Flood forecasting is a key issue in hydrology. The two factors which have the most significant effect on a flood forecast are the quantitative precipitation forecast and the rainfall-runoff nowcast. The quantitative precipitation forecast is normally derived from weather forecasts. The runoff nowcast is normally derived using measured or estimated evapo-transpiration and measured rainfall and the conceptual hydrological model, and may be improved by assimilating measured discharge data. Therefore, two particle filters, sequential importance resampling (SIR) and residual resampling (RR), and an Ensemble Kalman (EnKF) filter which is capable of handling dynamic non-linear/non-Gaussian models are compared to obtain an optimal runoff nowcast with a conceptual rainfall-runoff model HBV-96. All three methods are easy to implement in real flood forecasting systems. Under normal circumstances, the residual resampling algorithm RR and EnKF performed equally well. SIR performed the least satisfactorily. With all three filters, the model error on the rainfall could be estimated during a twin experiment.
Both SIR and RR may degenerate temporarily, characterised by many identical particles. SIR and RR do not involve state updating, which may lead to significant overestimation of the runoff as a result of snow pack formation and later release of snow melt. This can be avoided by decreasing the role of these filters in favour of the forward model through temporary strong reduction of the rainfall uncertainty, e.g. when the temperature drops below 3o C. Given its state updates, EnKF forecasts adhere more closely to the data, regardless of a much smaller ensemble size. The error with respect to the evaporation could not be estimated since it acts on the same states as the rainfall, and the rainfall error was dominant in this case. Further research on the effect of the assumptions on model uncertainties and measurement uncertainties is recommended. |
