Synthetic feature selection and time-division stacking for photovoltaic output forecast without surface solar radiation information
Solar radiation observatories are few and lacks surface solar radiance(SSR)information nationwide,thus making it difficult to accurately predict the photovoltaic power output.Given this,this paper proposes a method of prediction without SSR.First,the feature augmentation on the original data is performed,and the idea of time-division of the data is proposed to further enhance the relevance of important features.Then,the D-S evidence theory is proposed to synthesize the score of multiple feature scoring methods,and the n-ratio method is employed to determine the threshold value to realize the dispatching of features.Finally,the cross-validation method is proposed as well as the Box-Cox normal transformation of the input layer to realize the improvement of the Stacking model,and the integrated prediction of the segmented sample set.Our results show the accuracy(CR)and qualification rate(QR)on the selected prediction days are 0.948 and 1.000 respectively,which are 16.5％and 20.3％higher than those of the unprocessed method,delivering fairly good prediction performances and satisfying the requirements of PV output prediction.
solar radiationPV power forecastfeature augmentationD-S evidence theoryBox-Cox normal transformationtime-division prediction
国家科技期刊平台
NSTL
万方数据
