Weather variations that affect photovoltaic (PV) generation increase the demand for high temporal resolution power predictions and very short-term forecasts (1 to 30 min) to ensure stability and efficient supply planning. There are several studies based on artificial intelligence (AI), but it is difficult to identify ultra-short- and short-term studies that address the performance evaluation of prediction architectures. Therefore, the strength of this study is its stress test-like design, which compares leading architectures (Transformers, CNN-GRU, Bio-inspired Optimization) under an analysis of accuracy and computational efficiency of PV energy predictions at different ultra-short and short-term time horizons against high-resolution weather variations (every 10 s), providing a solid foundation for future optimization, practical application, and scalability. For a sunny day, the forecasts reached R 2 = 0.9986 and R 2 = 0.9724 on 1-min and 24-h horizons, respectively. For these same time horizons, on a cloudy day, performance decreased by 3.29% and 18.90%, respectively; on a rainy day, the decrease was 7.9% for the 1-min time horizon and 44% for the 12-h time horizon. Likewise, prediction times relative to processing time on sunny days were reduced by 75.38%, 68.95%, 52.83%, and 48.52% for 1-min, 30-min, 3-h, and 12-h horizons, respectively; similar behaviors were observed in the remaining algorithms and day types. • Five algorithms were evaluated for time-based training tests and four for accuracy tests in the prediction of PV power fluctuation with real data with 10-s resolution, under varied weather conditions, for ultra-short and short-term horizons (from 1 min to 24 h). • The best result in evaluation metrics was for the E algorithm, achieving high-precision values of R² for a 1-min time horizon of 0.9986 on a sunny day, 0.9657 on a cloudy day and 0.9194 on a rainy day. • The E algorithm, on a sunny day for the time horizons of 1 min, 30 min, 3 h and 12 h, regarding the processing time, the prediction time was reduced by 75.38%, 68.95%, 52.83% and 48.52%. • Algorithm D, with an accuracy of 99.5% for horizons of less than one hour, shows a reduction in calculation times of 56.18% compared to algorithm E for predictions of photovoltaic power fluctuations.
Fernandez-Fabian et al. (Sun,) studied this question.