Short-term Power Load Forecasting Based On Difference Fusion And Deep Learning

Electric energy is one of the most important energy sources in the technological progress and economic development of modern society,and as the basic energy of national life,it has a profound impact on People’s Daily work and life.With the proposal of the national "double carbon" goal,the rise of wind power generation,photovoltaic power generation and other new energy power generation methods,and the rapid growth of the number of intelligent equipment and new energy vehicles in recent years,the electricity situation is increasingly complex,and the difficulty of power dispatching and power load prediction is gradually increasing.Traditional power load prediction methods are often difficult to meet the requirements of power companies on the accuracy of load prediction.In order to further improve the prediction accuracy of short-term power load,reduce resource waste and excess carbon emissions caused by loss of load and excessive power generation as much as possible,improve energy utilization efficiency,economic and social benefits,and reduce the operation cost of power system,this paper proposes a shortterm power load prediction method based on differential fusion(DF)and deep learning.The main research contents are as follows:(1)Analyze the performance of traditional load forecasting method and a single deep learning method.Because the traditional load forecasting method has a simple model structure,it is difficult to deal with the electrical load with increasing periodicity,nonlinear and timely sequence and numerous influencing factors.In this paper,multiple load data sets are used to evaluate the differences between traditional load forecasting method and a single deep learning method.The two load forecasting methods are summarized and discussed,and the single model Long and short term memory(LSTM)network with the best performance is selected as the first step scheme and the benchmark model for further research.(2)In view of the fact that load sequences tend to have high non-stationary,strong randomness and time continuity,and have a certain trend of increasing or decreasing with time,which has a certain influence on the accuracy of load prediction,a short-term power load prediction method based on difference fusion,improved residual network(IResNet)and two-layer LSTM is proposed as the second step scheme.The ADF stationarity test was carried out on the load sequence.For the very unstable sequence,the difference method was used to remove the trend of the load sequence and reduce the nonstationarity of the load sequence.Meanwhile,Pearson correlation coefficient is used for feature selection,influential factors with high correlation are integrated,and IResNet,which has strong feature extraction ability,is used to extract spatial coupling features of multidimensional feature data.It improves the ability of two-layer LSTM time sequence feature extraction,avoids the gradient explosion and network degradation when deepening the network,and further improves the accuracy of load prediction compared with the existing deep learning methods.(3)Considering that the accuracy of load prediction is limited when IResNet and dual-layer LSTM network structure parameters are artificially set,an optimization algorithm is added on the basis of the second step scheme,and the average absolute error of test set is taken as fitness function to find the optimal network structure.After comparing the optimization model structure of particle swarm optimization algorithm(PSO),genetic algorithm(GA)and whale optimization algorithm(WOA),the improvement degree of load prediction accuracy was selected as the final model structure optimization algorithm of this paper.In addition,many deep learning models are still difficult to predict the peak and trough load well,usually with large errors.However,there is effective information in the error sequence.On the basis of improving the prediction accuracy of this model,LSTM can be used to mine the error sequence information and perform error correction(EC)on the prediction results,so as to improve the overall accuracy and robustness of the model.Reduce load forecasting error.The accuracy and effectiveness of the scheme are verified by comparative experiments.