Street lighting is one of the most energy-intensive municipal services, largely because most installations run at a fixed static output throughout the night regardless of conditions. This study asks a simple question: can streetlight-power demand be predicted from weather, traffic, and temporal context, and can such a model drive an adaptive-dimming controller capable of reducing consumption by a significant figure? Using four months of feeder-level data from twelve smart-lighting controllers in Delhi (30 November 2024 to 31 March 2025; approximately 143,800 fifteen-minute records), that can lead to algebraic deduction being removed and using only contextual features under a strict chronological eighty-twenty train-test split, a Random Forest predicts active power with a mean absolute error of about one kilowatt and an R-squared of 0.895, outperforming static, gradient-boosted, linear, and hour-mean baselines. Adding weather and traffic improves prediction only modestly. The model is then coupled to a luminance-anchored dimming controller whose floor equals the ratio of two road-lighting class luminances, and a luminance check verifies that road luminance never falls below the class minimum. On the held-out month, the controller reduces operating-hour power by about twenty-nine percent, a saving reproduced almost exactly when the same policy is applied to actual metered power, while remaining fully luminance-compliant. The value of the added data lies not in sharpening prediction but in enabling demand-aware control.

Keywords
Adaptive street lighting, power-demand prediction, Random Forest regression, demand-responsive dimming, and energy efficiency.