Time-Series Forecasting for Water Demand: Predicting Tomorrow's Consumption Today

Accurate demand forecasting allows water utilities to optimize pumping schedules, reduce energy costs, and prevent supply shortfalls. Here is how we built a forecasting system with < 4% MAPE.

Data 05/12/2025 12 min read

#forecasting #time series #XGBoost #water #machine learning

Why Demand Forecasting Matters

Water production is energy-intensive. Pumping water uphill to reservoirs accounts for 30–60% of a water utility's energy bill. If you can predict tomorrow's demand accurately, you can optimize pump schedules to avoid peak electricity rates – saving 15–25% on energy costs.

Features for Water Demand Prediction

- Historical consumption (hourly, 96 lags)
- Day of week, hour of day, public holidays
- Temperature forecast (demand correlates strongly with heat)
- Rainfall forecast (outdoor irrigation drops when it rains)
- Seasonal decomposition components

Model Comparison Results

Model	MAPE	Training Time	Inference
ARIMA	8.3%	< 1 min	< 1ms
Prophet	6.1%	2 min	< 5ms
XGBoost	3.8%	5 min	< 1ms
N-BEATS	3.2%	45 min	< 10ms

Production System Architecture

Daily retrain (2 AM) → Feature pipeline → XGBoost → Forecast API → Pump scheduler
                                                    ↓
                                            Confidence intervals
                                            for anomaly detection

Results in Production

After 12 months in production across 8 water utilities:
- Average MAPE: 3.8% on 24h horizon
- Energy cost reduction: 18%
- Zero supply shortfall incidents
- ROI payback period: 4 months