An Ensemble Wavelet Deep Learning Approach For Short-Term Load Forecasting

With the continuous development of power system and energy technology,electric vehicles,distributed renewable energy access,power market gradually mature and demand side response are seriously changing the load variation law.The load variation law presents higher uncertainty and volatility so that the difficulty of load forecasting increases and the traditional power load forecasting methods are facing great challenges.This paper studies the traditional load forecasting methods and summarizes the different classification of load forecasting methods as well as their advantages and disadvantages.Then this paper analyzes the challenges faced by the traditional methods in power system nowadays.In recent years,artificial intelligence plays an increasingly important role in the field of load forecasting.In this paper,an ensemble forecasting approach based on deep learning and wavelet transform is proposed.First,three sub prediction models are established,based on deep learning method including deep belief network,long-term memory neural network and multi-layer perceptron.Then,according to the different characteristics of the three deep learning methods,this paper builds an ensemble learner based on deep learning approaches by the simple average hybrid method,which effectively improves the accuracy and generalization ability of the forecasting model.Finally,according to the characteristics of the load in the current power system,in other words,multi period variation law,this paper uses wavelet transform to decompose the load sequence,and hybrid it with the ensemble deep learning forecasting approach mentioned above to establish an ensemble wavelet deep learning approach for short term load forecasting.In this paper,the real load data of a low-voltage substation in a city in east China is used to simulate and the structural parameters of the deep learning forecasting model are optimized by the approach of hidden layers joint setting using the test data.The case has proved that the proposed method has higher prediction accuracy than deep learning approaches,BP neural network and support vector machine in the ultra-short-term load prediction of multiple time scales.Finally,the simulation analysis of load prediction in different seasons proves that the proposed method has fine generalization performance.The work in this paper has much engineering application value.In the end,the paper summarizes the research work and gives the prospect of future work.