GC-140-126 Machine Learning Models for Solar Power Output Prediction: A Comparative Study with Adaptive PSO-Based Random Forest Tuning

GC-140-126 Machine Learning Models for Solar Power Output Prediction: A Comparative Study with Adaptive PSO-Based Random Forest Tuning

https://www.kennesaw.edu/ccse/events/computing-showcase/sp26-cday-program.php

Accurate prediction of solar power output is essential for energy scheduling, grid reliability, and efficient integration of renewable resources. Because photovoltaic generation is governed by changing atmospheric conditions, forecasting output is inherently a nonlinear learning problem. This study evaluates four machine learning models — Linear Regression, Random Forest, Multi-Layer Perceptron (MLP), and an Adaptive Particle Swarm Optimization-tuned Random Forest (Adaptive PSO-RF) — using irradiance, temperature, humidity, wind speed, cloud cover, and time-derived features drawn from a dataset of 6,738 observations. Random Forest achieved the strongest overall performance, with an RMSE of 1,816.24 and an R² of 0.9533. The Adaptive PSO-RF produced nearly equivalent results, confirming that the baseline Random Forest parameters were already near-optimal for this dataset. Feature interpretation was examined through both impurity-based and permutation importance. Whereas impurity-based importance erroneously ranked humidity as the dominant predictor, permutation importance yielded a physically meaningful ranking in which irradiance and hour emerged as the primary drivers. The study demonstrates that nonlinear ensemble methods substantially outperform linear regression for solar power forecasting and that careful interpretation of feature importance is necessary whenever predictors are correlated.