Accurate prediction of indoor illuminance distribution is essential for enhancing occupant visual comfort and reducing building energy consumption. Indoor illuminance exhibits a dual-domain characteristic defined by stochastic temporal variability dri...
Accurate prediction of indoor illuminance distribution is essential for enhancing occupant visual comfort and reducing building energy consumption. Indoor illuminance exhibits a dual-domain characteristic defined by stochastic temporal variability driven by changing sky conditions and spatial heterogeneity governed by building geometry, façade configuration, and surface reflectance. However, conventional single-algorithm models are structurally limited in capturing both dimensions simultaneously, constraining their effectiveness for multi-point indoor prediction and real-time lighting control.
To address this limitation, this study proposes and evaluates a hybrid prediction framework that integrates the temporal sequence learning capability of Long Short-Term Memory (LSTM) networks with the nonlinear feature interaction modeling of Extreme Gradient Boosting (XGBoost). The framework was developed and validated using 340,312 one-minute illuminance measurements collected over three years (August 2022–December 2024) from a multi-sensor network deployed in a small office testbed (7.9 m × 2.8 m × 3.1 m) in Seoul, South Korea, linked to a Daylight-Responsive Dimming System (DRDS). Workplane illuminance was predicted at three representative positions—window-side (E1), central area (E2), and corridor-side (E3)—using nine photosensors (wall-mounted, ceiling-mounted, and luminaire-integrated) and categorical sky condition codes as input variables.
Five models were comparatively evaluated: LSTM A, LSTM B, XGBoost, Hybrid A (LSTM A + XGBoost), and Hybrid B (LSTM B + XGBoost), assessed on a held-out test set using R², RMSE, and MAE metrics. Results demonstrated that optimal model performance is inherently location-dependent. At the window-side position (E1), where rapid daylight fluctuations are dominant, the temporally richer Hybrid B model achieved the highest accuracy (R²=0.931). In the central zone (E2), where direct daylight and reflected light coexist in moderated conditions, the lower-complexity Hybrid A model was most effective (R²=0.888). At the corridor-side position (E3), dominated by reflected and artificially redistributed light under relatively stable interior conditions, the standalone XGBoost model outperformed both hybrid variants (R²=0.896). Across all three positions, Hybrid A demonstrated the most consistent predictive performance, achieving an average R² of 0.903.
Model-specific feature importance analyses, including SHAP analysis, provided a physically interpretable explanation for these location-specific performance differences. At E1, predictive contributions were distributed across sky condition variables, luminaire-integrated sensors, ceiling-mounted sensors, and wall-mounted sensors, reflecting the complex interplay of direct daylight, luminaire output, and surface reflections near the façade. Moving toward the interior, the dominance of wall-mounted reflection sensors (S21, S22) increased progressively, indicating a systematic shift in the governing illuminance mechanism from direct daylight to spatially redistributed and reflected light. This shift provides a physically grounded explanation for why temporally responsive hybrid models are most effective near windows, while nonlinear spatial models suffice in the corridor zone. Performance stratification under varying sky conditions further confirmed that predictive robustness is spatially conditioned: the same external sky condition generated different prediction challenges depending on the measurement location and its underlying lighting regime.
The present findings establish that indoor illuminance prediction should be interpreted as a location-dependent spatiotemporal problem rather than a uniform regression task. No single modeling strategy performs optimally across all indoor zones, and model suitability is governed by how strongly each location is characterized by rapid temporal daylight variation versus stable spatial redistribution of light. These results offer a physically interpretable and empirically grounded basis for location-aware model deployment in Daylight-Responsive Dimming Systems (DRDS), Building Energy Management Systems (BEMS), and intelligent lighting control applications, supporting zone-specific dimming strategies that simultaneously prevent over-dimming and under-dimming to balance visual comfort and energy efficiency.