As photovoltaic (PV) installed capacity continues to rise, accurate forecasting of PV output power is essential for power system operation and control. This paper proposes an ultra-short-term PV power forecasting method based on multi-scale time-frequency features. First, weather types are classified, and typical daily PV power curves are constructed under three representative weather conditions. According to the weather type of the predicted period, the corresponding typical weather curve is selected. Second, Continuous Wavelet Transform (CWT) is applied to the 3 time sequences (including the past 4-hour time sequence, the past 24-hour time sequence, and the typical weather time sequence) to generate time-frequency feature maps. The 3 time sequences provide different perspectives for power output, such as time-series dependency within short time period (past 4-hour curve), power output feature similarity for two consecutive days (past 24-hour curve), and weather correlation for two days with the same weather (typical weather curve). Finally, a parallel Convolutional Neural Network (CNN) architecture is used to extract multi-scale time-frequency features, which are further processed by a Long Short-Term Memory (LSTM) network to capture temporal dependencies. Experiments are conducted using practical roof-top PV power system. The result demonstrates that, compared to traditional models with single-source sequential inputs, the proposed method improves prediction accuracy, validating the effectiveness of time-frequency feature fusion in enhancing PV power forecasting performance.
Zhang et al. (Sun,) studied this question.