Day-Ahead Solar Power Forecasting with Pattern Analysis and State Transition

This study focuses on the day-ahead forecasting of the power generation of the grid-connected photovoltaic (PV) system. In the literature, most PV power forecasting models depends on the information from the numerical weather prediction (NWP). The errors of the weather prediction might cause the performance of the PV power forecasting model to decrease. Therefore, in this paper, the proposed day-ahead PV power forecasting model only utilizes the historical PV power generations as the input data. The proposed forecasting model is composed of two parts, pattern analysis and the state transition. The pattern analysis evaluates the similarity of the time series of the daily PV power generations based on Euclidean distance and group the time series with similar patterns by k-means time series clustering algorithm. Each cluster is regarded as one state. Then, the probabilities of the state transitions between two consecutive days are calculated. Finally, the expected PV power pattern of tomorrow is obtained as the forecasted PV power outputs. In this study, the proposed model is compared with the traditional feed-forward neural network (FFNN). The results of the cast study demonstrate the proposed model leads to smaller errors in 10 sets of the testing data.