Short-term Load Forecasting Based On EEMD-SE-SA-DBN Model

Short-term load forecasting is an important part of power system economic dispatch,and also an important module of energy management system.According to the forecast results,the power grid dispatch department will work out a reasonable start-up plan,operation plan and unit maintenance plan,and load forecast is also one of the important conditions to ensure industrial production and social stability.The original load data will be abnormal data due to system failure and human factors,usually using iForest method for abnormal data testing,according to the results of the detection of abnormal data to correct and complement.The factors that affect load forecasting mainly include two types,one is the internal factors that cause load fluctuation due to the characteristics of load itself,the other is the external factors that cause load fluctuation due to the change of external environment,finally,according to the Pearson correlation coefficient method and the characteristic analysis of the load,the main influencing factors of the load are determined.The load data is characterized by large amount of data and strong non-linearity.The traditional load prediction model cannot effectively extract the deep features of the data.However,when faced with massive data,models based on deep learning theory will show good feature learning ability and generalization ability,so this paper establishes PSO-DELM combination forecasting model.Firstly,the input variable of the model is tested for dimensions,and the input variable dimension is determined to be 10 and the output variable dimension is 1.Because the number of hidden layers and the number of neuron nodes in each layer of the depth limit learning machine can not be accurately determined,the number of fixed neuron nodes to change the number of layers is used to find the number of layers with the least output error.At the same time,the particle swarm optimization is used to optimize the number of neurons in each hidden layer of DELM,and the optimal number of neurons is assigned to each hidden layer of DELM to improve the prediction accuracy of the model.In order to improve the forecasting ability of the model,the load series is decomposed into several IMF series by using the combination of ensemble empirical mode decomposition algorithm and sample entropy theory,the nonlinear,nonstationarity and random fluctuation of load series are reduced by EEMD-SE.At the same time,the SA-DBN load prediction model based on deep learning theory is constructed,the deep belief network is optimized by simulated annealing algorithm,the SA-DBN model was used to predict multiple sequences formed by EEMD-SE decomposition,and the prediction results of each sequence were superimposed to get the final prediction results.In order to verify the superiority of EEMD-SE-SA-DBN model,PSO-DELM and SA-DBN were used for comparative analysis.The simulation results show that the EEMD-SA-SE-DBN model has the highest fitting degree,which is improved by 4.17%compared with the SA-DBN model without EEMD-SE decomposition,and 5.68%higher than that of the PSO-DELM.The average relative error of EEMD-SA-SE-DBN model is 3.89%,while the average relative error of SA-DBN model,PSO-ELM model are 4.24%and 4.56%respectively.Therefore,the EEMD-SA-SE-DBN model has the strongest generalization ability and higher prediction accuracy.Figure[41]table[16]reference[68]...