CN-122026327-A - Intra-provincial power load prediction method based on multi-modal information fusion analysis

Abstract

The invention relates to the technical field of electric quantity prediction and discloses an intra-provincial power load prediction method based on multi-mode information fusion analysis, which comprises the steps of preprocessing collected historical load data and market data, extracting the Seasonal features and trend trend features of the data through a time sequence information extraction module, and extracting the public opinion features of market public opinion corpus through a TextCNN model; and then, performing multi-angle analysis on the three types of features through a multi-head attention module, obtaining a combined feature vector through normalization, affine transformation, dynamic weight calculation and weighted feature fusion, inputting the combined feature vector into a gate control unit for processing, outputting a prediction result through a feedforward neural network, and finally, measuring the prediction accuracy by adopting MAE and RMSE indexes. The invention combines multidimensional features, adopts a mode of combining deep learning and machine learning, effectively improves the precision, stability and anti-interference capability of power load prediction, and provides reliable support for the operation and management of a power system.

Inventors

• ZHANG YUNXIN
• LV SHILEI

Assignees

• 天津大学

Dates

Publication Date

20260512

Application Date

20260203

Claims (5)

1. 1. The method for predicting the power load in the province based on the multi-mode information fusion analysis is characterized by comprising the following steps of: S1, data collection and preprocessing, namely collecting historical load data, market data and market public opinion corpus information, and adopting maximum and minimum normalization processing to the historical load data, wherein a preprocessing formula is as follows: Wherein, the For the i-th data after normalization, As a result of the original data set, As a minimum value for the data set, The market data comprises meteorological data such as time-by-time temperature, humidity, wind speed, sunlight and the like, city economic data, power transmission line data and holiday arrangement information; When the data is preprocessed, the date information is encoded into discrete values, text news information and industry information are collected, corpus is segmented and word vectorization is carried out, and a data set containing various features is constructed; S2, model training data division, namely dividing a preprocessed data set into a training set and a verification set according to the proportion of 8:2, selecting a Mean Square Error (MSE) function as a loss function of model training, reducing errors by adjusting model parameters, and storing the model parameters when the error of the verification set is the lowest; s3, deep learning model construction and feature extraction, wherein the deep learning model comprises a time sequence information extraction module and a multi-head attention module, and the time sequence information extraction module is used for extracting a Seasonal feature of the seal and a trend trend feature: S3.1. Extracting the characteristics of the serial, namely calculating the local characteristics of the load through convolution kernel dimensions [1, T ] and convolution with the step length of 1, and calculating the global characteristics of the load through convolution kernel dimensions [1, T ] and convolution with the step length of a sliding window of T; Adopting a two-dimensional convolution operation with a convolution kernel dimension of [1, feature_size ] and a sliding window step length of 1, wherein feature_size represents the total number of feature dimensions in input data and is used for dynamically distributing weights for different market information features so as to inhibit the influence of unimportant features; The local and global features with different scales are aligned and spliced, so that the model can be more effectively focused on key feature information; the calculation formula of the one-dimensional convolution operation is as follows: As the i-th element of the convolution result, For the t-th time point element of the input sequence, The i-T elements of the convolution kernel, T is the sequence period; the two-dimensional convolution operation calculation formula is as follows: Wherein, the As a result element of the two-dimensional convolution, In order to input the value of the value, An element that is a two-dimensional convolution kernel; Merging alignment convolution results: Wherein, the Is the calculation result of different convolutions; s3.2. The tree feature extraction, namely removing non-stationary information in the input sequence, particularly the mean value and standard deviation of the examples, and returning the input normalized deleted information to the model to realize the inverse normalization operation of the model. Specifically, the sequence is decomposed into the following components by a time sequence information extraction module: In the formula, For the observation at time t, A trend component representing time t; A seasonal component representing time t; A residual term representing the time t; Trend component The calculated expression of (2) is: Seasonal component The calculated expression of (2) is: In the formula, 、 、 The y values at times t-1, t-v +1, t-v respectively, Representing a data set Is the observed value of the i-th time point in the (b), Representing a data set Average of all observations in (a); Representing the average of all time points in the data set; indicating seasonal period lengths; Respectively representing a trend constant term and a trend slope; Indicating the point in time corresponding to the i-th observation. S4, extracting public opinion features, namely converting market public opinion corpus information into word vectors through an embedding layer; after TextCNN model training, inputting the coding result into a full-connection layer, and outputting public opinion characteristics through a sigmoid function; S5, multi-head attention calculation, wherein the multi-head attention module carries out multi-head attention calculation on trend features extracted in S3, seasonal features and public opinion features extracted in S4, and calculates query, key and value matrix Wherein, among them, 、 And Is a weight matrix, and further calculates an attention score, expressed as: Wherein, the Is the dimension of the query and key vector by Calculating a multi-headed attention, wherein Is a matrix of weights that are to be used, Is an output projection matrix, passes through a feedforward neural network layer The treatment, wherein, Is a matrix of weights that are to be used, , Is a bias vector; s6, feature fusion and prediction, namely normalizing and affine transformation are carried out on the three output features, and dynamic weights among the features are calculated ; Wherein, the , Respectively represent a Seasonal feature, a public opinion feature and trend features, Is a learnable projection matrix; is the mean and standard deviation of the features, Is a learnable scaling and offset parameter by A weighted feature fusion is performed, where t represents a time step, The fused characteristic is input into a gate control unit for processing, and the gate control is calculated as: The state update formula is The gate control output is calculated as Wherein, the , , Respectively representing an input gate, a forget gate, an output gate and a candidate state, Representing a sigmoid activation function, tanh representing a hyperbolic tangent activation function, A state vector representing the last moment in time, Representing the bias vector, and finally outputting the prediction result through the feedforward neural network ; S7, estimating the prediction accuracy, namely measuring the prediction accuracy by using an average absolute error (MAE) index and a Root Mean Square Error (RMSE) index, wherein the calculation formulas are respectively as follows: Wherein, the And The real value and the predicted value at the moment i are respectively, and n is a number of data.
2. 2. The method for predicting the power load in the province based on the multi-modal information fusion analysis of claim 1, wherein the market data in S1 comprises weather data such as time-by-time temperature, humidity, wind speed, sunlight and the like, city economic data, power transmission line data and holiday arrangement information, and the date information is encoded into category variables during data preprocessing to construct a data set containing various characteristics.
3. 3. The method for predicting the intra-provincial power load based on multi-modal information fusion analysis as set forth in claim 1, wherein the one-dimensional convolution operation calculation formula in S3.1 is as follows , wherein, As the i-th element of the convolution result, For the t-th time point element of the input sequence, The i-T element is convolution kernel, T is sequence period, and the two-dimensional convolution operation calculation formula is , wherein, As a result element of the two-dimensional convolution, In order to input the value of the value, Merging the aligned convolution results to obtain , 、 Is the result of the calculation of the different convolutions.
4. 4. The method for predicting the intra-provincial power load based on the multi-modal information fusion analysis of claim 1, wherein the public opinion corpus information in S4 is collected through a plurality of channels and is converted into word vectors after preprocessing, and the word vectors contain information of market positive, negative or neutral attitudes.
5. 5. The method for predicting the intra-provincial power load based on the multi-modal information fusion analysis of claim 1, wherein in the model training process, the number of hidden layer neurons is set to be 64, the learning rate is set to be 0.001, and an Adam optimizer is adopted for parameter updating.

Description

Intra-provincial power load prediction method based on multi-modal information fusion analysis Technical Field The invention belongs to the technical field of electric quantity prediction, and particularly relates to an intra-provincial power load prediction method based on multi-mode information fusion analysis. Background In the complex architecture of a modern power system, short-term load prediction plays an irreplaceable role as a key link for guaranteeing stable and efficient operation of the system. The short-term load prediction usually takes hours and days as units, has profound significance, not only provides core basis for the power dispatching department to reasonably arrange a power generation plan and optimize a power grid operation mode, realizes accurate matching of power supply and demand by aid of power, reduces power generation cost and power grid loss, but also provides powerful support for strategy formulation of power trade participants in a power market environment, and enhances transparency and stability of market trade. The existing power load prediction method mainly collects and selects historical load data, a load prediction model is built based on statistical data, and prediction accuracy is ensured through effective model selection and building. However, these existing methods have some obvious limitations that the lack of consideration of the influence of interaction of various market information data on the power load ignores recent market public opinion corpus information such as policy regulations, notices, major economic activities and the like, and meanwhile, the power data generally presents obvious seasonal, periodic and trending characteristics, so that the existing methods fail to fully utilize space-time correlation in the power data, and the accuracy and stability of prediction are required to be improved. Therefore, in order to improve the accuracy of full-power-saving power load prediction and better meet the operation and management requirements of the power system, a new power load prediction method combining text and associated complex market data is urgently needed. Disclosure of Invention The invention aims to provide an intra-provincial power load prediction method based on multi-mode information fusion analysis, which is used for improving the accuracy, stability and anti-interference capability of power load prediction by fusing multi-dimensional characteristics and adopting a mode of combining deep learning and machine learning. The method specifically comprises the following steps: the method for predicting the power load in the province based on the multi-mode information fusion analysis comprises the following steps: S1, data collection and preprocessing, namely collecting historical load data, market data and market public opinion corpus information, and adopting maximum and minimum normalization processing to the historical load data, wherein a preprocessing formula is as follows: Wherein, the For the i-th data after normalization,As a result of the original data set,As a minimum value for the data set,The market data comprises weather data such as time-by-time temperature, humidity, wind speed, sunlight and the like, city economic data, power transmission line data and holiday arrangement information; When the data is preprocessed, the date information is encoded into discrete values, text news information and industry information are collected, corpus is segmented and word vectorization is carried out, and a data set containing various features is constructed; S2, model training data division, namely dividing a preprocessed data set into a training set and a verification set according to the proportion of 8:2, selecting a Mean Square Error (MSE) function as a loss function of model training, reducing errors by adjusting model parameters, and storing the model parameters when the error of the verification set is the lowest; s3, deep learning model construction and feature extraction, wherein the deep learning model comprises a time sequence information extraction module and a multi-head attention module, and the time sequence information extraction module is used for extracting a Seasonal feature of the seal and a trend trend feature: S3.1. Extracting the characteristics of the serial, namely calculating the local characteristics of the load through convolution kernel dimensions [1, T ] and convolution with the step length of 1, and calculating the global characteristics of the load through convolution kernel dimensions [1, T ] and convolution with the step length of a sliding window of T; Adopting a two-dimensional convolution operation with a convolution kernel dimension of [1, feature_size ] and a sliding window step length of 1, wherein feature_size represents the total number of feature dimensions in input data and is used for dynamically distributing weights for different market information features so as to inhibit the influence of unimporta