A Photovoltaic Power Prediction Method That Integrates Deep Feature Extraction and Uncertainty Modeling

Abstract

Driven by China' s carbon peaking and carbon neutrality goals, the high penetration of photovoltaic (PV) power places increasing demands on both the accuracy and reliability of short-term forecasting. However, strong nonlinearity, pronounced day - night zero inflation, and weather-driven uncertainty significantly limit the effectiveness of conventional point prediction models. This paper proposes a probabilistic PV forecasting framework that integrates multi-scale deep feature extraction with hybrid uncertainty modeling. A CNN - LSTM - Attention architecture is employed to capture local irradiance variations, temporal dependencies, and rapid weather transitions. On this basis, a hybrid output layer combining Gaussian negative log-likelihood and Zero-Masked Quantile Regression (GNLL+ZM-QR) is designed to jointly learn the conditional mean, variance, and quantiles while excluding nighttime zero samples from quantile optimization, thereby mitigating zero-inflation-induced bias and yielding physically consistent prediction intervals. Experiments on 13 months of real-world PV data demonstrate that the proposed method consistently outperforms LSTM, random forest, and CNN-LSTM baselines in both point accuracy and probabilistic reliability. The resulting prediction intervals remain stable under sunny, cloudy, and extreme weather conditions, supporting grid scheduling, energy storage management, and high-penetration renewable integration.