AI-driven day-ahead renewable energy forecasting: Assessing machine learning performance under blind and weather-informed paradigms

Accurate day-ahead renewable energy forecasting remains one of the most challenging problems in modern energy systems because photovoltaic and wind generation are not controlled by the operator, but by uncertain atmospheric forcing. This paper develops a complete NASA POWER-driven artificial intelligence framework for one-day-ahead renewable power forecasting under two sharply different operational assumptions. The first scenario is a strict blind forecast , where only historical information is available. The second is a weather-informed forecast , where day-ahead meteorological predictors at the target hour are introduced as proxy weather-forecast inputs. Hourly solar irradiance, wind speed, temperature, and humidity from NASA POWER are used to construct photovoltaic production, wind turbine generation, battery state of charge, persistence baselines, lagged features, residual-correction forecasts, daily-energy forecasts, and learning-based predictions. Three machine learning models, least-squares boosting, random forest bagging, and neural network regression, are compared with weighted ensemble variants. The obtained results show that historical information alone is insufficient for accurate hourly 24-hour-ahead prediction, with the best strict blind model achieving R 2 = 0.3551 and RMSE=445.50 kW. However, when day-ahead weather predictors are introduced, the LSBoost model reaches RMSE=2.29 kW and R 2 = 0.99998. This weather-informed result is interpreted as an idealized upper-bound experiment rather than as a directly deployable operational forecast, because the target renewable generation is synthetically constructed from the same meteorological variables and because realized target-hour NASA POWER variables are used instead of numerical weather prediction forecasts. This contrast does not merely improve accuracy; it exposes the physical reason behind renewable forecastability. Meteorological information is therefore a central driver of one-day-ahead renewable-power prediction. The proposed dual-scenario strategy therefore provides both a lower-bound historical evaluation and an upper-bound weather-informed benchmark for energy planning, microgrid scheduling, and battery-aware renewable management.