Numerical weather models as virtual sensors to data-driven rainfall forecasts in urban catchments

Abstract

Weather and rainfall forecasts play a key role in operational urban water management, since they allow anticipating the effects of storms and flash floods and initiating alarms in a timely manner. Short-term rainfall forecasts are commonly based on radar nowcasting techniques, which provide high-resolution forecasts limited to a temporal horizon of few hours. Numerical weather models can extend this horizon further in time, but their coarser spatial resolution is often a limit when working on small urban catchments. In this study, with the purpose of providing long-term and reliable rainfall forecasts in a small urban area, we investigate the use of numerical weather models as virtual sensors. The Weather Research and Forecast (WRF) model, fed with real-time Global Forecast System (GFS) data, is implemented for Singapore spatial domain and used to get a time-continuous and spatially-distributed monitoring of the atmospheric processes. The model state variables (e.g. humidity, air temperature, heat exchange etc.) are then processed by an automatic input variable selection algorithm to single out the most relevant variables used by a data-driven model to yield rainfall forecasts at the catchment scale. In this work, we explore different lead times (up to 12 hours) to evaluate the reliability of this approach, as well as different meteorological seasons. A comparison against the forecasts issued by WRF shows that the prediction accuracy of the input selection-based data-driven models can be improved, especially for long-term predictions (up to 12 hours). Results show that for short lead times (up to 3 hours) the heat flux is the most relevant driver, while for longer lead times a combination of drivers, such as wind and temperature, is selected. Also, such combination varies with the meteorological season. This techniques can thus be adopted to improve the accuracy of rainfall forecasts, although it must be noted that the overall accuracy is still influenced by the underlying numerical model. Indeed, if the numerical weather model does not adequately represent some events (e.g. small-scale convective storms), the selected variables cannot be successfully adopted to yield a rainfall forecast. Further research will focus on two different aspects. First, the rainfall forecasts will be included in an operational framework, in order to provide long-term inflow predictions to Marina Reservoir and to assess their impact on the barrage operation. Second, the results here discussed will be further investigated, by running the input selection algorithm on a larger datasets and by considering different combinations of lead times and meteorological seasons.