Rainfall intervention analysis for on-line applications

Rainfalls during summer season are a major cause of sporadic nonlinear transient drops in daily municipal water consumption. The nonhomogeneous, nonlinear effects induced by rainfall interventions on water use complicate time series model identification and estimation. An iterative computer algorithm, that employs a model-switching transfer function, is proposed for sequential estimation of the transient drops in the water consumption, so that they can be removed from the time-series data. Existing time series analysis' techniques, which are based on homogeneous, covariance stationary assumptions are not directly applicable, since the water-use time series never reaches statistical equilibrium. The practical data transformation procedure introduced in this paper is useful for achieving approximately homogeneous and stationary time series prior to model identification and estimation of rainfall intervention effects. The resulting empirical model is a transfer function of intervention time, number of uninterrupted raining days, a moving average of maximum daily temperatures, and a moving average of the most recent water-use observations. This rainfall intervention model can also be used for online prediction of temporal changes in daily water consumption of a city, given rainfall forecasts. An example is included to illustrate the applicability of this approach, using a record of municipal water use.