Power load forecasting method, system equipment and medium based on multi-source heterogeneous data feature fusion

Abstract

The present invention discloses a power load forecasting method, system device and medium based on multi-source heterogeneous data feature fusion, which relates to the technical field of power system load forecasting. The method includes obtaining a power load forecasting data set for preprocessing; respectively performing input feature processing on static data, graph data and time series data based on a multi-layer perceptron, a graph convolutional network and a gated recurrent unit network, and fusing the processed features based on a feature splicing method to obtain comprehensive feature information; constructing a full connection neural network with the comprehensive features as the input, inputting the training set data into the model and iteratively optimizing and adjusting the model parameters and hyperparameters according to the model effect; based on the parameter optimization, adjusting the model, and using the optimal model obtained by evaluation to forecast the power load in a future time period. The method of the present invention is superior to the traditional method in both accuracy and stability, can better adapt to complex and changeable load forecasting scenarios, and has broad application prospects and significant practical value.

Description

Technical Field

[0001] The present invention relates to the technical field of power system load forecasting, and specifically to a power load forecasting method, system equipment and medium based on the feature fusion of multi-source heterogeneous data. Background Art

[0002] Load forecasting is of great significance in power system planning and construction, determination of power market boundary conditions, and balance of power supply and demand relationships. First, in terms of power system planning and construction, accurate load forecasting can help grid companies and power generation enterprises formulate more scientific and reasonable power generation and power transmission and distribution plans, thereby optimizing resource allocation and improving system operation efficiency. In terms of determining power market boundary conditions, load forecasting provides key electricity demand information for market operators, helping them determine the supply and demand balance point of the power market, thereby optimizing market mechanism design and ensuring the healthy operation of the market. In terms of the balance of power supply and demand relationships, through load forecasting, each participating entity can identify the risks of supply-demand imbalance in advance and take corresponding adjustment measures to ensure the stability and reliability of the power system.

[0003] Existing load forecasting models still have some problems and challenges in practical applications: Incomplete input features: Currently, load forecasting models have few feature selections, resulting in limited feature information that the models can learn. Due to insufficient features, the load forecasting performance varies greatly in different scenarios and it is difficult to adapt to the changing market environment. The correspondence between input features and forecasting algorithms is not clear enough: For different data structures in the power grid and power market, existing forecasting algorithms lack certain selection criteria. Different data types and structures have different requirements for forecasting models. The lack of a unified standard limits the applicability and forecasting accuracy of the models. Insufficient combined forecasting ability based on forecasting results: In load forecasting, the combined forecasting of output results often has insufficient ability to extract feature information, especially when predicting load peaks and troughs. Existing models are difficult to effectively capture the extreme values of load changes, affecting the accuracy and reliability of forecasting. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is: the problem that existing models are difficult to effectively capture the extreme values of load changes, affecting the accuracy and reliability of forecasting.

[0006] To solve the above technical problem, the present invention provides the following technical solution: A power load forecasting method based on the feature fusion of multi-source heterogeneous data, including:

[0007] Obtain a power load forecasting data set for preprocessing, and after the preprocessing is completed, divide the entire data set;

[0008] The input features of static data, graph data, and time series data are processed by a multi-layer perceptron, a graph convolutional network, and a gated recurrent unit network respectively. The processed features are fused based on the feature concatenation method to obtain comprehensive feature information;

[0009] A prediction model with the comprehensive features as the input is constructed based on a fully connected neural network, and a loss function and an optimization algorithm are defined. The training set data is input into the model, and the model parameters and hyperparameters are iteratively optimized and adjusted according to the model effect;

[0010] The prediction accuracy of the model is evaluated based on the test set, the model is adjusted based on parameter optimization, and the optimal model obtained from the evaluation is used to predict the power load in the future time period.

[0011] As a preferred scheme of the power load prediction method based on multi-source heterogeneous data feature fusion described in the present invention, wherein: the acquisition of the power load prediction data set for preprocessing includes acquiring the power load prediction data set, which is divided into static multi-variable data, graph data, and time series data according to types;

[0012] Among them, the static multi-variable data includes date type and holiday name; the graph data includes power grid structure information; the time series data includes load historical data, load prediction data, weather historical data, and weather prediction data;

[0013] Data mapping is performed to map character information into numerical information and map power grid structure information into matrix information; data cleaning is performed, and the missing data is filled by forward and backward filling and linear filling; outlier detection is performed by statistical methods and machine learning methods, and outliers are processed by deletion and replacement methods; data normalization is performed by statistical methods, and all input data of the training set is scaled to the range of 0-1.

[0014] As a preferred scheme of the power load prediction method based on multi-source heterogeneous data feature fusion described in the present invention, wherein: the input feature processing includes a static data processing unit, which processes static data through a multi-layer perceptron MLP. The unit contains two fully connected layers, and ReLU activation function is used to extract static features; a graph data processing unit, which uses a graph convolutional network GCN to process power grid structure data. The GCNConv layer is used to aggregate neighbor node information, extract the relationship between nodes, and the input is the node feature matrix and edge connection information, and graph feature representation is obtained through two layers of GCN convolution;

[0015] The time series data processing unit GRU uses GRU to process time series data. The GRU layer learns the dynamic features of the time series, obtains the final hidden state through the GRU network, and extracts time series features through the fully connected layer fc.

[0016] As a preferred solution of the power load forecasting method based on multi-source heterogeneous data feature fusion according to the present invention, wherein: the fusion of the processed features to obtain comprehensive feature information includes static features , graph features , and time series data features ; Use the feature extraction units of static, graph, and time series data, and input various types of data into the corresponding units during forward propagation to obtain the corresponding feature representations;

[0017] Feature splicing, splicing the three types of features together along the feature dimension to form a comprehensive feature representation, and the formula is expressed as:

[0018]

[0019] The splicing formula is:

[0020]

[0021] Comprehensive feature dimension, the dimension of the spliced comprehensive feature is , then the comprehensive feature vector contains the information of all input features;

[0022] Generate a feature fusion layer, use a fully connected layer, and obtain the fused feature representation by linear transformation and non-linear activation of the spliced features;

[0023] Use the prediction output layer to represent the final output layer, map the fused features to the prediction target space, and generate the load forecasting result.

[0024] As a preferred solution of the power load forecasting method based on multi-source heterogeneous data feature fusion according to the present invention, wherein: the construction of a prediction model with the comprehensive feature as the input based on a fully connected neural network and the definition of a loss function and an optimization algorithm include inputting the fused features after being processed by the feature fusion layer into a fully connected neural network to generate a prediction output; assuming that the dimension of the input features is , adopt the structure of a multi-layer fully connected layer, and finally output the predicted value;

[0025] Select the MSE loss function and the Adam optimizer, and set the learning rate; perform forward propagation, input the input features into the model to generate the predicted value; for each new sample , input it into the trained model for forward propagation, and calculate the output of the model The formula is expressed as:

[0026] ,

[0027] wherein, Represents the forward propagation process of the model, which represents the current parameters of the model;

[0028] Calculate the MSE loss between the predicted value and the true load value; After receiving a new sample each time the model performs forward propagation to obtain the predicted value and calculates the loss function The commonly used loss function is the mean squared error MSE; Assume the predicted value is and the target value is The loss function is defined as:

[0029]

[0030] Calculate the gradient of the loss with respect to the output. Backpropagation starts by calculating the gradient from the output layer; Assume the output layer of the neural network is a single predicted value, and the gradient of the loss function with respect to the output is expressed by the formula:

[0031]

[0032] According to the chain rule, next calculate the gradient of each layer of network parameters; Assume the network contains multiple hidden layers, and calculate the gradient of each layer through the chain rule; Assume is the weighted input of the th layer, is the activation value of the th layer, the weight is and the bias is The activation function is ;

[0033] The gradient formula of the output layer is expressed as:

[0034] ,

[0035] where represents the derivative of the activation function;

[0036] For the gradient of the hidden layer, for each layer calculate forward from the last layer as follows:

[0037] ,

[0038] where represents the error term of the th layer, represents the th layer's weight matrix Backpropagation and optimization, calculate the gradient through backpropagation and update the model parameters using an optimizer;

[0039] Calculate the gradients of the weights and biases for each layer. Given the error terms , the gradients of the weights and biases for each layer can be calculated. The formula for the gradient of the weight is expressed as:

[0040]

[0041] Calculate the MSE loss between the predicted value and the true load value; optimize the model through backpropagation, calculate the gradients through backpropagation, and update the model parameters using an optimizer; calculate the gradients of the loss with respect to the model parameters of , let ;

[0042] The formula for the mean of the gradients is expressed as:

[0043]

[0044] The formula for the mean of the squared gradients is expressed as:

[0045]

[0046] The formula for bias correction is expressed as:

[0047]

[0048] The formula for parameter update is expressed as:

[0049] ,

[0050] where represents the current gradient, represents the learning rate, and represent the decay rate, represents a small value, represents the moving average of the gradient, represents the mean of the gradients at the previous time step, represents the mean of the squared gradients, represents the first moment estimate after bias correction, represents the second moment estimate after bias correction, represents the second moment estimate at the previous time step.

[0051] As a preferred embodiment of the power load forecasting method based on multi-source heterogeneous data feature fusion described in the present invention, wherein: the adjustment of the model parameters and hyperparameters includes, if underfitting and overfitting of the model are found on the validation set, optimizing by adjusting the model parameters and hyperparameters; selecting an initial combination of hyperparameters, training the model using the training set and recording the validation set loss;

[0052] Construct a surrogate model based on the evaluated hyperparameter combinations to predict the performance of other hyperparameter combinations, select the next sampling point, and the acquisition function is the expected improvement EI, and the calculation formula is expressed as:

[0053] ,

[0054] where, represents the predicted value of the surrogate model, represents the current best hyperparameters; E represents the expected value of future improvement; update the surrogate model, train the model with the new hyperparameter combination, update the validation set loss, and incorporate the results into the surrogate model, and repeat the iteration until the termination condition is met.

[0055] As a preferred solution of the power load forecasting method based on multi-source heterogeneous data feature fusion described in the present invention, wherein: the use of the evaluated optimal model to predict the power load in the future time period includes defining the safety upper limit parameter of the parameter power load , the risk threshold of system overload , the threshold for judging whether to start the standby power supply ;

[0056] In the low load situation, if the predicted load is less than the parameter , it means that the current load is low; the system starts the energy-saving mode and reduces the load output in non-critical areas; schedules non-critical load equipment, optimizes energy use, and reduces the overall load;

[0057] In the medium load situation, if the predicted load is between the parameters and , the system load is normal but close to the upper limit, and it maintains stable operation; the system continues to maintain normal load operation, but pays attention to the real-time load fluctuation; if there is redundancy in the power grid, start the standby power supply preparation to cope with future load fluctuations;

[0058] In the high load situation, if the predicted load is greater than the parameter , it means that the load is close to or exceeds the carrying capacity of the power system, and start load scheduling; start the standby power supply and allocate the excess load to the standby equipment or area; implement the load shedding strategy, schedule non-critical areas and equipment to reduce the load; alarm and notify relevant personnel, and manual intervention is required;

[0059] If the predicted load exceeds the parameter , judge that the system is facing overload, start the emergency backup power supply and trigger the system protection mechanism, implement load reduction measures for the entire network, start the emergency load cutting procedure to avoid system collapse; trigger the alarm system to provide real-time decision support and visual feedback to the operator;

[0060] When overload and high load situations occur, forecast parameter adjustments are made.

[0061] As a preferred solution of the power load forecasting method based on multi-source heterogeneous data feature fusion described in the present invention, the prediction parameter adjustment includes error optimization and hyperparameter adjustment according to the prediction results and the actual output of the system. Through error feedback, the system dynamically optimizes the prediction model to maintain good prediction accuracy. If the error Less than the preset threshold , indicating that the model is accurate and the current rule continues to be executed; if the error Increases and is less than the threshold , no need to adjust parameters, re-predict, and continue normal operation;

[0062] If the error Larger and greater than the threshold , indicating that the model prediction is inaccurate, and there may be underfitting or overfitting; adjust the learning rate and batch size hyperparameters through online optimization, and use dynamic learning rate adjustment rules.

[0063] As a preferred solution of the power load forecasting system based on multi-source heterogeneous data feature fusion described in the present invention, it includes: a data acquisition module, a data preprocessing module, a power load forecasting model training module, and a power load forecasting model prediction module;

[0064] Data acquisition module, which collects various relevant data from multi-source heterogeneous data, including weather data, historical load data, and power grid structure;

[0065] Data preprocessing module, including non-numerical feature numerical mapping module, data missing and anomaly detection module, and data normalization module; cleans the collected data, removes noise and outliers, and normalizes data of different scales;

[0066] The power load forecasting model training module includes an input data feature processing module and a feature fusion module; the input data feature processing module includes feature extraction units for static, graph and time series data; the system uses multiple algorithms to extract meaningful features from pre-processed data for different data features, and uses the fused feature data to train the forecasting model;

[0067] The power load forecasting model prediction module includes a data import module, a model prediction module, and a model prediction result visualization module. The trained model is applied to actual data for load forecasting, and the system will evaluate and analyze the prediction results to provide basis and support for relevant decisions.

[0068] A computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the power load forecasting method based on multi-source heterogeneous data feature fusion are implemented.

[0069] A computer-readable storage medium stores a computer program thereon. When the computer program is executed by a processor, the steps of the power load forecasting method based on multi-source heterogeneous data feature fusion are implemented.

[0070] Advantages of the present invention: The power load forecasting method based on multi-source heterogeneous data feature fusion provided by the present invention adopts the multi-source heterogeneous data feature fusion technology, which can simultaneously sample influencing factors in multiple dimensions, thereby extracting more valuable information. The most suitable algorithm is used for feature extraction for different types of data, which can capture the potential laws in various types of data to the greatest extent and improve the accuracy and reliability of prediction. The multi-source data feature fusion is not simply data splicing, but deep fusion at the feature level. By fusing data features from different sources, new and more representative features are formed. This fusion method can make full use of the relevance and complementarity between different data sources, enabling the model to more comprehensively understand and capture the complex factors affecting load changes. The load forecasting algorithm based on multi-source heterogeneous data feature fusion is superior to traditional methods in terms of accuracy and stability, can better adapt to complex and changeable load forecasting scenarios, and has broad application prospects and significant practical value. Description of the Drawings

[0071] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0072] Figure 1 It is the overall flowchart of a power load forecasting method based on multi-source heterogeneous data feature fusion provided by the first embodiment of the present invention.

[0073] Figure 2 It is the schematic diagram of module connection of a power load forecasting method based on multi-source heterogeneous data feature fusion provided by the third embodiment of the present invention. Detailed Embodiments

[0074] To make the above objects, features, and advantages of the present invention more apparent and understandable, the following provides a detailed description of the specific embodiments of the present invention in conjunction with the accompanying drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0075] Example 1. Referring to Figure 1 , which is an embodiment of the present invention, provides a power load forecasting method based on multi-source heterogeneous data feature fusion, including:

[0076] S1: Obtain a power load forecasting data set for preprocessing, and after the preprocessing is completed, divide the entire data set.

[0077] Furthermore, the obtaining of the power load forecasting data set for preprocessing includes obtaining the power load forecasting data set, which is divided by type into static multi-variable data, graph data, and time-series data.

[0078] Even further, the static multi-variable data includes date type and holiday name; the graph data includes power grid structure information; the time-series data includes historical load data, load forecasting data, historical weather data, and weather forecasting data.

[0079] Use the Pearson correlation analysis method and the physical characteristics of the load data to select the key indicators required for load forecasting, including static feature data such as date type and holiday name, graph domain data such as power grid structure and transmission capacity, and time-series data such as historical load data, historical weather data, and predicted weather data.

[0080] Even further, perform data mapping to map character information to numerical information and map power grid structure information to matrix information; perform data cleaning, and fill in the missing data by forward and backward filling and linear filling; perform outlier detection through statistical methods and machine learning methods, and handle outliers through deletion and replacement methods; perform data normalization through statistical methods, and scale all input data of the training set to the range of 0-1.

[0081] Even further, the data preprocessing module cleans the collected data to remove noise and outliers to ensure data quality. Next, normalize data of different scales to make the data conform to a unified standard, which is convenient for subsequent feature extraction and model training. Data preprocessing is a key step to ensure data quality and consistency and has an important impact on the final forecasting result.

[0082] Collect the power grid structure and line transmission level information, weather data, and load data within the predicted area, and construct a complete load prediction dataset for model training and validation.

[0083] Based on the data cleaning method, identify the missing values in the dataset. For the available data such as date types and holiday names, perform corresponding filling. For the indispensable missing data such as load data and weather data, perform linear filling or forward filling, and modify the abnormal data.

[0084] Based on statistical methods, normalize the data, scale the feature data to the range of 0 to 1, eliminate the influence of the differences in the order of magnitude and dimension between different features on model training, and thus improve the model prediction accuracy. The normalization method expression is as follows:

[0085] ,

[0086] In the formula, represents the original data, represents the normalized data, represents the minimum value of the original dataset, represents the maximum value of the original dataset.

[0087] It should be noted that the power load prediction system includes the following key modules: the provincial / regional power load prediction related feature data collection module, the data preprocessing module, the power load prediction model training module, and the power load prediction model prediction module. In the data acquisition module, the system collects various relevant data from multi-source heterogeneous data, including weather data, historical load data, power grid structure, etc. Data acquisition is the foundation of the entire system. Through the comprehensive collection of multi-source data, it provides rich information resources for subsequent data processing and analysis.

[0088] S2: Based on the multi-layer perceptron, graph convolutional network, and gated recurrent unit network, respectively perform input feature processing on static data, graph data, and time series data, and fuse the processed features based on the feature splicing method to obtain comprehensive feature information.

[0089] Furthermore, in the feature fusion and model training module, the system uses a variety of algorithms to extract meaningful features from the preprocessed data. Then, use these fused feature data to train the prediction model, and traditional machine learning algorithms or deep learning algorithms can be selected. In this step, the close combination of feature fusion and model training helps to improve the prediction accuracy and generalization ability of the model.

[0090] Furthermore, the input feature processing includes a static data processing module MLP that processes static data through MLP. The module contains two fully connected layers and extracts static features through the ReLU activation function; a graph data processing module GCN that processes power grid structure data using GCN. The GCNConv layer is used to aggregate neighbor node information, extract the relationships between nodes, and the input is the node feature matrix and edge connection information. The graph feature representation is obtained through two layers of GCN convolution.

[0091] Furthermore, a time series data processing module GRU processes time series data. The GRU layer learns the dynamic features of the time series, obtains the final hidden state through the GRU network, and extracts the time series feature update gate through the fully connected layer (fc). . Input time series data , where is the number of time steps.

[0092] The single-step update rule of GRU is as follows:

[0093] ,

[0094] Reset gate

[0095] ,

[0096] Candidate hidden state

[0097] ,

[0098] Hidden state update

[0099] ,

[0100] Among them, 、 、 are the weight matrices of GRU.

[0101] The method based on feature concatenation fuses the processed features to obtain comprehensive feature information, including static features 、graph features 、time series data features ; Use the feature extraction modules of static, graph, and time series data. When propagating forward, input various types of data into the corresponding modules and obtain the corresponding feature representations.

[0102] Feature concatenation, concatenate the three types of features along the feature dimension to form a comprehensive feature representation, and the formula is expressed as:

[0103] ,

[0104] The splicing formula is:

[0105]

[0106] For the comprehensive feature dimension, the comprehensive feature after splicing has a dimension of , then the comprehensive feature vector contains the information of all input features.

[0107] For the feature fusion layer, a fully connected layer is used to obtain the fused feature representation by linearly transforming and non-linearly activating the spliced features. The prediction output layer represents the final output layer, which maps the fused features to the prediction target space to generate the load prediction result.

[0108] The obtained feature vectors of different dimensions are linearly transformed to obtain feature matrices with the same number of rows, and then the matrices are spliced column by column to obtain the fused features as the input for the subsequent network; through continuous iteration and model training of the fully connected neural network, the optimal parameters are obtained to generate the prediction model. The single-step prediction value of the load prediction is obtained, and then the multi-step prediction method is used for the multi-step prediction of the load.

[0109] Furthermore, the trained optimal model is used to predict the power load data for the next day and the next week, and different indicators are used to evaluate the power load prediction model to evaluate the accuracy of the prediction.

[0110] Furthermore, the formula for each indicator is:

[0111] ,

[0112] ,

[0113] ,

[0114] Among them, represents the number of samples of the predicted value and the true value, represents the predicted value, represents the true value;

[0115] It should be noted that in the model prediction module, the trained model is applied to actual data for load prediction. The system will evaluate and analyze the prediction results to ensure the accuracy and reliability of the prediction. Model prediction is the final output of the entire system. Through the prediction of actual data, the system can provide accurate load prediction results to provide basis and support for relevant decisions.

[0116] S3: Build a prediction model with the comprehensive features as the input based on the fully-connected neural network, define the loss function and the optimization algorithm, input the training set data into the model, and iteratively optimize and adjust the model parameters and hyperparameters according to the model effect, and evaluate the fitting effect of the model based on the validation data set.

[0117] Further, the building of the prediction model with the comprehensive features as the input based on the fully-connected neural network and the definition of the loss function and the optimization algorithm include inputting the fused features after being processed by the feature fusion layer into a fully-connected neural network to generate the prediction output; assuming that the dimension of the input features is , a structure with multiple fully-connected layers can be adopted, and finally the predicted value is output.

[0118] Even further, define the network structure. Assume that the fully-connected network contains two hidden layers, and the activation function uses ReLU.

[0119] Even further, the model formula uses to represent the input layer:

[0120] ,

[0121] The formula for the hidden layer is expressed as:

[0122] ,

[0123] where, , .

[0124] The formula for the output layer is expressed as:

[0125] ,

[0126] where, , represents the dimension of the final output for each parameter .

[0127] Select the MSE loss function and the Adam optimizer, and set the learning rate; perform forward propagation, input the input features into the model to generate the predicted value;

[0128] Calculate the MSE loss between the predicted value and the true load value; after receiving each new sample , the model performs forward propagation to obtain the predicted value , and calculate the loss function , and the commonly used loss function is the mean squared error MSE. Assume that the predicted value is , and the target value is , the loss function is defined as:

[0129]

[0130] To calculate the gradient of the loss with respect to the output, backpropagation starts by calculating the gradient from the output layer. Assume that the output layer of the neural network is a single predicted value (regression problem), and the gradient of the loss function (with respect to the output ) is as follows:

[0131]

[0132] This gradient represents the magnitude of the prediction error and reflects the difference between the predicted value and the actual value.

[0133] According to the chain rule, the gradients of the network parameters for each layer are then calculated. Assume that the network contains multiple hidden layers, and the gradients for each layer are calculated using the chain rule. Assume is the weighted input of the th layer, is the activation value of the th layer, the weight is , the bias is , and the activation function is Then:

[0134] Gradient of the output layer:

[0135] ,

[0136] where represents the derivative of the activation function (for the ReLU activation function, if , otherwise 0).

[0137] Gradient of the hidden layer. For each layer starting from the last layer and calculating forward:

[0138] ,

[0139] where represents the error term of the th layer, represents the weight matrix of the th layer. Backpropagation and optimization: Calculate the gradients through backpropagation and update the model parameters using an optimizer.

[0140] Calculate the gradients of the weights and biases for each layer. Given the error term , the gradients of the weights and biases for each layer can be calculated. The gradient formula for the weights is expressed as:

[0141] ,

[0142] Among them, represents the output of the previous layer, represents the error term of the current layer. The gradient formula of the bias is expressed as:

[0143]

[0144] Parameter update is a key step in online learning. After calculating the gradient each time, an optimization algorithm is used to update the model parameters. The most commonly used optimization algorithms include Gradient Descent and its variants, such as Adam, RMSProp, etc.

[0145] Use ordinary gradient descent to update parameters. The update rule of ordinary gradient descent is as follows:

[0146] ,

[0147] ,

[0148] Among them, represents the learning rate, which controls the step size of each parameter update. and represent the weights and biases at the current time step . and represent the gradients of the weights and biases of this layer.

[0149] Among them, represents the error term of the current layer. Further, the mean formula of the gradient is expressed as:

[0150]

[0151] Further, the mean formula of the squared gradient is expressed as:

[0152]

[0153] Further, the bias correction formula is expressed as:

[0154]

[0155] The parameter update formula is expressed as:

[0156] ,

[0157] Among them, represents the current gradient, represents the learning rate, and represent the decay rate, Represents a tiny value, represents the gradient moving average, represents the mean of the squared gradients; represents the model parameters after gradient update.

[0158] Furthermore, the adjustment of the model parameters and hyperparameters includes, if underfitting and overfitting are found on the validation set, optimizing by adjusting the model parameters and hyperparameters; selecting an initial hyperparameter combination, training the model using the training set and recording the validation set loss.

[0159] The adjustment of the model parameters and hyperparameters includes, if underfitting and overfitting are found on the validation set, optimizing by adjusting the model parameters and hyperparameters; selecting an initial hyperparameter combination, training the model using the training set and recording the validation set loss;

[0160] Construct a surrogate model based on the evaluated hyperparameter combinations to predict the performance of other hyperparameter combinations and select the next sampling point. The acquisition function is the expected improvement El, and the calculation formula is expressed as:

[0161] ,

[0162] where, represents the predicted value of the surrogate model, represents the current best hyperparameters; E represents the expected value of future improvement; update the surrogate model, train the model using the new hyperparameter combination, update the validation set loss, and incorporate the results into the surrogate model, repeating the iteration until the termination condition is met.

[0163] It should be noted that train the model using the new hyperparameter combination, update the validation set loss, and incorporate the results into the surrogate model, repeating the iteration until the termination condition is met.

[0164] S4: If there is no underfitting and overfitting phenomenon, evaluate the prediction accuracy of the model based on the test set, adjust the model based on parameter optimization, and use the evaluated optimal model to predict the power load in the future time period.

[0165] Furthermore, the prediction of the power load in the future time period using the evaluated optimal model includes that the optimization objective of the feedback adjustment of the predicted value is to maximize the acquisition function, and the formula is expressed as:

[0166]

[0167] By maximizing the acquisition function, the Bayesian optimization method gradually approaches the global optimal hyperparameter combination for model hyperparameter adjustment; set a judgment threshold for the difference between the validation set and test set losses (such as ), used to judge whether there is an overfitting phenomenon:

[0168]

[0169] If the validation set loss is high, increase the number of nodes in the hidden layer; if the validation set loss is low but the test set loss is high, decrease the number of nodes in the hidden layer.

[0170] If underfitting occurs, increase the number of nodes in the hidden layer to improve the model's expressive ability.

[0171] It can be adjusted using the incremental ratio method or the fixed increment method. The formula for the incremental ratio method is:

[0172] ,

[0173] Fixed increment method:

[0174] ,

[0175] The formula for the decremental ratio method is:

[0176] ,

[0177] The formula for the fixed decrement method is:

[0178]

[0179] The prediction of the power load for a future time period using the optimal model obtained from the evaluation includes defining the safety upper limit parameter of the parameter power load , the risk threshold of system overload , and the threshold for determining whether to start the backup power supply .

[0180] In the case of low load, if the predicted load is less than the parameter , it indicates that the current load is low; the system starts the energy-saving mode to reduce the load output in non-critical areas; schedules non-critical load devices, optimizes energy use, and reduces the overall load.

[0181] In the case of medium load, if the predicted load is between the parameter and , the system load is normal but close to the upper limit, and it maintains stable operation; the system continues to maintain normal load operation but pays attention to real-time load fluctuations; if there is redundancy in the power grid, it starts to prepare the backup power supply to cope with future load fluctuations.

[0182] In the case of high load, if the predicted load is greater than the parameter , indicating that the load is approaching or exceeding the carrying capacity of the power system, initiate load dispatching; start the backup power supply and distribute the excess load to backup equipment or areas; implement load shedding strategies, dispatching non-critical areas and equipment to reduce the load; alarm and notify relevant personnel, and manual intervention is required.

[0183] If the predicted load exceeds the parameter , determine that the system is facing an overload situation, start the emergency backup power supply and trigger the system protection mechanism, implement load shedding measures across the network, start the emergency load cut-off procedure to avoid system collapse; trigger the alarm system to provide real-time decision support and visual feedback to the operator;

[0184] When overload and high load situations occur, adjust the prediction parameters based on the parameter optimization adjustment model.

[0185] The parameter optimization adjustment model includes optimizing the error and adjusting hyperparameters according to the prediction results and the actual output of the system. Through error feedback, the system dynamically optimizes the prediction model to maintain good prediction accuracy; if the error is less than the preset threshold , it means the model is accurate, and continue to execute the current rules; if the error increases and is less than the threshold , there is no need to adjust the parameters, re-predict, and continue normal operation.

[0186] If the error is large and greater than the threshold , it means the model prediction is inaccurate, and there may be underfitting or overfitting; adjust the learning rate and batch size hyperparameters through online optimization, and use dynamic learning rate to adjust the model.

[0187] It should be noted that update the model, apply the adjusted number of hidden layer nodes to the model structure, determine the current model structure and parameters as the final model, and use it for actual prediction tasks.

[0188] Example 2, the following is an embodiment of the present invention, which provides a power load prediction method based on multi-source heterogeneous data feature fusion. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0189] Dataset Division: After preprocessing, the dataset is divided into a training set, a validation set, and a test set, with the proportions being 70%, 15%, and 15% respectively. This division ratio ensures the balance between model training and validation. In the experiment, a multi-layer perceptron (MLP) is used to process static data and extract static feature representations; a graph convolutional network (GCN) is used to process power grid structure data to obtain the feature relationships between nodes; a gated recurrent unit (GRU) is used to process time series data to capture time-dependent features. The features after the above processing are static features, graph features, and time series features respectively. The three types of features are concatenated to form a comprehensive integrated feature representation, which is used as the input for the subsequent fully connected neural network model.

[0190] Using the integrated features as input, a fully connected neural network (FCNN) is constructed for load forecasting. The model is designed with a two-layer hidden layer structure, and the ReLU activation function is selected. During training, the mean squared error (MSE) is used as the loss function, and the Adam optimizer is used for parameter optimization. During training, iterative optimization is used to adjust the model parameters and hyperparameters, including the number of hidden layer nodes, learning rate, etc., to make the model achieve the best fitting effect on the validation set.

[0191] The fitting effect of the model is evaluated on the validation set. If the validation set loss is high, the model's expressive ability is increased by increasing the number of hidden layer nodes; if the validation set loss is low but the test set loss is high, the model complexity is reduced by decreasing the number of hidden layer nodes. After ensuring that the model is neither overfitting nor underfitting, the prediction accuracy of the model is evaluated on the test set to determine the model's stability. Finally, the model based on the best parameter configuration is used to predict the load data for the next week to verify the effectiveness of this method in practical applications.

[0192] During the training process, the mean squared error (MSE) is used as the loss function, and the Adam optimizer is used for parameter optimization. During training, the training set data is input into the model, and the model parameters and hyperparameters are adjusted through multiple iterations to minimize the loss function.

[0193] By continuously adjusting the model's hyperparameters (such as the number of hidden layer nodes, learning rate, etc.), it is ensured that the model achieves the best fitting effect on the validation set. If the loss of the validation set is high, an attempt is made to increase the number of hidden layer nodes to enhance the model's expressive ability; if the validation set loss is low but the test set loss is high, the number of hidden layer nodes is reduced to avoid overfitting caused by an overly complex model.

[0194] After the model training is completed, the fitting effect of the model is evaluated using the validation set. If overfitting or underfitting occurs on the validation set, the model is optimized by adjusting the hyperparameters (such as the number of hidden layer nodes and the learning rate). To ensure that the model has good generalization ability on unknown data, the loss and prediction accuracy on the test set must also meet certain criteria.

[0195] Through the evaluation on the test set, the stability and prediction accuracy of the model are further verified. Finally, the prediction error on the test set is calculated to confirm the applicability of the model.

[0196] When the parameters and hyperparameters of the model are optimized, it is applied to the power load prediction for the next week. Using the input data for the next week, the model gives the corresponding prediction results, which are compared with the actual load data to verify the effectiveness and accuracy of the model in actual prediction.

[0197] The power load prediction method based on multi-layer model fusion of the present invention shows superior accuracy and stability in load prediction. The prediction errors of the training set and the validation set are both lower than 3%. During the training process, by dynamically adjusting the number of hidden layer nodes, the model gradually improves its adaptation ability on the validation set. In addition, the prediction error on the test set also basically remains within 3%, indicating that the model has good generalization effect in actual scenarios and has a certain degree of robustness.

[0198] The experimental data shows that the load error in future prediction is between 2.4% - 2.7%, slightly lower than that in the training and validation stages, verifying the high accuracy of the present method in future data prediction. Compared with the traditional single-model prediction method, the present invention significantly improves the sensitivity to load fluctuations by separately processing static, graph structure, and time-series features. For example, when the wind speed and humidity change, the prediction load error still remains at a low level, showing the strong robustness of the present method to meteorological factors. Traditional methods often have higher errors when facing dynamically changing meteorological factors, while the present method effectively captures the dynamic relationship of time series through the GRU network and can still maintain high prediction accuracy in the case of changing meteorological factors.

[0199] Example 3, as Figure 2 shown, is an embodiment of the present invention, providing a power load prediction system based on multi-source heterogeneous data feature fusion, including a data acquisition module 100, a data preprocessing module 200, a power load prediction model training module 300, and a power load prediction model prediction module 400.

[0200] The data acquisition module 100 collects various relevant data from multi-source heterogeneous data, including weather data, historical load data, and power grid structure.

[0201] The data preprocessing module 200 includes a non-numerical feature numerical mapping module 201, a data missing and anomaly detection module 202, and a data normalization module 203; it cleans the collected data, removes noise and outliers, and normalizes data of different scales.

[0202] The power load prediction model training module 300 includes an input data feature processing module 301 and a feature fusion module 302; the input data feature processing module includes feature extraction units for static, graph, and time-series data; the system uses multiple algorithms to extract meaningful features from the preprocessed data in different data features, and uses the fused feature data to train the prediction model.

[0203] The power load prediction model prediction module 400 includes a data import module 401, a model prediction module 402, and a model prediction result visualization module 403. It applies the trained model to actual data for load prediction, and the system will evaluate and analyze the prediction results to provide a basis and support for relevant decisions.

[0204] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods in various embodiments of the present invention. The aforementioned storage medium includes: USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs, etc., which can store program codes.

[0205] The logic and / or steps represented in the flowchart or described in other ways herein, for example, can be considered as a definite sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch and execute instructions from the instruction execution system, apparatus, or device), or used in combination with these instruction execution systems, apparatus, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device.

[0206] More specific examples (a non-exhaustive list) of computer-readable media include the following: electrical connections (electronic devices) having one or more wirings, portable computer diskettes (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber devices, and portable compact disc read-only memory (CDROM). Additionally, the computer-readable media can even be paper or other suitable media on which a program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpretation, or, if necessary, other suitable processing, and then storing it in a computer memory.

[0207] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc. It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

[0208] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A power load forecasting method based on multi-source heterogeneous data feature fusion, characterized in that Including: Obtain a power load prediction dataset for preprocessing. After the preprocessing is completed, divide the entire dataset; Based on a multi-layer perceptron, a graph convolutional network, and a gated recurrent unit network, perform input feature processing on static data, graph data, and time series data respectively. Based on the method of feature splicing, fuse the processed features to obtain comprehensive feature information; Construct a prediction model with the comprehensive features as the input based on a fully connected neural network, and define a loss function and an optimization algorithm. Input the training set data into the model and iteratively optimize and adjust the model parameters and hyperparameters according to the model effect; Evaluate the prediction accuracy of the model based on the test set, adjust the model based on parameter optimization, and use the optimal model obtained from the evaluation to predict the power load in the future time period; Further, using the optimal model obtained from the evaluation to predict the power load in the future time period includes that the optimization objective of feedback adjustment of the predicted value is to maximize the acquisition function, which is expressed by the formula: By maximizing the acquisition function, the Bayesian optimization method gradually approaches the global optimal hyperparameter combination for adjusting the model hyperparameters; set a judgment threshold ∈ for the difference between the validation set loss and the test set loss, ∈ = 0.05, to determine whether there is an overfitting phenomenon: |L valid -L test |>∈ If the validation set loss is high, increase the number of hidden layer nodes; if the validation set loss is low but the test set loss is high, decrease the number of hidden layer nodes; If underfitting occurs, increase the number of hidden layer nodes to improve the model's expressive ability; It can be adjusted using the incremental ratio method or the fixed increment method. The formula for the incremental ratio method is expressed as: Fixed increment method: The formula for the decremental ratio method is expressed as: The formula for the fixed decrement method is expressed as: Defining the loss function and optimization algorithm includes calculating the MSE loss between the predicted value and the true load value; upon receiving a new sample {x t ,y t}, the model performs forward propagation to obtain the predicted value and calculates the loss function L t , where the loss function is the mean squared error MSE; assuming the predicted value is the target value is y t , the loss function L t is defined as: To calculate the gradient of the loss with respect to the output, backpropagation first calculates the gradient starting from the output layer; assuming that the output layer of the neural network is a single predicted value, the gradient of the loss function with respect to the output The formula is expressed as: According to the chain rule, the gradients of the network parameters for each layer are calculated next; assume the network contains multiple hidden layers, and the gradients for each layer are calculated using the chain rule; assume z (l) is the weighted input of the l-th layer, a (l) is the activation value of the l-th layer, the weight is W (l) and the bias is b (l) , and the activation function is σ; The gradient formula of the output layer is expressed as: where, σ′(z (L) ) represents the derivative of the activation function; For the gradient of the hidden layer, calculate forward from the last layer L for each layer l: δ (l) =(W (l+1) ) T δ (l+1) ·σ′(z (l) ) Among them, δ (l) represents the error term of the l-th layer, and W (l+1) represents the backpropagation and optimization of the weight matrix of the (l + 1)-th layer. Calculate the gradient through backpropagation and update the model parameters using an optimizer; Calculate the gradients of the weights and biases for each layer. Given the error term δ (l) , the gradients of the weights and biases for each layer can be calculated. The gradient formula for the weights is expressed as: Calculate the MSE loss between the predicted value and the true load value; optimize the model through backpropagation, calculate the gradient through backpropagation, and update the model parameters using an optimizer; calculate the gradient of the loss with respect to the model parameter θ Let The formula for the mean of the gradient is expressed as: m t = β1m t-1 + (1 - β1)g t The formula for the mean of the square of the gradient is expressed as: The formula for bias correction is expressed as: The formula for parameter update is expressed as: Among them, g t represents the current gradient, α represents the learning rate, β1 and β2 represent the decay rates, ∈ represents a small value, m t represents the moving average of the gradient, m t-1 represents the mean of the gradients at the previous time step, v t represents the mean of the squared gradients, represents the first moment estimate after bias correction, represents the second moment estimate after bias correction, v t-1 represents the second moment estimate at the previous time step; The parameter-optimized adjustment model includes error optimization and hyperparameter adjustment based on the prediction results and the actual output of the system. Through error feedback, the system dynamically optimizes the prediction model to maintain good prediction accuracy; if the error ∈ t is less than the preset threshold ∈ th , it indicates that the model is accurate and the current rule continues to be executed; if the error ∈ t increases and is less than the threshold ∈ th , there is no need to adjust the parameters, re-predict, and continue normal operation; If the error ∈ t is large and greater than the threshold value ∈ th , it indicates that the model prediction is inaccurate and there may be underfitting or overfitting; adjust the learning rate and batch size hyperparameters through online optimization, and use dynamic learning rate to adjust the model.

2. The power load forecasting method based on multi-source heterogeneous data feature fusion according to claim 1, characterized in that: The preprocessing of obtaining the power load prediction dataset includes obtaining the power load prediction dataset, which is divided by type into static multivariate data, graph data, and time series data; Among them, the static multivariate data includes date type and holiday name; the graph data includes power grid structure information; the time series data includes load historical data, load prediction data, weather historical data, and weather prediction data; Perform data mapping to map character information to numerical information and map the power grid structure information to matrix information; perform data cleaning, and fill in the missing data by forward and backward filling and linear filling; perform outlier detection through statistical methods and machine learning methods, and process outliers through deletion and replacement methods; perform data normalization processing through statistical methods to scale all input data of the training set to the range of 0 - 1.

3. The power load forecasting method based on multi-source heterogeneous data feature fusion according to claim 1 or 2, characterized in that: The input feature processing includes a static data processing unit that processes static data through a multi-layer perceptron (MLP). The unit contains two fully connected layers and extracts static features through a ReLU activation function; a graph data processing unit that processes power grid structure data using a graph convolutional network (GCN). The GCNConv layer is used to aggregate neighbor node information, extract the relationships between nodes, with the input being the node feature matrix and edge connection information, and obtaining the graph feature representation through two layers of GCN convolution; a time series data processing unit that processes time series data using a gated recurrent unit network (GRU). The GRU layer learns the dynamic features of the time series, obtains the final hidden state through the GRU network, and extracts time series features through a fully connected layer (fc).

4. The power load forecasting method based on multi-source heterogeneous data feature fusion according to claim 1, wherein: The fusion of the processed features to obtain comprehensive feature information includes static Feature H static 、Graph Feature H graph 、Time Series Data Feature H temporal ; Using feature extraction units for static, graph, and time series data, input various types of data into the corresponding units during forward propagation to obtain corresponding feature representations; feature concatenation, where the three types of features are concatenated along the feature dimension to form a comprehensive feature representation. The formula is expressed as: H combined = concat(H static , H graph , H temporal ) The concatenation formula is: H combined = [H static | H graph | H temporal ​ Comprehensive feature dimension, the combined comprehensive feature H combined has a dimension of N × (3·d hidden ), then the comprehensive feature vector contains the information of all input features; generating a feature fusion layer that uses a fully connected layer to obtain the fused feature representation through linear transformation and non-linear activation of the concatenated features; using a prediction output layer to represent the final output layer, mapping the fused features to the prediction target space to generate the load prediction result.

5. The power load forecasting method based on multi-source heterogeneous data feature fusion according to claim 4, wherein: The prediction model constructed based on the fully connected neural network with the comprehensive feature as the input includes inputting the fused feature H combined after being processed by the feature fusion layer into a fully connected neural network to generate a prediction output; assuming that the dimension of the input feature is d input , adopting the structure of multiple fully connected layers, and finally outputting the predicted value; Selecting the MSE loss function and Adam optimizer and setting the learning rate; performing forward propagation by inputting the input features into the model to generate prediction values; For each new sample x t , input it into the pre-trained model for forward propagation and calculate the output of the model The formula is expressed as: Among them, f(x t ; θ) represents the forward propagation process of the model, and θ represents the current parameters of the model.

6. The power load forecasting method based on multi-source heterogeneous data feature fusion according to claim 5, wherein: The adjustment of model parameters and hyperparameters includes optimizing by adjusting model parameters and hyperparameters if underfitting and overfitting are found on the validation set; selecting an initial hyperparameter combination, training the model using the training set, and recording the validation set loss; constructing a surrogate model based on the evaluated hyperparameter combinations to predict the performance of other hyperparameter combinations and selecting the next sampling point. The acquisition function is the expected improvement (EI), and the calculation formula is expressed as: El(θ) = E[max(f(θ + ) - f(θ), 0)] Among them, f(θ) represents the predicted value of the surrogate model, and θ + represents the current best hyperparameter; E represents the expected value of future improvement; update the surrogate model, train the model with a new combination of hyperparameters, update the validation set loss, incorporate the results into the surrogate model, and repeat the iteration until the termination condition is met.

7. The power load forecasting method based on multi-source heterogeneous data feature fusion according to claim 1, wherein: The prediction of the power load for future time periods using the evaluated optimal model includes defining the safety upper limit parameter P1 of the power load, the risk threshold P2 of system overload, and the threshold P3 for determining whether to start the standby power supply; Under low load conditions, if the predicted load is less than parameter P1, it indicates that the current load is low; the system starts the energy-saving mode, reduces the load output in non-critical areas; schedules non-critical load devices, optimizes energy use, and reduces the overall load; Under medium load conditions, if the predicted load is between parameters P1 and P2, the system load is normal but close to the upper limit, and stable operation is maintained; the system continues to operate at normal load, but pays attention to real-time load fluctuations; if there is redundancy in the power grid, prepare to start the standby power supply to cope with future load fluctuations; Under high load conditions, if the predicted load is greater than parameter P2, it indicates that the load is approaching or exceeding the carrying capacity of the power system, and load dispatching is initiated; start the standby power supply and allocate the excess load to standby equipment or areas; implement a load shedding strategy to schedule non-critical areas and equipment to reduce the load; give an alarm and notify relevant personnel, requiring manual intervention; If the predicted load exceeds parameter P3, it is determined that the system is facing an overloaded situation. The emergency backup power supply is started and the system protection mechanism is triggered to implement load shedding measures across the network, start the emergency load cut-off procedure to avoid system collapse; trigger the alarm system to provide real-time decision support and visual feedback to the operator; when overloading and high load conditions occur, adjust the prediction parameters and optimize the model based on parameter optimization.

8. A system for a power load forecasting method based on multi-source heterogeneous data feature fusion, characterized in that: It includes a data acquisition module (100), a data preprocessing module (200), a power load prediction model training module (300), and a power load prediction model prediction module (400); The data acquisition module (100) collects various relevant data from multi-source heterogeneous data, including weather data, historical load data, and power grid structure; The data preprocessing module (200) includes a non-numerical feature numerical mapping module (201), a data missing and anomaly detection module (202), and a data normalization module (203); cleans the collected data, removes noise and outliers, and normalizes data of different scales; The power load forecasting model training module (300) includes an input data feature processing module (301) and a feature fusion module (302); the input data feature processing module includes feature extraction units for static, graph, and time-series data; in different data features, the system uses multiple algorithms to extract meaningful features from the preprocessed data, and uses the fused feature data to train the forecasting model; The power load forecasting model prediction module (400) includes a data import module (401), a model prediction module (402), and a model prediction result visualization module (403). The trained model is applied to actual data for load forecasting, and the system will evaluate and analyze the prediction results to provide basis and support for relevant decisions.

9. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, the steps of the power load forecasting method based on multi-source heterogeneous data feature fusion are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, the steps of the power load forecasting method based on multi-source heterogeneous data feature fusion are implemented.

Images