A Grey Wolf Optimized LSTM for Short-Term Solar Photovoltaic Power Forecasting

Abstract

Accurate short-term forecasting of photovoltaic (PV) power is essential for the stable and economic integration of solar energy into modern power grids. PV output is highly nonlinear and strongly influenced by rapidly changing cloud cover, which makes short-horizon prediction difficult and often leaves simple persistence methods hard to beat. This paper presents a hybrid forecasting model in which a long short-term memory (LSTM) network is optimized by the Grey Wolf Optimizer, a swarm intelligence algorithm that automatically tunes the network hyper-parameters separately for each forecast horizon. The model is trained on field measurements from a grid-connected PV plant sampled every fifteen minutes, using irradiance, module and ambient temperature, an engineered temperature gradient, cyclic time features, and an empirical clear-sky index that captures cloudiness. A contiguity-aware windowing procedure prevents training sequences from crossing overnight gaps, removing a subtle source of leakage. Evaluated at horizons of fifteen minutes, one hour, and three hours, the optimized model consistently outperforms persistence, with the forecast skill rising from about twelve percent at fifteen minutes to fiftyeight percent at three hours. The results confirm that swarmoptimized deep networks deliver their greatest advantage at the longer horizons most relevant to grid dispatch and reserve scheduling.