Research On Short-term Power Load Forecasting Based On VMD And Improved LSTM

The power industry is one of the important basic industries of the country,and the accuracy of power load forecasting is also directly related to the supply-demand balance and operating costs of the power grid.The traditional single prediction method has a poor fitting effect when faced with non-stationary and nonlinear fluctuation load sequences.In this paper,a combination of deep learning and signal decomposition methods,combined with actual data,is used to combine short-term power load prediction the study.The main work of this article includes:(1)This article introduces the concept,characteristics and basic process of load forecasting in detail,focusing on the analysis of the time,climate and other influencing factors of power load forecasting.(2)The advantages of Long Short-Term Memory(LSTM)for complex data fitting and time correlation analysis of complex data.This article compares and analyzes the fitting effects and prediction accuracy of different neural network models in detail.Due to the problem that the parameters of the LSTM neural network cannot be determined,the Adaptive Particle Swarm Optimization(APSO)with adaptively adjusted inertial weights and learning factors is introduced to count and iterate the hidden layer neuron nodes in the LSTM neural network model.The number of times and the learning rate are used to optimize the network parameters,and the three test functions are used to verify the superiority of the APSO algorithm in optimization accuracy and speed.(3)In this paper,the load sequence is regarded as a random fluctuation signal,and the signal processing method is introduced into the Variational Mode Decomposition(VMD)algorithm to decompose the original sequence into sub-sequences with limited bandwidth.Compared with the traditional Empirical Mode Decomposition(EMD)method,the advantages of VMD in mode aliasing and end effect are analyzed and discussed in detail.At the same time,K-means clustering is used to calculate the Euclidean distance of each sub-modal component,a calculation method of sub-sequence prediction results is weighted and reconstructed to obtain the final result,and a short-term power load prediction model based on weighted VMD-APSO-LSTM is constructed.(4)To verify the validity and practicability of the model in this paper.Combined with the actual data,the simulation experiment was carried out in Matlab,and the appropriate values of important parameters of VMD were determined through several experiments.The original data is decomposed into sub-modal sequences of limited bandwidth by VMD.The weighted VMD-LSTM,weighted VMD-PSO-LSTM,weighted VMD-APSO-LSTM and weighted VMD-BP are used to compare the prediction results.The experiment shows: In terms of prediction results of subsequences and total curves,the prediction accuracy of the weighted VMD-APSO-LSTM model is generally higher,especially the curve fitting degree is better at extreme points or turning points.In order to further verify the superiority of the model,the model prediction results are compared for whether weighted and whether to use VMD decomposition.The models in this paper have achieved good prediction results.