| Household power load is an important part of the load.Accurate short-term household load forecasting is of great help to issues such as power price formulation,demand side response,or power transmission.Compared with electric load,household electric load has greater volatility,stronger randomness,and more difficult prediction.The purpose of this article is to improve the accuracy of household load forecasting,using signal decomposition and feature selection as data preprocessing methods,and neural network model based on the attention mechanism for forecasting research.This article mainly carried out the following work:(1)The data set used in this article contains multiple variables,and the information contained in each variable has a different degree of influence on the forecasting model.Too many redundant variables will lead to an increase in forecasting time and unsatisfactory forecasting performance,so it is necessary to screen the characteristics of the variables to filter out redundant variables to improve the forecasting accuracy.Therefore,this article uses recursive feature elimination(RFE)to select variables,and the filtering model used by RFE is random forest(RF).(2)For household power load data,it contains a lot of noise,which is reflected in the strong volatility and randomness on the load curve.To extract effective information,this article uses variational mode decomposition(VMD)to decompose load data.The decomposed sub-signals contain periodic information of the load.The sub-signals are added to the data set to form a new data set,which makes the new data set more informative and helpful for the establishment of forecasting models.(3)With the rise of big data,the neural network is better than traditional linear prediction in the forecasting of time series data.In this article,we propose a forecasting model based on long short-term memory(LSTM)combined with self-attention mechanism(SAM)and attention mechanism(AM).When traditional LSTM deals with sequence problems,the weights given to the inputs are the same.When dealing with long sequence problems,the computational complexity of LSTM model greatly increases and the prediction effect decreases.When SAM and AM process long sequence data,they give different weights to each sequence,update the weights through training,and select information processing with significant weights to improve forecasting accuracy.(4)The preprocessed data is input into the forecasting model for testing,mean absolute percentage error(MAPE)and root mean square error(RMSE)are used as evaluation functions to evaluate the prediction effect of the forecasting model.The method proposed in this article can improve the forecasting accuracy through experimental comparison. |
