An efficient QR-BiMGM model for probabilistic PV power forecasting

Abstract

The inherent volatility and randomness of PV output has posed a severe danger to the safe and stable functioning of power systems in recent years, as grid-connected PV capacity has increased. Short-term probabilistic PV power predictions can provide power system decision-makers with more detailed information. A novel probabilistic PV power forecast model based on the fuzzy C-mean clustering algorithm and the QR-BiMGM model is proposed in this research. The MGM (Minimal Gated Memory) model reduces training time while enhancing prediction accuracy by simplifying the neuron structure. The model’s data mining capabilities is further enhanced by the bidirectional propagation structure. To provide probabilistic prediction results, the Quantile Regression (QR) model is employed to integrate the BiMGM model. Under diverse weather conditions, a comparative analysis using actual measurement data with four benchmark models shows that the proposed model has higher prediction performance and good generalization potential. The QR-BiMGM model’s mean integrated interval evaluation metric WC is 31.38 percent lower than QR-MGM, 50.21 percent lower than QR-LSTM, 44.85 percent lower than QR-BiLSTM, and 44.83 percent lower than QR-GRU, according to simulation results. The mean probabilistic prediction evaluation metric CRPS of the QR-BiMGM model is 3.5 percent lower than that of QR-MGM. 3.6 percent lower than QR-MGM, 20.91 percent lower than QR-LSTM, 17.56 percent lower than QR-BiLSTM, and 20.36 percent lower than QR-GRU. In addition, the probability density function based on the conditional quantile of PV power is estimated using a modified kernel density estimation approach.