Research On Combined Forecasting Of The Short-term Wind Speed Based On LSTM

Along with the considerable progress of industrialization in our country,there have been lots of major production safety accidents caused by toxic gas leakage in many places,which have seriously endangered the lives and property of the people.After the leakage of poisonous gas,the reasons for its propagation path and diffusion trajectory are complicated.Among them,the decisive role is wind.Wind in natural wind field has turbulent and intermittent characteristics,which increase the difficulty of locating toxic gas leaks.It will provide a valuable clue for locating the toxic gas leakage to forecast the changes of complex wind fields accurately.Based on the shortcomings of traditional wind speed forecasting models,this paper establishes wind speed forecasting model with LSTM deep learning model.Moreover,this paper puts forward an adaptive iterative reinforcement high-precision combined forecasting method for short-term wind speed.The specific works of this paper is as follows:Firstly,SVR wind speed forecasting model,ARIMA wind speed forecasting model and BP neural network wind speed forecasting model are established by using three classical prediction methods: Support Vector Machine(SVR),Time Series method and Artificial Neural Network(ANN)respectively.Furthermore,we forecast short-term wind speed by these classic wind speed forecasting models.The results show that SVR forecasting model has strong generalization,but it will encounter difficulties when the training data is large and the forecasting effect is not very satisfactory.ARIMA forecasting model can transform increasingly volatile time series into mathematical model.However,the forecast result of ARIMA model may not ideal in dealing with strongly nonlinear data.BP neural network forecasting model has strong adaptability and fault tolerance.So BP neural network forecasting model is more suitable for analysing nonlinear time series that are highly intermittent and fluctuant.Among three classical forecasting models,BP neural network forecasting model achieves the best forecasting result.But it will encounter many problems such as being trapped in local minimum values and vanishing gradient problem.Secondly,aiming at the problems that feedforward neural network can’t makes full use of temporal information and traditional recurrent neural network may faces vanishing gradient problems and exploding gradient problems,LSTM neural network is adopted to establish a wind speed forecasting model.The results show that the forecast effect of LSTM forecasting model is better than that of three classical forecasting models.So it is able to reflect the continuous change between the wind speed data and mine internal information of wind speed time series.And we can understand its fluctuation structure and internal laws by LSTM forecasting model.Thirdly,aiming at the problem that single wind speed forecasting method can’t achieve the best results on all application object,we propose an adaptive iterative reinforcement high-precision combined forecasting method for short-term wind speed.This method firstly optimize several single forecasting methods,secondly adjust combined coefficient of each forecasting method adaptively and take it as a weak regression predictor,then constantly adjust the weights of input samples and we can get multiple weak regression forecaster,finally combine the results of all weak regression forecasters into high-precision short-term wind speed forecast value.The experimental results show that the forecast effect of adaptive iterative reinforcement high-precision short-term wind speed combined forecasting methods is superior to that of the entropy weight method combined forecasting method and generalized regression neural network combined forecasting method.