Regime-Aware Forecasting of Customer Water Consumption: Segmented Evaluation, Fair Benchmarking, and Significance Testing for Classical and LSTM Models

Abstract

Planning clean-water production and distribution at a Regional Drinking-Water Utility (PDAM) requires reliable consumption forecasts at the per-customer-address level—a series with many small and zero values that conventional percentage metrics assess poorly. Prior water-forecasting studies typically report aggregate accuracy and claim the superiority of complex models without fair baselines, per-demand-pattern evaluation, or significance testing. This study proposes a regime-aware evaluation framework for forecasting customer water consumption at PDAM Tirta Langkisau Batang Kapas (1,407 customers; 47,547 records; January 2023–December 2025). Each customer is classified into one of four demand regimes (smooth, intermittent, erratic, lumpy) using the Syntetos–Boylan scheme, after which nine models are compared fairly: Naive, Seasonal-Naive, Moving Average, Croston, SBA, TSB, and three multi-input LSTM variants (MSE, two-stage, and Tweedie loss). Models are evaluated with scaled metrics (WAPE, MASE, RMSSE) across multiple seeds and tested with the Diebold–Mariano test. The Naive baseline proves very strong (aggregate WAPE 29.90%) and remains superior on the smooth, intermittent, and erratic regimes; intermittent-demand methods (Croston/SBA/TSB) do not help because intermittency is mild (median ADI ≈ 1). The Tweedie-loss LSTM achieves the lowest RMSE (6.54 m³) and is the only model that significantly outperforms Naive on squared error (DM = −5.367; p < 0.001), while address embedding provides no measurable advantage. The main contribution is not a new algorithm but a fair, segmented, regime-aware framework combining demand-regime classification, multi-seed scaled benchmarking, and significance testing—making claims of model superiority verifiable and transferable to other utilities.