In the race against time to predict and mitigate flash floods, a new systematic literature review published in *Intelligent Construction and Sustainable Cities* (formerly *Smart Construction and Sustainable Cities*) offers a roadmap for leveraging machine learning (ML) to improve forecasting accuracy. Led by Leonardo B. L. Santos of the National Center for Monitoring and Early Warning of Natural Disasters (Cemaden), the study highlights both the promise and challenges of ML-based hydrological models, with significant implications for the energy sector and disaster management.
Flash floods are among the most destructive natural disasters, causing billions in damages annually. Traditional physically based hydrological models often fall short at the small spatiotemporal scales typical of flash floods due to their computational complexity and reliance on detailed local data, which are rarely available during emergencies. Enter machine learning, a field that has seen explosive growth in recent years, thanks to the increasing availability of hydrological data.
Santos and his team reviewed over 1,200 papers published until January 2024, narrowing them down to 50 studies that met stringent criteria, including fine-scale temporal data (less than 6 hours) and a focus on forecasting rather than post-flood damage assessment. The results reveal a sharp rise in ML-based flash flood research, with China leading the charge at 38%.
“Nearly all studies rely on rainfall, discharge, and water level data—often in combination,” Santos explains. “Long short-term memory (LSTM) networks dominate the landscape, accounting for 60% of the models reviewed.” However, the study also uncovers a significant barrier to progress: only 10% of the selected studies provide access to their datasets. This lack of transparency hampers reproducibility, inhibits fair comparative evaluation of models, and ultimately slows methodological advancements.
The energy sector, in particular, stands to benefit from improved flash flood forecasting. Power plants, transmission lines, and other critical infrastructure are vulnerable to flood damage, and accurate predictions could help utilities mitigate risks and minimize downtime. “The variability in model performance is likely influenced by regional hydrological characteristics, input data quality, and the length of the forecast horizon,” Santos notes. “This underscores the need for benchmarks, open data practices, and stronger multidisciplinary collaboration.”
The review also highlights the lack of a one-size-fits-all solution. No single method consistently outperforms others, suggesting that regional and contextual factors play a significant role in model effectiveness. This variability presents both a challenge and an opportunity for the energy sector, which must tailor forecasting tools to their specific needs and geographic locations.
Looking ahead, Santos and his team recommend several key steps to advance the field. These include integrating ML-based models with early warning systems, creating standardized benchmarks for model evaluation, promoting open data practices, and fostering collaboration between hydrologists, data scientists, and energy sector professionals.
As the world grapples with increasingly frequent and severe weather events, the insights from this systematic literature review offer a crucial step forward in the fight against flash floods. For the energy sector, the implications are clear: improved forecasting tools could save lives, protect infrastructure, and enhance resilience in the face of climate change.