Research on time-series based and similarity search based methods for PV power prediction

The improved accuracy of photovoltaic (PV) power prediction facilitates the stable operation of power systems. The PV power prediction methods based on time-series and similarity search are proposed in this paper. Firstly, the PV power data is categorized based on the volatility of time-series. The decomposed PV power data serves as the data basis for the research of prediction methods. Then, an improved bidirectional long short-term memory (BiLSTM) method for time-series prediction is proposed. The hyperparameters of BiLSTM are optimized using improved crayfish optimization algorithm (ICOA). Finally, the PV power prediction method based on similarity search is proposed by researching the application scenarios for FastDTW (Fast Dynamic Time Warping) and HNSW (Hierarchical Navigable Small World). The effectiveness and superiority of above two methods are verified through experiments. The optimal RMSE (Root Mean Square Error) and R2 (Coefficient of determination) of BiLSTM-ICOA are 0.02081 kW and 99.898%, respectively. The optimal RMSE and R2 of FastDTWHNSW are 0.01645 kW and 99.923%, respectively. The applicable scenarios of proposed methods are further presented through experiments. BiLSTM-ICOA is suitable for ultra-short-term and short-term power prediction with an optimal RMSE of 0.0243 kW. FastDTWHNSW is suitable for medium-term and long-term power prediction with an optimal RMSE of 0.0165 kW. •An effective prediction method based on time-series is proposed.•A superior prediction method based on similarity search is provided.•The suitable application scenarios for the two proposed methods are demonstrated.