Quantile-Based Short-Term Probabilistic Forecasting of Solar PV Power Under Data Loss

Accurate forecasting of solar photovoltaic (PV) power is essential for grid stability and efficient energy management. This study investigates the impact of missing data, introduced through both random and block removal at varying percentages, on the performance of hybrid neural network architectures, Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM), LSTM Autoencoder, and standalone LSTM models, for short-term PV power prediction. Using two real-world PV datasets representing semi-arid and desert climates, the focus is placed on non-parametric probabilistic forecasting, leveraging Kernel Density Estimation and quantile regression to generate confidence intervals that capture forecast uncertainty. Experimental results demonstrate that while all models experience performance degradation with increasing levels of missing data, hybrid architectures such as CNN-LSTM and LSTM Autoencoder exhibit greater robustness in capturing temporal patterns and maintaining forecast reliability. Notably, the CNN-LSTM model achieved a 19.35% reduction in quantile loss compared to the standalone LSTM under 30% random data removal. The findings underscore the importance of evaluating both predictive accuracy and uncertainty calibration under data scarcity scenarios to enable dependable solar energy forecasting and resilient energy management.