The growing role of solar power has underscored the need for accurate solar energy generation forecasting to address its variability. Machine learning models, particularly tree-based algorithms like CatBoost, have demonstrated significant potential in enhancing forecast accuracy. This study focuses on developing CatBoost models to predict future solar power production. Data from one of The City of Calgary’s solar photovoltaic projects was utilized. Various temporal and statistical features were extracted to improve a day-ahead forecast precision, contributing to more reliable grid management and energy system integration. A comparison of three different feature set configurations was conducted. The results revealed that the CatBoost model incorporating temporal and statistical features was the best performer. SHAP analysis was applied to identify the most influential features for the top-performing model. The analysis found that the past lag was the most influential feature, with the top five features being a combination of temporal and statistical attributes. This indicates that extracting and utilizing such features is highly beneficial.