Time Series Prior Effect on Water Demand Estimation

The current study intends to evaluate the impact of using time series models as priors on the demand estimation process. Two different priors were compared: an informationless prior and a novel approach for prior generation based on the explicit propagation of the estimated demand uncertainty through autoregressive (AR) models. The proposed novel AR-type model can be understood as a natural extrapolation of the classic AR model that considers the estimated demands as random variable inputs instead of point estimates. The proposed models were tested utilizing a realistic case study containing one week of observed measurements. In addition, since sensor failures are commonly observed in real systems, the robustness of the adopted methods was evaluated under different levels of missing flow measurements. The results indicate that the uncertain AR prior was clearly superior to the informationless prior for both failure and no failure situations resulting in consistently lower errors and much smaller confidence intervals without significant losses in terms of reliability.