Research On Transfer Learning-based Wind Power Prediction

With the rapid development of national sustainable energy industry,the main problem of the scaled use of sustainable energy is to realize the large-scale consumption by the power grid.Wind energy is more difficult to use than other energy sources because of uncertainty,randomness,and intermittency.In addition,due to the continuous increase in wind power capacity in China,a large number of wind farms lack sufficient historical data to model wind power forecasts,resulting in low overall power generation and grid efficiency.In response to this problem,this paper proposes a series of methods to improve the forecasting effect on wind power and resolve problems in data insufficiency from the perspective of physical analysis,feature extraction,transfer learning and model integration.Therefore,we designed a wind power prediction method based on transfer learning and achieved good results.Firstly,we acquire wind power data and numerical weather predict data from multiple sources.Spatial correlation analysis is used to estimate the real-time wind speed for each wind turbine.By fitting the wind power characteristic curve,wind power forecasting problem is converted into wind speed forecasting.By analyzing the uncertainty of wind speed,we found amplitude modulation effect and multi-scale effect between variance of wind speed and average wind speed.We used wavelet technics to analyze the variation of wind speed at different time resolutions and frequencies,and performed predictability analysis on the uncertain components of wind speed,then used autocorrelation function to analyze the predictability of wind speed over different forecast lengths.Upon these works,several feature subsets that are easy to perform regression analysis are extracted from the original wind speed data.Secondly,a wind power prediction model is constructed by using transfer learning strategy.We adopted a second-order ANFIS model as basic learning unit,and used TransferAdaboost algorithm as the optimization strategy.The ANFIS model is optimized by reducing the weight of the data with large regression error in source domain training data.The experimental results show that wind power forecasting using transfer learning has a significant improvement over the traditional models.In addition,as increases in the length of forecasting length,model performance gradually deteriorates,but the transferbased model is still better than non-transferred learning models.When models are trained in training data from 1% to 90%,the TransferAdaboost model is better than the nontransferred ANFIS and SVM models in most training data groups.Finally,a weighted ensemble model based on transfer learning is proposed.We decomposed 72-hour forecast length into a number of different size intervals.Each base model is trained for each specific interval length to guarantee best results in this interval.Then genetic algorithm is used to integrate all based models by weighting average.Experimental results show that the integrated model weighted by the genetic algorithm is better than the unweighted integrated model and non-integrated model,especially in short-term forecast.