Machine learning-based real-time prediction of lighting classes in adaptive road lighting systems

This study presents a machine learning-based framework for real-time prediction of road lighting classes in adaptive lighting systems. A field implementation was conducted on Buyukdere Street in Istanbul using two years of traffic and meteorological data. Lighting scenarios were developed in line with EN 13201 standards, and simulations with dimmable LED luminaires ensured compliance with lighting quality requirements. Hourly traffic flow and vehicle speed were predicted using Extreme Gradient Boosting and mapped to appropriate lighting classes via a rule-based matrix. This predictive approach enables the field-installed system to determine the lighting class in advance for each hour rather than after the period has elapsed. Six months of field data were used to validate the approach across seasonal and daily variations. The models achieved R ² scores of 0.864 and 0.822 for flow and speed predictions, respectively. Predicted lighting classes reached 92.1% and 89.7% agreement with reference classes in two directions. The resulting energy savings closely matched those obtained from measured traffic-based classification, reaching 32.1% compared with non-adaptive M2-level lighting. The results demonstrate the feasibility of data-driven adaptive lighting for energy-efficient urban road operation while meeting the lighting class requirements prescribed by the standard.