model the single feature pattern of each leaf node, and electricity consumption prediction is performed for each leaf node separately.
[0042] The step of using an electricity consumption feature set as input and predicting the predicted electricity consumption value of each leaf node in the hierarchical tree structure of electricity consumption through a multi-scale LSTM model fused with a KAN network includes: The electricity consumption feature set is input into a multi-scale LSTM model. The electricity consumption feature set is split into preset time scales, and key time-series features at each time scale are extracted and fused. Using the fusion result as input to the KAN network, the predicted power consumption value of each leaf node in the power consumption hierarchical tree structure is obtained through the KAN network.
[0043] The training and prediction processes of the multi-scale LSTM model with fused KAN network are based on the electricity consumption feature set. In the process of model training and electricity prediction, the multi-scale LSTM model with fused KAN network has the same processing flow for the input data. It first splits the electricity consumption feature set by a preset time scale, extracts the key time series features at each time scale and fuses them, and then uses the fused result as the input of the KAN network to obtain the corresponding prediction result.
[0044] Specifically, the electricity consumption feature set is first divided into a training dataset and a prediction dataset according to time sequence. The training dataset is then divided into a training set and a validation set according to the ratio. The training set is input into a multi-scale LSTM model that integrates the KAN network to learn the temporal correlation and nonlinear mapping relationship between features and electricity consumption. The model training effect is then monitored through the validation set. Furthermore, by adjusting the model parameters, such as the hidden layer dimension of the multi-scale LSTM and the number of spline nodes in the KAN network, overfitting or underfitting is avoided, ultimately forming a multi-scale LSTM model that integrates the KAN network.
[0045] Then, based on the multi-scale LSTM model of the formed fused KAN network, the corresponding prediction dataset is input to obtain the predicted electricity consumption value of each leaf node in the final electricity consumption hierarchical tree structure.
[0046] In the prediction process, feature extraction and fusion are first performed using a multi-scale LSTM mechanism. Specifically, the electricity consumption feature set is split into preset time scales, and key time-series features at each time scale are extracted and fused, including: The electricity consumption feature set is divided into several sub-time series based on a preset time scale; The multi-scale LSTM model extracts features from the sub-time series corresponding to each time scale through a corresponding gating mechanism, thereby obtaining the key time series features at the corresponding time scale. By using a fully connected layer, key temporal features across all time scales are fused to obtain fused temporal features.
[0047] The preset time scales include fine-grained hourly, medium-grained daily, and coarse-grained weekly scales. Minute-level or monthly extreme scales can also be added as needed. Using these preset scales as the basis for segmentation, the predicted dataset of electricity consumption characteristics is split into corresponding sub-time series sequences, so that the electricity consumption patterns of different time granularities are separated from the mixed original sequences, avoiding the dilution of hourly short-term peaks by daily or weekly macro trends, or the obscuring of weekly overall trends by fine-grained fluctuations.
[0048] The multi-scale LSTM part uses the combined effect of branching structure and adaptive gating mechanism to assign independent LSTM branches to subsequences at each time scale, and the gating parameters of each branch are dynamically adjusted according to the corresponding scale characteristics.
[0049] The multi-scale LSTM part handles time-series dependencies through gating mechanisms such as forget gates, input gates, and output gates. The specific computational expression is as follows: ; ; ; ; ; ; in, Indicates the current time step Input, and These represent the current time step. and the previous time step The hidden state of the LSTM and There are two activation functions, one for the forget gate and one for the forget gate. Input gate and output gate , This represents the memory state at the current time step. For the previous time step memory state, Indicates candidate memory units. , , and These are the weight matrices for the forget gate, input gate, output gate, and candidate memory units, respectively. , , and These are the bias vectors for the forget gate, input gate, output gate, and candidate memory units, respectively. This represents the operation of multiplying element by element.
[0050] Then, the fusion of multi-dimensional features is completed based on the fully connected layer, and its calculation expression is: ; in, This represents the hidden state after merging. Indicates the number of time scales. , indicating the first All time steps at each time scale The hidden state, and These are the weight matrix and bias vector for feature fusion.
[0051] While multi-scale LSTM fusion features have captured temporal patterns at different time scales, such as hourly charging peaks, daily weekday differences, and weekly trends, LSTM inherently relies on traditional activation functions like tanh and sigmoid to achieve nonlinear transformations, limiting its ability to fit complex nonlinear mappings. Furthermore, there is a highly complex nonlinear correlation between the energy consumption of charging station leaf nodes and the fused features. For example, under the same hourly charging peak feature, factors such as random charging durations, device charging mode switching, and sudden changes in surrounding short-term passenger flow can all cause nonlinear fluctuations in energy consumption. Multi-scale LSTM fusion features only achieve structured extraction of temporal patterns but struggle to accurately fit such strong nonlinear correlations, easily leading to fitting deviations between features and actual energy consumption. In contrast, the KAN network (Kolmogorov-Arnold Network), based on the nonlinear approximation properties of spline functions, can adaptively fit nonlinear mappings of arbitrary complexity using piecewise smooth spline basis functions, accurately capturing the nonlinear logic between fused temporal features and leaf node energy consumption.
[0052] Therefore, based on the fusion features of multi-scale LSTM, the fusion result is used as the input to a KAN network. The KAN network then predicts the predicted electricity consumption values for each leaf node in the hierarchical tree structure of electricity consumption. The execution steps include: Construct a two-layer KAN network, each layer of which is composed of a univariate spline function matrix; The fused temporal features of each leaf node are input into the first layer of the KAN network. The fused temporal features are then subjected to nonlinear transformation based on the corresponding univariate spline function matrix. The fused temporal features after nonlinear transformation are accumulated and summed to obtain the intermediate feature vector. The intermediate feature vector is input into the second-layer KAN network, and a nonlinear transformation is performed through the corresponding univariate spline function matrix. The intermediate feature vectors after the nonlinear transformation are accumulated and summed to obtain the predicted power value of the corresponding leaf node.
[0053] The KAN layer is defined as a single-variable function. The matrix formed , , ,in and Indicates the number of inputs and outputs.
[0054] With the first Taking the KAN layer as an example, its expression is: ; in, The output of the KAN layer, As the input to the KAN layer, For the first One KAN layer, For the first One KAN layer, This is the first KAN layer.
[0055] In this embodiment, a two-layer KAN network is specifically set up. First, the fusion temporal features of each leaf node obtained by multi-scale LSTM are input into the first-layer KAN network. Based on the corresponding univariate spline function matrix, the fusion temporal features of each leaf node are subjected to nonlinear transformation. After accumulation processing, they are input into the second-layer KAN network for a second nonlinear transformation. Finally, the fusion temporal features of each leaf node are obtained through accumulation processing.
[0056] The computational expression for a two-layer KAN network is: ; Among them, the input dimension of the first layer ,for The dimensions of the first layer's output and the second layer's input are: The output dimension of the second layer is 1, indicating that the output prediction result is output. It is the first-level output node For input nodes Similarly, for a single-variable function, It is a single-variable function of the second level.
[0057] Based on the predicted power values of each leaf node, and relying on the hierarchical relationship characteristics of the tree structure, the fine-grained and accurate prediction of the leaf nodes is transformed into an overall prediction of the entire level and the entire region.
[0058] Specifically, the predicted electricity consumption based on each leaf node is accumulated from bottom to top according to the hierarchical order of the electricity consumption hierarchy tree structure to obtain the overall electricity consumption prediction result for the target area, including: Iterate through each charging station's sub-nodes, filter out the corresponding associated leaf nodes through the edge set, sum the predicted power values of the associated leaf nodes, and obtain the predicted power value of each charging station's sub-node. Iterate through each administrative region sub-node, filter out the corresponding associated charging station sub-nodes through the edge set, sum the power prediction values of the associated charging station sub-nodes, and obtain the power prediction value of each administrative region sub-node. By filtering out the associated administrative region child nodes of the root node through the edge set, the power prediction values of the associated administrative region child nodes are summed to obtain the overall power consumption prediction result of the target area.
[0059] The multi-scale LSTM that integrates KAN has controlled the prediction error to the smallest granular unit, the leaf node. The prediction errors of each leaf node are independent and small in magnitude. The bottom-up accumulation logic can naturally neutralize these scattered fine-grained errors in the hierarchical summarization, reducing the overall prediction error magnitude.
[0060] Furthermore, this hierarchical accumulation method can further enhance the adaptability to low-data scenarios. In low-data scenarios, the prediction error of some leaf nodes may be relatively large, but the hierarchical relationship of the tree structure dilutes these errors during the accumulation process. The prediction results of multiple leaf nodes at the same level will complement each other, and the error of a single node will not dominate the accumulation result of the upper-level node, thus ensuring the accuracy of the final power prediction result.
[0061] Another aspect of this embodiment provides a power prediction system based on an improved multi-scale LSTM model, including: The hierarchical processing module is used to construct a corresponding hierarchical tree structure of electricity consumption based on the electricity consumption characteristics and basic distribution information of the target area. The data processing module is used to extract multi-dimensional features of the target area based on historical data and construct an electricity consumption feature set; The prediction module takes the electricity consumption feature set as input, uses a multi-scale LSTM model that integrates the KAN network to predict the electricity consumption of each leaf node in the electricity consumption hierarchical tree structure, and accumulates the predicted electricity consumption of each level node from bottom to top according to the hierarchical order of the electricity consumption hierarchical tree structure to obtain the overall electricity consumption prediction result of the target area.
[0062] The data processing module further includes: The feature filtering unit is used to calculate the correlation coefficient between feature data of each dimension and electricity, and to filter key features based on the correlation coefficient calculation results and preset coefficient thresholds.
[0063] The hierarchical processing module, data processing module, and prediction module are all equipped with computers, microprocessors, and other devices with corresponding data processing capabilities. They carry relevant algorithm programs for constructing the corresponding electricity consumption hierarchical tree structure, feature extraction, and electricity prediction, which can achieve high-precision and efficient electricity prediction.
[0064] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Other variations and modifications are possible without departing from the technical solutions described in the claims.
Claims
1. A power prediction method based on an improved multi-scale LSTM model, characterized in that, include: Construct a corresponding hierarchical tree structure for electricity consumption based on the electricity consumption characteristics and basic distribution information of the target area; Based on historical data, multi-dimensional features of the target area are extracted to construct an electricity consumption feature set; Using the electricity consumption feature set as input, the predicted electricity consumption value of each leaf node in the hierarchical tree structure of electricity consumption is obtained by using a multi-scale LSTM model that integrates KAN network. Based on the predicted electricity consumption values of each leaf node, the predicted electricity consumption of each level node is accumulated from bottom to top according to the hierarchical order of the electricity consumption hierarchy tree structure to obtain the overall electricity consumption prediction result of the target area.
2. The power prediction method based on the improved multi-scale LSTM model according to claim 1, characterized in that, The process of using a set of electricity consumption features as input and a multi-scale LSTM model fused with a KAN network to predict the predicted electricity consumption values for each leaf node in the hierarchical tree structure of electricity consumption includes: The electricity consumption feature set is input into a multi-scale LSTM model. The electricity consumption feature set is split into preset time scales, and key time-series features at each time scale are extracted and fused. Using the fusion result as input to the KAN network, the predicted power consumption value of each leaf node in the power consumption hierarchical tree structure is obtained through the KAN network.
3. The power prediction method based on the improved multi-scale LSTM model according to claim 2, characterized in that, The process of splitting the electricity consumption feature set into preset time scales, extracting key time-series features at each time scale, and then fusing them includes: The electricity consumption feature set is divided into several sub-time series based on a preset time scale; The multi-scale LSTM model extracts features from the sub-time series corresponding to each time scale through a corresponding gating mechanism, thereby obtaining the key time series features at the corresponding time scale. By using a fully connected layer, key temporal features across all time scales are fused to obtain fused temporal features.
4. The power prediction method based on the improved multi-scale LSTM model according to claim 3, characterized in that, The process of using the fusion result as input to the KAN network and predicting the predicted electricity consumption values for each leaf node in the electricity consumption hierarchical tree structure through the KAN network includes: Construct a two-layer KAN network, each layer of which is composed of a univariate spline function matrix; The fused temporal features of each leaf node are input into the first layer of the KAN network. The fused temporal features are then subjected to nonlinear transformation based on the corresponding univariate spline function matrix. The fused temporal features after nonlinear transformation are accumulated and summed to obtain the intermediate feature vector. The intermediate feature vector is input into the second-layer KAN network, and a nonlinear transformation is performed through the corresponding univariate spline function matrix. The intermediate feature vectors after the nonlinear transformation are accumulated and summed to obtain the predicted power value of the corresponding leaf node.
5. The power prediction method based on the improved multi-scale LSTM model according to claim 1, characterized in that, The construction of the corresponding hierarchical tree structure for electricity consumption based on the electricity consumption characteristics and basic distribution information of the target area includes: Define a tree structure, with the total electricity consumption of the target area as the root node, and initialize the node attributes; Based on administrative regions, regional sub-nodes are divided to form a first-level sub-node hierarchy; Under each regional-level sub-node, charging station sub-nodes are divided according to the physical location of the charging station, forming a two-level sub-node hierarchy; Under each charging station site sub-node, charging equipment sub-nodes are divided according to the equipment identifier, and the charging equipment sub-nodes are the leaf nodes of the tree structure. Based on the root node, regional child nodes, charging station site child nodes, and leaf nodes, a hierarchical tree structure of electricity consumption is obtained by combining the corresponding hierarchical distribution, and corresponding social factors and electricity consumption behavior characteristics are embedded in each node.
6. The power prediction method based on the improved multi-scale LSTM model according to claim 5, characterized in that, The predicted electricity consumption values based on each leaf node are accumulated from bottom to top according to the hierarchical order of the electricity consumption hierarchy tree structure to obtain the overall electricity consumption prediction result for the target area, including: Iterate through each charging station's sub-nodes, filter out the corresponding associated leaf nodes through the edge set, sum the predicted power values of the associated leaf nodes, and obtain the predicted power value of each charging station's sub-node. Iterate through each administrative region sub-node, filter out the corresponding associated charging station sub-nodes through the edge set, sum the power prediction values of the associated charging station sub-nodes, and obtain the power prediction value of each administrative region sub-node. By filtering out the associated administrative region child nodes of the root node through the edge set, the power prediction values of the associated administrative region child nodes are summed to obtain the overall power consumption prediction result of the target area.
7. The power prediction method based on the improved multi-scale LSTM model according to claim 1, characterized in that, The process of extracting multi-dimensional features of the target area based on historical data to construct an electricity consumption feature set includes: Based on preset dimensions, corresponding feature data are extracted, and the correlation coefficient between feature data of each dimension and electricity is calculated; Based on the correlation coefficient calculation results, key features are selected by combining the preset coefficient threshold, and an electricity consumption feature set is constructed based on the selected key features.
8. The power prediction method based on the improved multi-scale LSTM model according to claim 1, characterized in that, The multi-dimensional features include at least time features, weather features, and derived features.
9. A power prediction system based on an improved multi-scale LSTM model, used to execute the power prediction system according to any one of claims 1 to 8, characterized in that, include: The hierarchical processing module is used to construct a corresponding hierarchical tree structure of electricity consumption based on the electricity consumption characteristics and basic distribution information of the target area. The data processing module is used to extract multi-dimensional features of the target area based on historical data and construct an electricity consumption feature set; The prediction module takes the electricity consumption feature set as input, uses a multi-scale LSTM model that integrates the KAN network to predict the electricity consumption of each leaf node in the electricity consumption hierarchical tree structure, and accumulates the predicted electricity consumption of each level node from bottom to top according to the hierarchical order of the electricity consumption hierarchical tree structure to obtain the overall electricity consumption prediction result of the target area.
10. The power prediction system based on the improved multi-scale LSTM model according to claim 9, characterized in that, The data processing module further includes: The feature filtering unit is used to calculate the correlation coefficient between feature data of each dimension and electricity, and to filter key features based on the correlation coefficient calculation results and preset coefficient thresholds.