ABSTRACT Graphical abstract combining two example water-demand time-series under concept drift with a comparison of forecast accuracy and runtime. The top panel shows Area 1, a period with reduced variability, and the second panel shows Area 2, a shape-change period with changing demand patterns over time. Two summary boxes report average MAE of 391.2 and runtime of 27.6 seconds. A scatter plot of average MAE versus runtime shows online models clustered at much lower runtimes than batch models, with the Online Ensemble near the lower-left region, indicating strong accuracy with high computational efficiency. This study examines whether incremental learning yields operationally superior 1-hour-ahead urban water demand forecasts under nonstationarity compared with periodic batch retraining. Using 6 years of hourly consumption and meteorological data from a German utility, we conduct a deployment-faithful prequential evaluation in which forecasts are generated sequentially, and models are updated either (i) incrementally after each observation or (ii) via weekly batch retraining under rolling and expanding windows. Beyond accuracy mean absolute error (MAE)/root mean squared error (RMSE) and Diebold–Mariano tests, we quantify update efficiency via end-to-end runtime to characterize the accuracy-latency trade-off relevant for real-time operation. Across two evaluation areas with distinct drift signatures (variance/level drift versus distributional/shape drift), an online averaging ensemble of three incremental learners achieves the lowest MAE/RMSE and mitigates high-error episodes that dominate RMSE. While several expanding-window batch models are statistically indistinguishable from the online ensemble in Area 1, the online ensemble remains consistently faster (seconds versus minutes/hours) and retains strong accuracy in Area 2, where most batch baselines degrade. Overall, the findings clarify how adaptation speed and computational cost interact with drift intensity, informing the choice between incremental updating and periodic batch retraining in operational water demand forecasting.
Hans et al. (Thu,) studied this question.