Research On Short-term Electric Load Forecasting Model Based On RNN

Recurrent Neural Network(RNN)is one of the most promising networks of deep learning.Many of the improved structures derived from RNN have been widely used in speech recognition,machine translation and DNA sequence analysis.In the smart grid field,the collection of user behavior data and prediction analysis by RNN or RNN's improved Long Short-Term Memory(LSTM)has become the mainstream of load forecasting.In the highly competitive power market,the prediction of short-term power load is of vital importance to both consumers and producers.The predictions with high accuracy can not only help power companies rationally plan and distribute resources,but also enable them to take appropriate actions in a timely manner.At the same time,such predictions could balance the supply and demand relationship of electricity,and provide consumers with cost-effective power services as well.Based on the basic theory of power load forecasting,this paper proposes an improved RNN network(long-short-term memory network Long Short-Term Memory,LSTM)combined with attention mechanism as a new power load forecasting model.The traditional LSTM network is compared with the new model proposed in this paper in terms of accuracy and speed.In this paper,the actual load data has been simulated in the above two models.It is verified that the new network structure proposed in this paper has improved in accuracy compared with the LSTM network,and the training duration is also shortened.This paper mainly completed the following work:(1)According to the characteristics of historical load data and the correlation between input vectors,two different LSTM power load prediction models based on RNN network are built.Aiming at the problem that the gradient disappears due to the convergence of the activation function after iteration,a solution combining the traditional LSTM network and the Attention mechanism is proposed to make up for the long-term caused by the gradient or even disappearing of the traditional LSTM network.Different from the traditional LSTM network adopting the structure of forgetting gate to ease the disappearance of gradient,the proposed model mainly relies on the dynamic weight distribution of input vectors to complete the real-time control of the signal to weaken the influence of gradient disappearance.(2)Data preprocessing and normalization are performed for all input vectors,and the neural network structure is designed according to the characteristics of the sample data.Since the acquired historical data has vacancies and noise,the input vector is added and denoised,which reduces the possibility that the sample data will produce prediction bias due to non-compliance with the model.By equalizing the input vector to a certain range,the adverse effects caused by the unconventional sample data are reduced,and the over parameters of the neural network are determined by the control variable method.(3)The processed sample data has been input into the traditional LSTM network and the LSTM network combined with the Attention mechanism to simulate the two models,and the Relative Error(RE)and Absolute Error(AE)are selected as the evaluation indicators.By comparing the prediction results of the two models,the conclusion that the LSTM error of the Attention mechanism is relatively smaller is obtained,which proves that the LSTM with Attention mechanism can improve the prediction accuracy and shorten the training time in the power load forecasting.