Comparative analysis of machine learning techniques for workplane illuminance estimation and photosensor placement assessment in DLCs

Daylight-linked control systems (DLCs) can provide several advantages. Nonetheless, several variables (e.g. sensor location and commissioning configuration) have a substantial influence on the behaviour of the system. One way to find good correlations between the illuminance measured on one surface and the one on the workplane is to use artificial intelligence (AI) techniques, namely machine learning (ML) approaches. This study develops and compares seven ML models for workplane illuminance estimation in a DLCs context. The analysis explicitly evaluates the effect of photosensor location on estimation performance by considering two ceiling photosensor cases (C1 and C2). The ML algorithms, including XGBoost, gradient boosting, LSTM, FFANN, random forest, decision tree and KNN, are implemented and evaluated in terms of key performance metrics such as MAE, RMSE and R2. To align the comparison with DLCs operation, a range-based assessment has also been performed in the control-relevant 300–1000 lx band, and a setpoint-based dim/no-dim decision evaluation has been conducted at 500 lx. The results reveal that the XGBoost model provides the best full-range regression performance across the complete illuminance dataset, achieving test R2 greater than 0.97 for different test scenarios. Moreover, the setpoint-based analysis highlights that KNN achieves the highest dimming decision accuracy near the 500 lx threshold in both sensor cases, indicating that models with similar regression metrics can differ in control reliability. These findings support a control-oriented and deployment-relevant interpretation of model selection for DLCs, beyond generic algorithm benchmarking.