Automated novel real-time framework for rainfall data imputation in flood early warning systems

Real-time flood warning systems play a crucial role in mitigating impacts of flooding. However, their performance is highly dependent on input data, which can often contain missing values. While data imputation techniques have been widely applied in pre-processing stages, their integration into real-time operations remains underexplored. This study presents a real-time automated decision support system that integrates a soft-voting stacked data mining ensemble model comprising decision tree, K-nearest neighbour, Naive Bayes, Neural Network, Support Vector Machine, Discriminant Analysis, and Gaussian Regression. The system also incorporates hydrological–hydraulic event identification, external benchmarking, and a multi-data fuzzy weighted spatial imputation framework. The effectiveness of the proposed method was evaluated through a real-world case study involving a flood early warning system in an urban drainage network in London, UK. Comparative analyses were conducted against well-established artificial intelligence model, and a sensitivity analysis was performed for further assessment. Results showed that all types of missing data were correctly identified with a precision exceeding 90 % and were accurately imputed - particularly in situations where other models failed to recognise current rainfall values during the onset, peak, and falling limb of events (with no reduction in accuracy compared to the best-performing benchmark models). For the 3-h-ahead flood forecasting, the proposed method reduced the normalised root mean square error by up to 30 % compared to alternative approaches. To ensure the generalisability of the approach, additional locations across the UK were used for validation, which demonstrates the stability and robustness of the system, with only minor error variations.