This study presents a novel approach for urban flood forecasting in drainage systems using a dynamic ensemble-based data mining model which has yet to be utilised properly in this context. The proposed method incorporates an event identification technique and rainfall feature extraction to develop weak learner data mining models. These models are then stacked to create a time-series ensemble model using a decision tree algorithm and confusion matrix-based blending method. The proposed model was compared to other commonly used ensemble models in a real-world urban drainage system in the UK. The results show that the proposed model achieves a higher hit rate compared to other benchmark models, with a hit rate of around 85% vs 70 % for the next 3 h of forecasting. Additionally, the proposed smart model can accurately classify various timesteps of flood or non-flood events without significant lag times, resulting in fewer false alarms, reduced unnecessary risk management actions, and lower costs in real-time early warning applications. The findings also demonstrate that two features, “antecedent precipitation history” and “seasonal time occurrence of rainfall,” significantly enhance the accuracy of flood forecasting with a hit rate accuracy ranging from 60 % to 10 % for a lead time of 15 min to 3 h.

Urban flooding is a major problem for cities around the world, with significant socio-economic consequences. Conventional real-time flood forecasting models rely on continuous time-series data and often have limited accuracy, especially for longer lead times than 2 hrs. This study proposes a novel event-based decision support algorithm for real-time flood forecasting using event-based data identification, event-based dataset generation, and a real-time decision tree flowchart using machine learning models. The results of applying the framework to a real-world case study demonstrate higher accuracy in forecasting water level rise, especially for longer lead times (e.g., 2–3 hrs), compared to traditional models. The proposed framework reduces root mean square error by 50%, increases accuracy of flood forecasting by 50%, and improves normalised Nash–Sutcliffe error by 20%. The proposed event-based dataset framework can significantly enhance the accuracy of flood forecasting, reducing the occurrences of both false alarms and flood missing and improving emergency response systems.