Interpretability Verification of Real-time Deviation Correction Probability Prediction Model of Photovoltaic Output
In order to solve the problem that the photovoltaic output(PV)probabilistic prediction model may generate model drift over time and lack of interpretability,an interpretability verified real-time corrective probabilistic prediction model for PV output is proposed.Firstly,a PV output probabilistic prediction model based on natural gradient boosting(NGBoost)is constructed under the premise of considering seasons,and the PV output probabilistic prediction models of different seasons are obtained through hyper-parameter tuning.Secondly,by analyzing the temporal specificity of PV out-put objects,a simple and effective real-time bias correction strategy of PV output probabilistic prediction is proposed,so as to solve the current problem of model drift over time and lack of interpretability.Moreover,we combine the Shapley additive explanations(SHAP)method to explore the influences of dominant factors on the uncertainty of PV output,aim-ing to explore the contribution of each input feature to the uncertainty of PV output in different seasons through global and local explanations,and further validate the reasonableness of the prediction process and the correction strategy of the constructed model.Finally,the proposed model is validated by simulations using public data sets.The results show that the continuous probability ranking scores of the proposed model are in the range of 9.89～24.01 kW,which are more accurate compared with other models and can analyze the complex prediction process,providing effective theoretical supports for the prediction of PV output uncertainty.
万方数据
维普
