Multi-Horizon Infrastructure Forecasting for Urban Planning: A Case Study of Bengaluru

This paper proposes a horizon-aware evaluation framework comparing six forecasting models — SARIMAX, Random Forest, LightGBM, LSTM, N-BEATS, and Temporal Fusion Transformer — across four urban indicators (congestion, electricity demand, water demand, PM2.5) for Bengaluru, India, at 1-, 3-, and 5-year planning horizons. Using leakage-free rolling-origin evaluation and a novel catastrophic failure rate metric, results show that model suitability shifts fundamentally with horizon: SARIMAX degrades fourfold by H60 while deep learning models maintain stable performance, and reliability diverges sharply from accuracy at strategic timescales. The findings provide concrete, horizon-aligned model selection guidance for urban infrastructure planners. Keywords: urban forecasting · multi-horizon evaluation · LSTM · TFT · N-BEATS · LightGBM · infrastructure planning · Bengaluru