Аңдатпа
The growing integration of photovoltaic systems into modern power grids requires accurate solar power generation forecasting to support reliable supply-demand balance and efficient energy management. In recent years, the rapid expansion of machine learning applications in this field has led to substantial methodological diversity, which complicates direct comparison across studies. This paper presents a systematic review of peerreviewed research on machine-learning-based solar power forecasting. The analysis focuses on commonly used input features, modeling approaches, evaluation metrics, and reporting practices across different geographic regions. The review shows that deep learning methods, particularly long short-term memory (LSTM) networks, are increasingly adopted, while traditional approaches such as linear regression and random forest remain widely utilized. Meteorological variables - including temperature, solar irradiance, humidity, and wind speed - consistently form the core input features. However, considerable inconsistencies in reporting practices are observed. Geographic information is often insufficiently detailed, and standardized or normalized evaluation metrics that would enable cross-study comparison are rarely applied. Several important research gaps are identified, including limited model generalization across locations, insufficient incorporation of uncertainty quantification, and inadequate validation under real operational conditions. In addition, Central Asian countries, including Kazakhstan, are entirely absent from the reviewed literature despite their significant solar energy potential. Overall, the findings highlight the need for more standardized evaluation protocols, expanded geographic representation, and improved reporting transparency in future research.