Neural Network Short-Term Power Load Forecasting Based On Intelligent Optimization

Power load forecasting accurately is a prerequisite for developing the field of power supply and development.and its reliability is enough to eliminate the dilemma caused by its inherent irregularity and randomness and non-stationarity,to achieve effective power supply and dispatch balance,energy conservation and emission reduction standards and economic benefits.So this thesis proposes a short-term power load forecasting model based on CDM-ADPSO-LSTM-COR,which has spectral clustering feature fusion,error compensation strategy and deep learning combination.The main contents of this article include:1)The classification of power load forecasting is discussed,to determine the thesis load forecasting cycle category and summarize the characteristics of load series.This thesis analyzes power load case from two aspects of time factor and date factor,to form a preliminary judgment on different types of factors affecting power load.2)For the complex fluctuation of power load,the CEEMDAN algorithm is introduced to decompose to weaken the non-stationarity of power load.Comparing the EMD and EEMD algorithms,CEEMDAN algorithms has the adavantages of solving pattern aliasing and difficult alignment of different orders subsequences.Combined with the DTW＿K-Mediods algorithm determines the number of clusters to realize the reconstruction of the decomposed components of CEEMDAN algorithm,thereby reducing the multiple random errors introduced by the traditional algorithm in modeling the decomposed components.MIC（Maximum Information Coefficient）algorithm is used to measure the correlation between various influencing factors and each reconstructed sequence,to determine the best features of each reconstructed sequence.Therefore,a CDM spectral clustering feature fusion algorithm model consisting of CEEMDAN,DTW＿K-Mediods and MIC is established,to decompose and reconstruct and feature the payload data.3)Aiming at the defect of RNN gradient vanishing,the working principle of LSTM and its structural variant bidirectional LSTM network is discussed.Due to the strong correlation between power load and time,bidirectional LSTM can more effectively capture the intrinsic correlation between historical and future load sequences.For the random initialization of LSTM network parameters,two improvements are made: one is to improve Adam’s second-order moment estimation,and adabelief algorithm is introduced to optimize the iteration speed of LSTM network;Second,the adspo（Adaptive Particle Swarm Optimization algorithm,ADPSO）algorithm is proposed without considering the inertia weight in PSO algorithm and introducing a dynamic function to the learning parameters.The benchmark function test results show that the ADPSO algorithm effectively improves the ability of particles to jump out of the local minimum trap and the convergence speed.Therefore,ADPSO-LSTM model is established for preliminary prediction.In order to estimate the uncertainty associated with the preliminary results,an error compensation strategy for LSTM network is proposed.4)A model framework based on CDM-ADPSO-LSTM-COR is bulit.On the public data set,experimental simulation analysis is carried out from decomposition and clustering,different improved algorithms of LSTM,fusion influencing factors and error compensation.Finally,the proposed model is compared with the other four models,the experimental simulation results show that the predicted value of the proposed model CDM-ADPSO-LSTM-COR is highly fitting to the acrual load.