PV System Performance Analysis and Forecasting Using Deep Learning and Statistical Methods

Abstract

This study investigates the long-term performance degradation and forecasting of three silicon-based photovoltaic technologies—polycrystalline (pc-Si), monocrystalline (mc-Si), and amorphous silicon (a-Si)—using a seven-year dataset (2015–2021) from a semi-arid climate. Degradation rates are quantified through seasonal-trend decomposition and Arrhenius analysis, revealing distinct mechanisms: pc-Si exhibits the lowest annual degradation (0.36%/year), followed by a-Si (0.57%/year), while mc-Si shows the highest (0.77%/year), with a notable thermal annealing effect partially compensating degradation in a-Si. For forecasting performance ratio, four models are compared, where long short-term memory networks achieve the highest accuracy by capturing nonlinear temporal dependencies, while SARIMA offers robust, interpretable results with lower complexity. Beyond predictive performance, the study establishes links between model behavior and underlying physical processes such as degradation and annealing, and analyzes prediction uncertainty in relation to temperature variability and dust accumulation. These findings highlight trade-offs between accuracy, interpretability, and deployment feasibility, providing a framework for PV performance forecasting under univariate, semi-arid conditions, with future work directed toward multivariate, physics-informed approaches across broader technologies and climates.