Research On Short-term Power Load Forecasting Based On Reinforcement Learning And LSTM

With the advancement of technology in the energy sector,electricity can be generated,transported and stored more conveniently,efficiently and economically.Smart grids are designed to create automated,efficient energy transmission networks to improve the reliability and quality of power transmission,as well as network security,energy efficiency,and demand-side management aspects.Forecasting power demand is a basic task in power system management.Overestimating or underestimating demand may lead to instability of the power grid and insufficient utilization of resources,leading to high economic costs.Power system infrastructure planning requires long-term load forecasts(months to a year).However,operational decisions for smart grids must be made within a short period of time and require short-term load forecasting(STLF)(hours to days).Accurate short-term load forecasting is critical to the efficient operation of the power sector.Due to the high volatility and the uncertainty of the load,it is difficult to predict the load with fine granularity such as a single home or building.The power load in this paper is based on the overall load forecast in some areas.This paper focuses on the following two issues: 1.Constructe short-term power load forecasting algorithms that combine reinforcement learning ideas with long short-term memory networks;2.Verify the effect of deep learning network structure on the accuracy of short-term power load forecasting.The main contents of this article are as follows:First,construct short-term power load feature sets,perform data preprocessing operations such as missing value completion and outlier removal on historical power load data,and select influencing factors related to power load,such as temperature,date and time information,etc.In the circumstances,select an appropriate pre-processing method,serialize the processed features,learn from the word embedding method,and concatenate into a brand new time series to construct model input data.Secondly,construct a combined forecasting model,combine 1D convolutional neural network 1D-CNN with long and short-term memory networks,and then further improve the network structure to build a short-term power load forecasting model based on reinforcement learning and LSTM.The model is mainly composed of two parts,LSTM and 1D Convolutional Neural Networks(1D-CNN),and the Inception network structure is introduced,combining the performance advantages of LSTM and CNN,combining the order sensitivity of LSTM with the speed and lightness of convolutional neural networks,the prediction frequency is 1 hour,that is,one load forecast value per hour,and the load of the next day is predicted.Finally,the feasibility of a short-term power load forecasting model based on reinforcement learning and long-term and short-term memory network is studied.Experiment with real data set based on deep learning framework Pytorch,and multiple evaluation indicator is set up.The performance comparison shows that the short-term power load prediction model based on reinforcement learning and LSTM has greatly improved the prediction accuracy and accuracy.