Research On Electricity Load Forecasting Based On Improved Echo State Network

The stable supply of electricity is the basis of people’s lives,factory operations and agricultural production.Accurate electric load forecasting helps power companies adjust the amount of power generation and maintain a balance between supply and demand.It is an important part of automation management,so it is far-reaching to study high-precision and high-performance electric load forecasting models.In recent years,artificial neural networks have become an important research direction because of their excellent prediction performance.Among them,Echo State Network improves the middle layer of traditional recurrent neural networks and solves the problem that recurrent neural networks are easily overfit problems and reduce the complexity of the algorithm.Because of its excellent performance,the ESN model quickly became one of the most popular research directions in time series prediction.However,faced with the characteristics of non-linear,non-stationary and easy mutation of electric load data,classic ESN can no longer meet the requirements of the accuracy.This paper starts from ESN expansion and reservoir improvement,and carries out research on electric load forecasting based on improved Echo State Network.The main work includes the following three aspects:(1)Aiming at the problem of electric load multi-step prediction,a voltage multi-step prediction method based on phase space reconstruction is proposed.This method fits the nonlinear voltage data by reconstructing the phase space.On this basis,the results of direct multi-step prediction and iterative multi-step prediction are fused by weighted.Experiments show that,compared with direct multi-step prediction and iterative multi-step prediction,this method obtains higher prediction accuracy,and the improvement range is between 10% and50%.(2)Aiming at the multitimescale structures of electric load,a multi-reservoir echo state network model is proposed to obtain multitimescale structures of electric data.The model consists of a multitimescale encoder and a convolutional decoder.The encoder contains multiple reservoirs with different time spans,which maps low-dimensional data to high-dimensional space,and the multi-reservoir structure encodes the history of time series as the echo state vector.The convolution decoder is composed of a convolution layer and a fully connected layer.The convolution layer is used to extract features in the echo state vector,and the fully connected layer combines all the feature values to obtain the model output.(3)The electric load forecasting experiment in Yangzhou shows that the multi-reservoir echo state network can effectively extract the multitimescale features,and its performance is better than existing algorithms.It can be seen from the visual analysis that the model has short-term memory capacity and the ability to process complex nonlinear electric load data.