Machine Learning Approaches to Short-Term Solar Irradiance Forecasting: Performance Comparison of Random Forest and Multilayer Perceptron Models

This study presents a comprehensive framework for short-term solar irradiance forecasting in the Adana Organized Industrial Zone, one of Türkiye’s most solarrich regions. High-resolution irradiance data collected throughout 2024 using ISO 9060–compliant pyranometers were analyzed to capture seasonal and daily variations, with annual cumulative global horizontal irradiance measured at ~1346 kWh/m2. Two machine learning approaches were applied to predict hourly irradiance: Random Forest (RF) regression and multilayer perceptron (MLP) neural networks. Input features included persistence-based irradiance values, diurnal and seasonal indicators, and time-series variables. Model performance was evaluated using mean absolute error, root mean square error (RMSE), and R2 metrics. Results indicated that RF achieved superior accuracy (R2 = 0.891, RMSE = 0.472 kWh/m2) compared to MLP (R2 = 0.874, RMSE = 0.514 kWh/m2), highlighting the robustness of ensemble methods for short-term forecasting. Uncertainty analysis confirmed that measurement errors and cloudy-day conditions remain key challenges. Unlike previous studies, this work provides one of the first high-resolution, region-specific datasets for industrial-scale solar forecasting in Türkiye. Overall, the findings demonstrate that data-driven forecasting can effectively support photovoltaic (PV) system operation, smart grid integration, and regional energy planning, providing valuable insights for the sustainable energy transition in Türkiye.