Probabilistic online flood forecasting for flash flood prone catchments in lower mountain ranges

快速响应集水区洪水预报的可靠性和短暂有效性问题一直是水文气象学中最具挑战性的问题之一。我们通过使用严格的气象/水文/水力建模以及计算效率很高的人工智能技术来考虑水文和气象模型的不确定性来解决这个问题。气象部门开发了一套从天到分钟的整体降水预报，通过(a)一种新的方法来改善国家气象局的业务预报，将提前期从18小时缩短到6小时，(b)使用物理初始化将最佳集合成员的有效性延长到3小时，以及(c)基于雷达特征跟踪的整体预报的临近预报，提高了准确性。对集合的不确定性进行量化是气象工程部分的一项主要任务。水文部分主要通过蒙特卡罗降雨径流和洪水路径模拟来处理气象预报的不确定性。包括水文模型的不确定性采用扰动方法来建立一个基于物理的随机集水区模型。随后，准随机人工神经网络（ANN-S）充分描绘了这一点，该网络源于基于不确定土壤数据的广泛训练，以及使用随机集水区模型模拟所有与洪水相关的暴雨情景。将得到的水文模型不确定性与气象不确定性相结合，可以将ANN- s与洪水路径ANN- f相耦合，ANN- f本身由水动力模型训练。因此，耦合ANN-S和ANN-F的蒙特卡罗模拟可以相对简单和快速地预测具有回水效应的河流超过临界水位的概率。

The questionable reliability and brief validity of flood forecasting for fast responding catchments remains one of the most challenging problems in hydrometeorology. We tackle this problem by considering the hydrological and meteorological model uncertainties using rigorous meteorological/hydrological/hydraulic modelling together with computationally highly efficient artificial intelligence techniques. The meteorological part develops a suite of ensemble precipitation forecasts from days to minutes with increasing accuracy by (a) a novel approach to improve operational ensembles from the national weather service for lead times from 18 hours down to 6 hours, (b) extending the validity of best ensemble members using physical initialisation for lead times down to 3 hours and (c) nowcasting based on ensemble forecasts from radar feature tracking. Quantifying the uncertainties from the ensembles is a major task of the meteorological project part. The hydrologic part processes the meteorological forecast uncertainty basically by Monte Carlo rainfall-runoff and flood routing simulations. The inclusion of the hydrologic model uncertainties employs a perturbation approach for setting up a physically based stochastic catchment model. This is subsequently fully portrayed by a quasi-stochastic artificial neural network (ANN-S), which originates from an extensive training on the basis of uncertain soil data together with simulations of all flood relevant rainstorm scenarios using the stochastic catchment model. Combining the resulting hydrologic model uncertainty with the meteorological uncertainty allows coupling the ANN-S to a flood routing ANN (ANN-F), which itself is trained by a hydrodynamic model. The Monte Carlo simulations of the coupled ANN-S and ANN-F thus allow a relatively simple and fast prediction of the probability of exceeding critical water levels also in rivers with backwater effects.