Research On Photovoltaic/Wind Power Forecasting Based On Deep Learning

With the increasing popularity of photovoltaic/wind in the global power grid,the forecasting of photovoltaic/wind power and the improvement of the accuracy of the forecasting model are crucial for the planning,management and operation of power systems.However,this is a difficult task due to the intermittency and instability of the wind/photovoltaic.Up to now,researchers have developed a variety of methods for photovoltaic/wind power forecasting and are committed to improving the accuracy of photovoltaic/wind power forecasting.Deep learning,as a promising artificial intelligence method,can discover the inherent non-linear features and high-level invariant structures in data,which makes it popular in various fields and has made corresponding achievements.However,deep learn-based forecasting research in the field of photovoltaic/wind power forecasting is not mature enough.In this paper,photovoltaic and wind are taken as research objects,and deep neural networks are used to build the photovoltaic/wind power forecasting model.The problem of photovoltaic/wind power forecasting based on deep learning network is studied,in order to realize the application of deep learning in photovoltaic/wind power forecasting,and improve the accuracy of the forecasting.The main work contents and contributions of this article are as follows:1.Aiming at the problems of poor robustness,complicated forecasting model and low forecasting accuracy of existing photovoltaic power forecasting models,two photovoltaic power forecasting models based on one-dimensional convolutional neural network and long short-term memory neural network were proposed.Compared with the two-dimensional convolutional neural network,the one-dimensional convolutional neural network used in the research does not require data conversion,no information loss,and less calculation on the photovoltaic power forecasting problem;while the long short-term memory neural network can obtain the long-term dependence characteristics of photovoltaic data and improve the accuracy of photovoltaic power forecasting.For photovoltaic power forecasting,when the length of historical data is too short,it cannot reflect changes in factors affecting photovoltaic power such as irradiance,temperature,and wind speed,etc.When the length of historical data is too long,it will cause data redundancy.Therefore,in the study,the influence of the length of historical data on the forecasting accuracy was explored and experimental verification and analysis were carried out.Through the comparison and analysis of the forecasting results of the proposed models,the effectiveness of the deep neural network in photovoltaic power forecasting is verified,and it provides a reference for the selection of photovoltaic power forecasting model and historical data length in practice.2.A hybrid deep neural network model based on convolution and long short-term memory neural network is proposed,which can better extract the temporal and spatial feature of photovoltaic data and improve the accuracy of photovoltaic power forecasting.In the study,starting from the spatio-temporal characteristics of photovoltaic data,the influence of the reasonable combination sequence of the hybrid model on the accuracy of photovoltaic forecasting was also considered,and two hybrid neural network models are proposed and the forecasting results of the two hybrid models and the single model are compared and verified.The experimental results show that the hybrid model of “extracting the temporal features of the data first and then the spatial features of the data”is superior to the hybrid model of “extracting the spatial features of the data first and then extracting the temporal features of the data”.In the research,the corresponding experiments were carried out on the variation of the number of layers of each model of the hybrid model.The experimental results provide a reference for the choice of the number of layers of the hybrid model.3.Aiming at the problems of insufficient feature extraction of large amounts of data and low accuracy of forecasting models faced by current wind power prediction methods,a wind power forecasting model based on deep neural networks is proposed.Due to the seasonality and instability of the wind data itself,and different seasons show different probability distribution characteristics,the wind power has significant instability.If this feature of wind data is not considered,it will lead to inaccurate of wind power forecasting model.In view of this problem,this chapter proposes a clustering algorithm-based input data processing model for wind power forecasting,and finally realizes high-precision wind power forecast.In the study,five mainstream deep neural network models were selected for wind power forecasting and comparative analysis,and the forecasting results were compared with the shallow model.The experimental results show that the proposed model has higher forecasting accuracy in all seasons than the shallow model,and different deep neural networks show different forecasting effects under different data distribution characteristics.4.Compared with single-step forecasting,multi-step forecasting is more important for ensuring the reliability and controllability of wind power systems.However,the current multistep forecasting research on wind is mostly focused on the multi-step wind speed forecasting.For wind farms,the high-precision conversion rules for obtaining wind power indirectly from wind speed are very complicated and difficult to describe,and as the prediction step increases,errors will inevitably become larger and larger.To solve this problem,this paper extends the wind power single-step forecasting model to a multi-step wind power forecasting model,and proposes a long short-term memory neural network model based on encoder-decoder for multistep wind power forecasting.Experimental research shows that the multi-step wind power forecasting model provides better forecasting accuracy and higher robustness than other deep forecasting model.The proposed model can extract more information from long-term wind power historical data,so as to obtain better multi-step wind power forecasting results,and with the increase of the forecasting step,the advantages are more and more obvious,but it also requires a higher amount of calculation.