Operational Performance Assessment of PV-Powered Street Lighting: A Comparative Study of Different Machine Learning Prediction Models

The widespread adoption of photovoltaic (PV) energy contributes to carbon emissions reduction, cost minimization, and climate change mitigation, particularly in applications such as street lighting. However, efficiency is negatively impacted by variations in weather and temperature. Despite the growing interest in PV-powered street lighting, a significant research gap remains in predicting the operational performance of such systems, as research on PV performance prediction is constrained by the scarcity of accurate data on energy consumption values corresponding to varying lighting intensities and the impact of environmental factors. In this paper, a comparative analysis of five machine learning algorithms—Linear Regression (LR), Expectation Maximization (EM), Random Forest (RF), Deep Belief Networks (DBNs), and Deep Neural Networks (DNNs)—is conducted to predict the operational conditions of PV-powered streetlights based on luminance levels and energy consumption. A comprehensive database is created and used to train the proposed algorithms. The results indicate that DNNs and DBNs algorithms achieve the lowest error rate (2.5%) and highest accuracy (97%) with high-quality data. LR, meanwhile, is distinguished by its rapid response time, enabling predictions with minimal training on large datasets. The computational expense differs among algorithms; DNNs and DBNs require significant resources and have lengthy training periods, whereas RF and LR are faster and better suited for real-time applications. DNNs demonstrate superior performance in recall, F1 scores, and ROC-AUC, while LR shows the lowest performance.