This paper presents a methodology to forecast short-term water demands either offline or online by combining seasonal autoregressive integrated moving average (SARIMA) models with data assimilation. In offline mode, the method frequently reestimates the models using the latest historical data. In online mode, the method applies a Kalman filter to optimally and efficiently update the models using a realtime feed of data. The tailoring process consists of identifying, estimating, and validating the models, along with exploring how the length of demand history used in fitting can improve forecast performance. A suite of models are obtained that are adequate for 15-min, hourly, and daily demands having daily and weekly periodicities. The model output is analyzed across temporal resolutions, periodicities, and forecasting modes. The study finds that the normalized forecast deviations range from approximately 4.2 to 1.3%, in correspondence to a decrease in temporal granularity. Models of the weekly-seasonal type are found to more efficiently remove the autocorrelations with respect to models of the daily-seasonal type. In terms of the forecasting mode, the online implementation is shown to produce a higher performance specially for models with higher temporal resolution. Finally, a case study is conducted where forecasts are compared to the actual water production volumes of the local water utility. The results indicate that a significant improvement may be obtained in estimating the production of water based on the model output.