Multi-dimensional power load prediction method and system based on LSTM
By combining stochastic matrix theory and LSTM networks, the problems of signal-noise separation and time dependency capture in load forecasting are solved, achieving high-precision multidimensional power load forecasting and improving the grid's dispatching and management capabilities.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-04
- Publication Date
- 2026-03-27
AI Technical Summary
Existing load forecasting methods suffer from low prediction accuracy and impact power grid security when dealing with high-dimensional, nonlinear, and strongly coupled power load data. This is due to issues such as difficulty in distinguishing between signals and noise during feature engineering, insufficient ability of the model to capture complex time dependencies, and poor generalization ability for special scenarios.
The random matrix theory is used to separate the signal and noise in multi-source data, and a multi-dimensional power load forecasting model based on LSTM is constructed. The time dependence is captured by the long short-term memory network, and users are trained by clustering algorithm to construct a differentiated load forecasting model.
It improves the accuracy and adaptability of load forecasting, enabling precise load forecasting across different user groups and regions, and supporting refined power grid scheduling and demand-side management.
Smart Images
Figure CN121749115A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of power system load prediction, and particularly relates to a multi-dimensional power load prediction method and system based on LSTM. BACKGROUND
[0002] The penetration rate of intermittent energy sources represented by wind and light in the power system continues to rise, which leads to unprecedented diversity, randomness and volatility of the load side of the power grid. Traditional load prediction methods mainly rely on time series models (such as ARIMA) or shallow machine learning models (such as support vector machines), and these methods face the following serious challenges when dealing with current high-dimensional, nonlinear and strongly coupled load data: 1. "Noise" pollution and "signal" omission of feature engineering: Load is affected by multiple factors such as weather, holidays, user behavior, etc. The existing technology adopts methods such as principal component analysis (PCA) or Pearson correlation coefficient for feature selection. However, the linear dimension reduction method such as PCA has strong assumptions on data distribution, and its basis for selecting principal components (such as cumulative variance contribution rate threshold) is subjective, which makes it difficult to effectively distinguish between real correlation "signals" and random "noises" in the data, and may lead to key weak signals being removed as noise, or noise being mistaken for effective features, thereby polluting the model input and affecting the prediction accuracy.
[0003] 2. Insufficient ability of the model to capture complex time-dependent relationships.
[0004] 3. Poor generalization ability for special scenarios: existing general models often treat holidays and the like as ordinary classification features and fail to deeply analyze the complex nonlinear coupling mechanism between them and weather, user behavior. Therefore, when facing special scenarios such as statutory holidays and consecutive extreme high-temperature weather, the model's generalization ability is severely insufficient due to the sparsity of such samples in the training data, and the prediction error will increase sharply, posing a threat to the safe and stable operation of the power grid.
[0005] In summary, developing a new prediction method that can automatically and robustly extract real power consumption features from high-dimensional data and accurately capture long-term and short-term dynamics and global dependencies in load sequences is a technical problem that needs to be solved in the current power industry. SUMMARY
[0006] In view of the deficiencies in the prior art, the application provides a multi-dimensional power load prediction method and system based on LSTM, which separates signals and noises in multi-source data using random matrix theory to construct input data for the prediction model, and the prediction model uses LSTM to effectively improve the accuracy of load prediction.
[0007] The application provides the following technical solutions: The primary objective of this invention is to provide a multidimensional power load forecasting method based on LSTM, comprising: Acquire historical power load data, historical meteorological data, and historical calendar data from power users, and preprocess the acquired historical data; Preliminary electricity consumption characteristic indicators are extracted from the preprocessed data; Random matrix theory was used to analyze preliminary electricity consumption characteristics in order to construct typical electricity consumption characteristics; A load forecasting model consisting of a hierarchical fusion of long short-term memory networks was constructed, and the load forecasting model was trained using typical electricity consumption characteristics to obtain the trained load forecasting model. Real-time power load data, real-time meteorological data, and real-time calendar data are acquired. After preprocessing, preliminary extraction of electricity consumption characteristic indicators, and analysis using random matrix theory, real-time typical electricity consumption characteristics are obtained. These real-time typical electricity consumption characteristics are then input into the trained load prediction model to obtain the predicted power load value.
[0008] As a further improvement of the present invention, the method of analyzing preliminary electricity consumption characteristic indicators using random matrix theory to construct typical electricity consumption characteristics includes: A feature matrix is constructed based on preliminary electricity consumption characteristic indicators and the number of electricity users; The feature matrix is standardized, and the covariance matrix of the standardized feature matrix is calculated. The covariance matrix is decomposed into eigenvalues to obtain multiple eigenvalues of the covariance matrix and the eigenvectors corresponding to each eigenvalue, so as to project the preliminary electricity consumption characteristics onto the feature space. By comparing the eigenvalues of the covariance matrix with the theoretical boundary of the Marchenko-Pastur law, the principal components of the signal representing the actual electricity consumption pattern are identified and separated to construct typical electricity consumption characteristics.
[0009] By standardizing the feature matrix, the calculation validity of the covariance matrix is ensured. Combined with the Marchenko-Pastur theorem, an objective standard for signal / noise discrimination is provided. This eliminates the limitations of subjective thresholds in traditional feature selection, effectively separates real electricity consumption "signals" from random "noise," and makes the constructed typical electricity consumption features purer and more reflective of the essence of electricity consumption. This improves the quality of model input from the source and solves the problem of signal-noise confusion in traditional feature engineering.
[0010] As a further improvement of the present invention, the eigenvectors corresponding to the eigenvalues of the covariance matrix that are greater than the theoretical boundary of the Marchenko-Pastul law are identified as the principal components of the signal used to construct typical electricity consumption characteristics.
[0011] As a further improvement of the application, the load prediction model comprises an input layer, two long short-term memory network layers and a fully connected output layer connected in sequence; the long short-term memory network layer is used to capture data time dependence.
[0012] As a further improvement of the application, after the typical power consumption feature is constructed, the following is further included: Based on the typical power consumption feature, a clustering algorithm is used to group power users, and an adaptive load prediction model is constructed and trained for different user groups.
[0013] The user is grouped according to the core power consumption mode, so that the model can be trained differently according to the power consumption characteristics of different groups, avoid the insufficient adaptation of the general model to the special power consumption mode, and improve the prediction accuracy of the model in different user groups.
[0014] As a further improvement of the application, the clustering algorithm is a K-means algorithm, and the optimal clustering number is determined by a silhouette coefficient method.
[0015] As a further improvement of the application, the preliminary power consumption feature index includes mean, standard deviation, peak-valley difference, load rate, peak-valley ratio, skewness, kurtosis, trend, autocorrelation, temperature correlation, humidity correlation and holiday effect.
[0016] The preliminary feature index covers multiple dimensions, can comprehensively depict the core attributes of power consumption behavior, provides rich and effective original data for subsequent feature purification of the random matrix theory, avoids omission of key power consumption features, and ensures the integrity and representativeness of the typical power consumption feature.
[0017] The second object of the application is to provide a multi-dimensional power load prediction system based on LSTM, which is used for the above-mentioned prediction method, and characterized in that it comprises a data preprocessing module, a preliminary feature extraction module, a typical feature mining module, a model training module and a load prediction module connected in sequence. The data preprocessing module is configured to obtain historical power load data, historical meteorological data and historical calendar data of power users, and preprocess the obtained historical data. The preliminary feature extraction module is configured to extract preliminary power consumption feature indexes from the preprocessed data. The typical feature mining module is configured to analyze the preliminary power consumption feature indexes by using the random matrix theory to construct a typical power consumption feature. The model training module is configured to construct a load prediction model fused by long short-term memory network layers, train the load prediction model with the typical power consumption feature, and obtain a trained load prediction model. The load prediction module is configured to: acquire real-time power load data, real-time meteorological data and real-time calendar data, obtain real-time typical power consumption characteristics after pre-processing, preliminary power consumption characteristic index extraction and random matrix theory analysis of the real-time data, input the real-time typical power consumption characteristics into the trained load prediction model, so as to obtain a power load prediction value.
[0018] As a further improvement of the application, the system further comprises a clustering module configured between the typical feature mining module and the model training module, and the clustering module is configured to group power users by using a clustering algorithm based on the typical power consumption characteristics.
[0019] As a further improvement of the application, the model training module constructs and trains an adaptive load prediction model for different power user groups based on the power user grouping.
[0020] Compared with the prior art, the application has the following beneficial effects: The application of random matrix theory to power load feature engineering provides a signal and noise separation criterion based on strict mathematical statistics theory, avoids the subjectivity and uncertainty of manually setting a threshold in the traditional method, makes the extracted typical power consumption characteristics more pure and better reflect the essence of the problem, and guarantees the quality of model input data from the source.
[0021] The application provides an end-to-end and systematic solution from data processing, feature mining, user clustering to model prediction. The framework has good modularity and scalability, can be conveniently applied to load prediction tasks of different regions, different user groups and different time granularities, and provides a powerful decision support tool for fine scheduling of power grids, demand side management and future planning. BRIEF DESCRIPTION OF DRAWINGS
[0022] Figure 1 A flowchart of the prediction method provided by the application is shown in the figure; Figure 2 A schematic diagram of the component eigenvalue results obtained by using the random matrix theory for mining is shown in the figure; Figure 3 A schematic diagram of the component explained variance ratio results obtained by using the random matrix theory for mining is shown in the figure; Figure 4 A schematic diagram of the component composition ratio obtained by using the random matrix theory for mining is shown in the figure; Figures 5 to 10 The component analysis diagrams of signal component 1 to signal component 6 are shown in the figures, respectively; Figures 11 to 14 The comparison diagrams of the load prediction values and actual values of the industrial users at the hour, day, week and month scales are shown in the figures, respectively; Figures 15 to 18The comparative chart of the predicted value and the actual value of the load of the residential user at the hour, day, week and month scale, respectively; Figures 19 to 22 The comparative chart of the predicted value and the actual value of the load of the agricultural user at the hour, day, week and month scale, respectively; Figures 23 to 26 The comparative chart of the predicted value and the actual value of the load of the commercial user at the hour, day, week and month scale, respectively. DETAILED DESCRIPTION
[0023] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.
[0024] The present application will be described in further detail below with reference to the drawings: As shown in the figure, the present embodiment provides a multi-dimensional power load prediction method based on LSTM, comprising: Figure 1 Obtaining historical power load data, historical meteorological data and historical calendar data of power users, and preprocessing the obtained historical data.
[0025] The historical meteorological data is high-resolution data, including temperature, humidity, wind speed, light intensity and rainfall; the historical calendar data includes weekdays, weekends, statutory holidays and special weekdays and rest days caused by national rest policies.
[0026] The preprocessing comprises: Data cleaning: using 3σ criterion or isolated forest algorithm to identify and correct abnormal peaks or depressions; Missing value interpolation: for missing data points, the weighted average value of the corresponding time of several cycles (days or weeks) before and after is used for filling to retain the periodicity of the data; Data normalization: using the Min-Max Scaling method to linearly map all numerical features to the [0, 1] interval to eliminate the dimension effect and accelerate the model convergence.
[0027] The preliminary electricity consumption characteristic indexes are extracted from the preprocessed data, and the preliminary electricity consumption characteristic indexes cover multiple dimensions such as electricity consumption scale, volatility, distribution form, time trend and external factor correlation, and specifically include 12 dimensions such as mean, standard deviation, peak-valley difference, load rate, peak-valley ratio, skewness, kurtosis, trend (linear regression slope), autocorrelation (such as first-order autocorrelation coefficient), temperature correlation, humidity correlation, holiday effect (holiday and weekday mean difference rate), etc. The preliminary electricity consumption characteristic indexes at least include two of the above-mentioned 12 dimensions.
[0028] The random matrix theory is used to analyze the preliminary electricity consumption characteristic indexes to construct typical electricity consumption characteristics, specifically: Based on the preliminary electricity consumption characteristic indexes (a total of 12 dimensions) and the number of power users, a characteristic matrix is constructed ; wherein, is the number of users; The characteristic matrix is standardized, and the covariance matrix of the standardized characteristic matrix is calculated ; The covariance matrix is subjected to eigenvalue decomposition to obtain 12 eigenvalues ( ) and eigenvectors corresponding to the eigenvalues; According to the Marchenko-Pastur (M-P) law in the random matrix theory, the theoretical boundary of the eigenvalues of the covariance matrix in the case of completely random data (i.e. pure noise) is calculated , which is determined by the data dimension N and T (here, the total number of characteristics is 12); The actually calculated eigenvalues are compared with the theoretical boundary , if , it indicates that the eigenvalue is a signal principal component carrying true and non-random information, and the eigenvector corresponding to the eigenvalue reveals the core electricity consumption mode of the user group; the remaining eigenvalues are regarded as noise principal components and are discarded.
[0029] Through this method, the original 12 characteristics are projected into a low-dimensional and pure typical electricity consumption characteristic space, and the signal principal component is identified from the space, so that the extracted typical electricity consumption characteristics are more pure and can better reflect the essence of the problem, thereby ensuring the quality of the model input data from the source.
[0030] A load prediction model is constructed by hierarchical fusion of long short-term memory networks, and the load prediction model is trained by the typical electricity consumption characteristics to obtain a trained load prediction model; wherein the load prediction model includes an input layer, two long short-term memory network layers and a fully connected output layer connected in turn.
[0031] The specific functions of each layer of the model are as follows: Input layer: the pre-processed historical data is constructed into a supervised learning sample; LSTM layer: the input sequence is first processed through one or more LSTM layers. The LSTM layer utilizes its unique gating mechanism (input gate, forget gate, output gate) to process at each time step of the input sequence, effectively capturing and encoding the local dynamics and overall time-dependent relationships in the load data, and compressing the sequence information into a series of context-aware hidden state vectors; Output layer: the output of the LSTM encoder is passed through a fully connected layer (or multi-layer perceptron) to finally map to the load prediction value for one or more future time steps.
[0032] Model training includes: dividing the dataset into training set, validation set and test set according to 8:1:1; using optimizers such as Adam, taking root mean square error (RMSE), mean absolute error (MAE), R 2 , maximum error, and mean absolute percentage error (MAPE) as the objective function, training the model end-to-end; by the performance on the validation set, select the optimal hyperparameter combination, such as time step length 28 days, learning rate 0.005, etc.
[0033] Real-time power load data, real-time weather data, and real-time calendar data are obtained. After preprocessing, preliminary electricity consumption feature extraction, and random matrix theory analysis, real-time typical electricity consumption features are obtained. The real-time typical electricity consumption features are input into the trained load prediction model to obtain the power load prediction value.
[0034] Further, in order to improve the adaptability of the prediction model to different electricity user groups, after constructing the typical electricity consumption features and before model construction, based on the typical electricity consumption features, a clustering algorithm is used to group the electricity users, and then an adaptive load prediction model is constructed and trained for different user groups; wherein the clustering algorithm is the K-means algorithm, and the optimal clustering number is determined by the silhouette coefficient method.
[0035] Grouping users according to core electricity consumption patterns enables the model to differentiate training for different groups of electricity consumption characteristics, avoiding the lack of adaptation of general models to special electricity consumption patterns, and improving the prediction accuracy of the model in different user groups.
[0036] Taking the power load prediction of commercial, industrial and other multi-industry users in a certain area of Quzhou City, Zhejiang Province as an application scenario, the implementation steps of the prediction method provided in this embodiment are described in detail.
[0037] 1. Data preparation: The user's electricity load data (15 minutes per point) in 2024, meteorological data (temperature, humidity, wind speed, etc.) and calendar data are collected throughout the year; the data is cleaned, interpolated and normalized; the data set is divided into training set, validation set and test set according to the ratio of 8:1:1.
[0038] 2. Feature engineering: Calculate the mean, standard deviation, peak-to-valley difference and other 12 preliminary features for each user, and input the matrix composed of the 12 features into the typical feature mining module based on random matrix theory, which performs the following operations: Calculate the feature covariance matrix and decompose it to get 12 eigenvalues, which are verified to be in the range of [0.0015, 3.9164]; According to the M-P law, the eigenvalue theoretical boundary of the pure noise matrix is calculated 1.0272; Compare the empirical eigenvalue with the theoretical boundary It is found that there are 7 eigenvalues (3.9164, 2.0416, 1.8348, 1.0020, 1.0016, 0.9946, 0.5730). Three conditions are used to make a comprehensive judgment: for each principal component, if it meets multiple "noise-like" conditions (falling within the MP interval, explaining small variance, and being located after 90% cumulative variance), it is classified as noise; otherwise, it tends to be retained as a signal. The results are: the first three eigenvalues are much larger than the upper bound of MP, which must be signals; the 4th to 7th components, although not all exceed 1.0272, even some eigenvalues fall within or below the MP interval, but because the explained variance is still large and before reaching 90% cumulative variance, they do not meet the "at least two noise conditions", so they are also judged as signals, thus obtaining 7 signal components.
[0039] The results of mining eigenvalues using random matrix theory are shown in Figures 2 to 4 There are 7 signal principal components and 5 noise components; among them, Figure 2 The eigenvalues of each component are shown, Figure 3 The explained variance ratio of each component is shown, Figure 4 The composition ratio of the component is shown.
[0040] Figures 5 to 10 The composition analysis chart of each signal principal component, the horizontal coordinate is 12 basic features, specifically mean, standard deviation, peak-to-valley difference, load rate, peak-to-valley ratio, skewness, kurtosis, trend, autocorrelation, temperature, humidity, holiday effect, and the vertical coordinate is the coefficient of each basic feature; Figures 5 to 10 Each signal component is composed of which basic features, and what is the coefficient of each basic feature.
[0041] Signal component 1: eigenvalue 3.9164, explained variance 32.6%, dominant features include standard deviation, peak-to-valley difference, and trend; Signal component 2: eigenvalue 2.0416, explained variance 17.0%, dominant features include kurtosis, skewness, and load rate; Signal component 3: eigenvalue 1.8348, explained variance 15.3%, dominant features include holiday effect, autocorrelation, and load rate; Signal component 4: eigenvalue 1.0020, explained variance 8.4%, dominant features include temperature correlation, humidity correlation, and peak-to-valley ratio; Signal component 5: eigenvalue 1.0016, explained variance 8.3%, dominant features include peak-to-valley ratio, humidity correlation, and temperature correlation; Signal component 6: eigenvalue 0.9946, explained variance 8.3%, dominant features include peak-to-valley ratio, humidity correlation, and temperature correlation; By analyzing the composition of the signal principal components, these signal principal components can be given physical interpretation, for example: signal component 1 (eigenvalue 3.9164) is mainly contributed by standard deviation, peak-to-valley difference, and trend, which can be named as "electricity consumption scale and volatility" feature; signal component 3 (eigenvalue 1.8348) is mainly contributed by holiday effect and autocorrelation, which can be named as "periodicity and holiday response" feature.
[0042] 3. User clustering: The above 7 purified signal principal components are used as input, and K-means algorithm is used to cluster users, and the optimal clustering number is determined to be 4 by the silhouette coefficient method, and all data is divided into industrial, commercial, residential, and agricultural four typical electricity consumption modes.
[0043] 4. Model prediction: An LSTM model is constructed, and the configuration of the model is as follows: after optimization on the validation set, the best input time step is determined to be 28 days (i.e., using the past 28 days of data to predict future load); the LSTM layer is set to 2 layers with 64 hidden units; the Transformer encoding layer is set to 2 layers with 4 attention heads.
[0044] Model training and prediction: the model is trained on the training set, and the performance is evaluated on the test set, Figures 11 to 26 The comparison between the true value and the predicted value of the load of the model on the test set of industrial, residential, agricultural, and commercial user types is shown, and the specific test results are as follows: As Figures 11 to 14As shown, the model's hourly prediction RMSE, MAE, R², maximum error, and MAPE (%) on the industrial user test set are 1.072, 1.019, 0.943, 1.535, and 0.216%, respectively; daily prediction RMSE, MAE, R², maximum error, and MAPE (%) are 6.153, 4.164, 0.995, 39.233, and 0.772%, respectively; weekly prediction RMSE, MAE, R², maximum error, and MAPE (%) are 27.340, 20.987, 0.909, 107.591, and 3.645%, respectively; and monthly prediction RMSE, MAE, R², maximum error, and MAPE (%) are 30.589, 23.943, 0.885, 115.482, and 4.165%, respectively.
[0045] like Figures 15 to 18 As shown, the model's hourly prediction RMSE, MAE, R², maximum error, and MAPE (%) on the residential user test set are 2.178, 1.900, 0.927, 3.410, and 1.134%, respectively; daily prediction RMSE, MAE, R², maximum error, and MAPE (%) are 2.455, 1.970, 0.996, 9.728, and 1.128%, respectively; weekly prediction RMSE, MAE, R², maximum error, and MAPE (%) are 5.711, 4.261, 0.982, 21.088, and 2.154%, respectively; and monthly prediction RMSE, MAE, R², maximum error, and MAPE (%) are 6.458, 4.153, 0.979, 31.372, and 2.000%, respectively.
[0046] like Figures 19 to 22 As shown, the model's hourly prediction RMSE, MAE, R², maximum error, and MAPE (%) on the agricultural user test set are 2.544, 1.900, 0.351, 4.584, and 1.044%, respectively; daily prediction RMSE, MAE, R², maximum error, and MAPE (%) are 1.070, 0.808, 0.999, 3.916, and 0.633%, respectively; weekly prediction RMSE, MAE, R², maximum error, and MAPE (%) are 3.903, 3.150, 0.976, 9.601, and 2.481%, respectively; and monthly prediction RMSE, MAE, R², maximum error, and MAPE (%) are 10.624, 7.570, 0.787, 36.300, and 5.621%, respectively.
[0047] like Figures 23 to 26As shown, the RMSE, MAE, R², maximum error, MAPE (%) of the model for hourly-level prediction on the commercial user test set are 3.015, 2.630, 0.684, 4.536, and 0.807%, respectively; the RMSE, MAE, R², maximum error, MAPE (%) of daily-level prediction are 17.381, 10.570, 0.976, 55.941, and 2.100%, respectively; the RMSE, MAE, R², maximum error, MAPE (%) of weekly-level prediction are 22.695, 14.013, 0.961, 73.798, and 2.704%, respectively; and the RMSE, MAE, R², maximum error, MAPE (%) of monthly-level prediction are 20.483, 12.433, 0.968, 65.590, and 2.441%, respectively.
[0048] The embodiment provides a multi-dimensional power load prediction system based on LSTM, which is used for the above-mentioned prediction method and comprises a data preprocessing module, a preliminary feature extraction module, a typical feature mining module, a model training module, and a load prediction module connected in sequence. The data preprocessing module is configured to acquire historical power load data, historical meteorological data, and historical calendar data of power users, and preprocess the acquired historical data. The preliminary feature extraction module is configured to extract preliminary power consumption feature indexes from the preprocessed data. The typical feature mining module is configured to analyze the preliminary power consumption feature indexes by using the random matrix theory to construct typical power consumption features. The model training module is configured to construct a load prediction model, wherein the load prediction model comprises an input layer, two long short-term memory network layers, and a fully connected output layer connected in sequence, and the load prediction model is trained by using the typical power consumption features to obtain a trained load prediction model. The load prediction module is configured to acquire real-time power load data, real-time meteorological data, and real-time calendar data, obtain real-time typical power consumption features after preprocessing, preliminary power consumption feature index extraction, and random matrix theory analysis of the real-time data, input the real-time typical power consumption features into the trained load prediction model, and obtain a power load prediction value.
[0049] The system further comprises a clustering module configured between the typical feature mining module and the model training module, wherein the clustering module is configured to group power users by using a clustering algorithm based on the typical power consumption features; and the model training module constructs and trains an adaptive load prediction model for different power user groups based on the grouping of the power users.
[0050] In summary, the application provides an end-to-end and systematic solution from data processing, feature mining, user clustering to model prediction. The framework has good modularity and scalability, and can be conveniently applied to load prediction tasks of different regions, different user groups and different time granularities, providing a powerful decision support tool for fine scheduling of power grids, demand side management and future planning.
[0051] The preferred embodiments of the application have been described above with the preferred embodiments; the application is not limited to this and can be variously changed and modified by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the application shall be included in the protection scope of the application.
Claims
1. A multidimensional power load forecasting method based on LSTM, characterized in that, include: Acquire historical power load data, historical meteorological data, and historical calendar data from power users, and preprocess the acquired historical data; Preliminary electricity consumption characteristic indicators are extracted from the preprocessed data; Random matrix theory was used to analyze preliminary electricity consumption characteristics in order to construct typical electricity consumption characteristics; A load forecasting model consisting of a hierarchical fusion of long short-term memory networks was constructed, and the load forecasting model was trained using typical electricity consumption characteristics to obtain the trained load forecasting model. Real-time power load data, real-time meteorological data, and real-time calendar data are acquired. After preprocessing, preliminary extraction of electricity consumption characteristic indicators, and analysis using random matrix theory, real-time typical electricity consumption characteristics are obtained. These real-time typical electricity consumption characteristics are then input into the trained load prediction model to obtain the predicted power load value.
2. The prediction method according to claim 1, characterized in that, The method employs random matrix theory to analyze preliminary electricity consumption characteristics to construct typical electricity consumption characteristics, including: A feature matrix is constructed based on preliminary electricity consumption characteristic indicators and the number of electricity users; The feature matrix is standardized, and the covariance matrix of the standardized feature matrix is calculated. The covariance matrix is decomposed into eigenvalues to obtain multiple eigenvalues of the covariance matrix and the eigenvectors corresponding to each eigenvalue, so as to project the preliminary electricity consumption characteristics onto the feature space. By comparing the eigenvalues of the covariance matrix with the theoretical boundary of the Marchenko-Pastur theorem, the principal components of the signal representing the actual electricity consumption pattern are identified and separated to construct typical electricity consumption characteristics.
3. The prediction method according to claim 2, characterized in that, The eigenvectors corresponding to the eigenvalues of the covariance matrix that are greater than the theoretical boundary of the Marchenko-Pastul law are identified as the principal components of the signal used to construct typical electricity consumption characteristics.
4. The prediction method according to claim 1, characterized in that, The load forecasting model comprises an input layer, two long short-term memory (LSTM) network layers, and a fully connected output layer connected in sequence; the LSTM network layers are used to capture the time dependence of the data.
5. The prediction method according to claim 1, characterized in that, After constructing typical electricity consumption characteristics, the following is also included: Based on typical electricity consumption characteristics, a clustering algorithm is used to group electricity users, and load prediction models are built and trained for different user groups.
6. The prediction method according to claim 5, characterized in that, The clustering algorithm is the K-means algorithm, and the optimal number of clusters is determined by the silhouette coefficient method.
7. The prediction method according to claim 1, characterized in that, The preliminary electricity consumption characteristics indicators include mean, standard deviation, peak-to-valley difference, load factor, peak-to-valley ratio, skewness, kurtosis, trend, autocorrelation, temperature correlation, humidity correlation, and holiday effect.
8. A multidimensional power load forecasting system based on LSTM, used to implement the forecasting method as described in any one of claims 1 to 7, characterized in that, It includes a data preprocessing module, a preliminary feature extraction module, a typical feature mining module, a model training module, and a load prediction module connected in sequence; The data preprocessing module is configured to: acquire historical power load data, historical meteorological data, and historical calendar data of power users, and preprocess the acquired historical data; The preliminary feature extraction module is configured to extract preliminary electricity consumption characteristic indicators from the preprocessed data. The typical feature mining module is configured to: use random matrix theory to analyze preliminary electricity consumption characteristic indicators in order to construct typical electricity consumption characteristics; The model training module is configured to: construct a load forecasting model by hierarchical fusion of long short-term memory networks, train the load forecasting model with typical electricity consumption characteristics, and obtain the trained load forecasting model. The load forecasting module is configured to: acquire real-time power load data, real-time meteorological data, and real-time calendar data; after preprocessing, preliminary extraction of electricity consumption characteristic indicators, and analysis of random matrix theory, real-time typical electricity consumption characteristics are obtained; and the real-time typical electricity consumption characteristics are input into the trained load forecasting model to obtain the power load forecast value.
9. The prediction system according to claim 8, characterized in that, The system also includes a clustering module configured between the typical feature mining module and the model training module. The clustering module is configured to group electricity users based on typical electricity consumption characteristics using a clustering algorithm.
10. The prediction system according to claim 9, characterized in that, The model training module is based on electricity user grouping, and constructs and trains adapted load prediction models for different electricity user groups.