The role of event identification in translating performance assessment of time-series urban flood forecasting

Today, urban flood forecasting becomes a hydrological hot topic due to urbanisation, population growth, the staggering rise in weather extremes and its significant consequences such as economic, social and infrastructural losses. To address this, multiple time-series urban flood forecasting models have been developed recently. These models inevitably require relatively long-range and continuous time-series data. However, this database usually is mixed by the large share of unnecessary data, particularly dry weather conditions in which there is no flood occurrence. In this situation, the conventional model performance is assessed based on the entire database and consequently is diluted by prediction results of unnecessary data in practice. To overcome this, this paper aims to propose a framework to identify the three different approaches including rainfall-based, water level -based and hybrid method. In each method, model performance is determined based on a precise portion of data entitled events to remove the effects of dry weather conditions. Events are defined as a part of the database which represents urban flooding. Differences in these methods are illustrated through a real case study of urban flood forecasting in Ruislip?s drainage system. The nonlinear autoregressive network with exogenous model is used for predictions of 15-minute, 1-hour, 2-hour and 3-hour steps ahead. Furthermore, mean absolute error, root mean square error, coefficient of determination and Nash?Sutcliffe model efficiency coefficient are evaluated as performance assessment indicators. The results reveal the role of event identification in the performance assessment of these models. While the conventional method shows the best performance, indicators are expected to become worth in the rainfall-based, the water level-based and hybrid methods, respectively.