Next-gen urban flood management: AI, IoT, and smart decision systems lead the way

With the intensifying threat of climate-induced flash floods in urban centers, researchers in Taiwan have developed a groundbreaking AI-powered framework to revolutionize flood management. The study, titled "Intelligent Urban Flood Management Using Real-Time Forecasting, Multi-Objective Optimization, and Adaptive Pump Operation," was published in Smart Cities. It introduces a novel, fully integrated solution known as the Intelligent Flood Control Decision Support System (IFCDSS), tested in Taipei’s Zhongshan Pumping Station catchment to enhance real-time forecasting, pump operations, and energy efficiency in flood-prone urban zones.

The IFCDSS is a three-tiered AI system that combines forecasting, optimization, and adaptive control to manage urban flood responses dynamically. Its architecture is built on four core components: a CNN-BP hybrid model for short-term water level forecasting; NSGA-III for multi-objective optimization of pumping operations; TOPSIS for selecting optimal pumping strategies; and ANFIS for adaptive pump operation predictions.

Using real-time IoT sensor data from the Zhongshan Pumping Station area, the CNN-BP model forecasts sewer, forebay, and river water levels up to 60 minutes in advance. These predictions feed into NSGA-III, which identifies optimal pump activation thresholds by balancing flood hazard mitigation, water level stability, and minimal pump switching. The TOPSIS framework selects the best strategy among hundreds of Pareto-optimal solutions, and the ANFIS model predicts the number of pumps needed within 10-minute intervals.

This integration allows for forecasting in under five seconds, supporting proactive flood mitigation during extreme weather events. Notably, during Typhoon Mitag and a major rainfall event in June 2021, the optimized operation rules outperformed manual strategies in energy efficiency and water level stability, while slightly trailing in peak flood minimization.

What performance gains were achieved through multi-objective optimization?

Traditional pump station operations rely heavily on operator experience, focusing primarily on minimizing flood hazards. IFCDSS, however, aims for a broader performance profile. By running simulations across 24 training and 14 testing rainfall events, researchers used NSGA-III to optimize three key objectives: reduce flood hazard exposure, stabilize water levels, and minimize pump switching frequency.

The optimization yielded 255 non-dominated solutions. The TOPSIS method shortlisted the best ones, which featured early pump activations at low thresholds and delayed activations at higher thresholds, allowing better energy management and more consistent operation under varying flow conditions. Compared to baseline operations, the optimized rules improved water level stability by 50.5% in training and 45.4% in testing. Pump switching was reduced by 82% in training and 28% in testing. These gains translate into substantial reductions in mechanical wear, energy consumption, and system fatigue.

Evaluation during real storm events confirmed the system’s practical benefits. For instance, during Typhoon Mitag, optimized pump strategies achieved earlier and fewer activations, resulting in lower final internal water levels and better system recovery. Even when overestimations occurred due to unmodeled gravity drainage, the system's anticipatory behavior proved advantageous.

Can AI-powered systems scale and adapt to future climate scenarios?

Beyond improved performance, the IFCDSS lays the groundwork for future-ready, autonomous flood management. By integrating real-time data analytics, AI-based forecasting, and decision-making automation, the framework provides a blueprint for smart infrastructure systems that can respond flexibly to unpredictable climate scenarios.

The ANFIS component was particularly effective in simulating human decision logic. Trained on a diverse dataset, it achieved an R2 score of 0.81 and maintained prediction errors within a margin of one pump. While the model slightly underperformed during rapid surges, its robustness during moderate and stable flow conditions indicates high practical value. Visualizations in the study showed the model reliably forecasted pump demand across a range of events, reinforcing its suitability for integration into live systems.

The researchers acknowledge some limitations, including the absence of gravity drainage effects in modeling and challenges in capturing high-variability edge cases. Future improvements could involve integrating high-resolution rainfall forecasts and direct pump data into the forecasting module.