Comparative Analysis of Hybrid Machine Learning Models for Photovoltaic Energy Output Prediction: A Case Study from Nigeria's South-South Region

Abstract

This study evaluated machine learning models for predicting photovoltaic (PV) energy output using daily meteorological data (temperature, relative humidity, wind speed, and solar radiation) from Nigeria’s South-South region over a decade (2012–2022). Data from Uyo, sourced from the University of Uyo’s Geography Department, were aggregated with PV output metrics and processed into a time-series dataset. Models including support vector machine (SVM), multiple linear regression (MLR), artificial neural network (ANN), deep neural network (DNN), gradient boosted decision tree (GBT), and hybrid architectures—convolutional neural network with long short-term memory (CNN-LSTM), reinforcement learning with LSTM (RL-LSTM), and their attention-enhanced variants (CNN-A-LSTM and Reinforcement Learning with Attention-enhanced LSTM (RL-A-LSTM)—were assessed using error metrics (MSE, MAE, RMSE, MAPE) and success rate (SR). The RL-A-LSTM model outperformed others, achieving the lowest MSE (0.004), MAPE (2.2%), MAE (0.004), and RMSE (0.063), with the highest SR (0.97), demonstrating exceptional accuracy in daily PV energy forecasting. CNN-A-LSTM followed closely (MSE: 0.005, SR: 0.96), while traditional models like SVM (MSE: 0.015) and MLR (MSE: 0.025) lagged due to limited nonlinear adaptability. The results accentuate the superiority of reinforcement learning-based hybrid models in capturing complex spatiotemporal dependencies, enabling precise predictions even under fluctuating humidity and irradiance conditions. Practically, these models enhance grid stability and operational efficiency by reducing forecast errors by up to 83% compared to conventional methods, enabling proactive energy management and cost savings. This study advocates integrating RL-A-LSTM into real-world PV systems to optimize renewable energy utilization, aligning with global decarbonization goals. The findings highlight the transformative potential of attention mechanisms and reinforcement learning in advancing sustainable energy systems.