Deep reinforcement learning based interpretable photovoltaic power prediction framework

Accurate prediction of photovoltaic (PV) power is of great significance to the stability and economic operation of renewable power systems. However, very few studies have been reported on explainable deep reinforcement learning and outliers in PV power prediction. To this end, a novel deep reinforcement learning-based interpretable framework considering outliers is proposed to predict PV power. First, a categorical boosting (CatBoost) model is utilized to train and predict the occurrence of PV power outliers. Then, a novel twin adaptive convolutional deterministic policy gradient (TACDPG) model is proposed with CatBoost classification prediction results taken into account to predict PV power values. The proposed TACDPG can improve prediction accuracy by introducing the twin critic network, convolutional neural network, and adaptive reward function. In addition, to improve PV power prediction performance, a feature selection method based on random forest is proposed for CatBoost and TACDPG predictive models to find the main relevant influencing features. Finally, the relationship between relevant features and output variables for the TACDPG is analyzed using an interpretable method based on the SHapely additive explanation. The proposed prediction framework is verified on the actual operation data of the PV plant in Northwest China. It is calculated that the minimum and maximum mean absolute error values of the proposed prediction framework are 0.0025 and 4.2848, with an average value of 0.8235. Compared with the TACDPG model without outlier prediction, the mean absolute error of the proposed prediction framework is reduced by 0.186. The experimental results also demonstrate that the proposed deep reinforcement learning framework can explain the importance of relevant features and that the TACDPG model is superior to the benchmarks.