Research On Photovoltaic Short-term Power Prediction Method Based On Deep Learning

Large-scale interconnection of renewable energy represented by solar energy and wind energy brings unprecedented challenges and difficulties to the planning and operation of power grid.Accurate short-term forecasting of renewable energy output plays a very important role and value in power system dispatching and operation,spot trading decision making,etc.In view of the shortcomings in photovoltaic short-term power point prediction and interval prediction,this paper carried out related work based on deep learning theory.To solve this problem,based on extensive review of relevant literature,this paper analyzed the influence of irradiance error characteristics on reducing NWP data error,proposing photovoltaic short-term power point prediction LSTM model and interval prediction model based on similar daily irradiance error correction and scenarios respectively.Firstly,the maximum mutual information coefficient method was applied to analyze the influencing factors,and the XGBoost feature screening method was conducted to realize the feature screening of the influencing factors.Secondly,in view of the strong time-series characteristics of short-term photovoltaic power,a short-term photovoltaic power point prediction model based on LSTM was constructed.On this basis,since the NWP data is not capable to forecast rapid formation and extinction of small-scale clouds accurately,causing inaccuracy,this paper introduced a concept of irradiance error innovation,used to describe the fluctuation phenomenon of ground surface irradiance under the influence of rapid cloud formation and extinction,and built the model and method of forecasting daily irradiance error estimation based on K-Means++clustering method.Then,a short-term photovoltaic power point prediction LSTM model based on similar daily irradiance error correction model was devised.Thirdly,aiming at the error distribution characteristics of photovoltaic short-term power point prediction,the student t distribution was chosen to replace the normal distribution in the traditional method as the best probability density distribution of the prediction error of fitting points,and the confidence interval was solved,proposing the scenario-based photovoltaic short-term power interval prediction model.Finally,the model and method presented in this paper were verified by combining the actual data of one certain place,and the validity and reliability were demonstrated.