Probabilistic forecasting of photovoltaic power with interval prediction and error correction via DTW-PAM clustering and improved TCN quantile regression

Accurate probabilistic forecasting of photovoltaic (PV) power is crucial for grid stability under volatile weather conditions. This paper proposes a three-stage hybrid framework. First, daily power curves are clustered using dynamic time warping partitioning around medoids to capture shape-based similarities, yielding weather-type labels for model specialization. Second, a squeeze-and-excitation temporal convolutional network with quantile loss and interval width penalty directly outputs the 5th, 50th, and 95th percentiles for multi-step prediction, capturing both short-term fluctuations and long-term trends. Third, an irradiance-segmented error correction module, optimized by an enhanced differential evolution algorithm, learns segment-specific correction factors to refine prediction intervals while balancing coverage and sharpness. Experiments on two real-world 15-min resolution datasets demonstrate that the proposed method outperforms state-of-the-art benchmarks across all weather types. On Dataset 1 under sunny conditions, the model fuzzy C-means achieves a mean absolute error of 0.7435, R2 of 0.9900, and an interval score of 5.2778, with a prediction interval coverage probability of 92.85% and normalized average width of 0.0823. Under cloudy and rainy conditions, the method also maintains superior accuracy and interval quality. The framework demonstrates robust generalization across two distinct datasets, offering an effective solution for ultra-short-term PV power probabilistic forecasting.