Modernising Flood Forecasting with Ensemble Hydrological Modelling and AI

Flood forecasting and preparedness have become increasingly critical as extreme weather events become more intense and widespread due to climate change. We learned this the hard way… Traditional flood forecasting approaches, which rely heavily on physics-based numerical models, often struggle to provide accurate and timely predictions, especially for flash floods and in data-scarce regions. However, the rapid advancements in ensemble hydrological modelling and artificial intelligence (AI) have opened up new opportunities to modernize flood forecasting and enhance its predictive capabilities.

Now, this might seem counterintuitive…

Hybrid Flood Forecasting Approaches

One of the promising avenues for improving flood forecasting is the use of hybrid modelling approaches that combine the strengths of physics-based and data-driven models. These hybrid systems employ a diverse range of techniques, including:

1. Statistical-Dynamical Models: These models typically use the outputs from numerical weather prediction (NWP) models or Earth system models (ESMs) as inputs to a statistical or machine learning (ML) model, which then produces the final flood forecast. This is the most common type of hybrid model.

2. Serial Models: These models combine data-driven and physics-based models sequentially, often with additional intermediate steps like hydrological modelling.

3. Coupled or Parallel Models: In these approaches, the data-driven and physics-based models are run in parallel, with the outputs combined to produce the final forecast.

By leveraging the complementary strengths of these different modelling approaches, hybrid flood forecasting systems can better capture the complex interactions between atmospheric, hydrological, and anthropogenic factors that contribute to flood events.

Ensemble Hydrological Modelling

Ensemble modelling is another key component of modern flood forecasting. Instead of relying on a single simulation, ensemble models generate multiple forecasts based on different initial conditions, model parameterizations, or input datasets. This ensemble approach provides a more comprehensive representation of the uncertainty associated with flood predictions, allowing for better risk assessment and decision-making.

One example of an operational ensemble hydrological forecasting system is the European Flood Awareness System (EFAS), which uses an ensemble of hydrological models driven by meteorological forecasts from multiple NWP models. This ensemble-based approach has shown improved skill in predicting flood events across Europe compared to standalone hydrological models.

Artificial Intelligence for Flood Forecasting

The rapid advancements in AI and machine learning have also had a significant impact on flood forecasting. Data-driven models, such as long short-term memory (LSTM) neural networks, have demonstrated the ability to outperform traditional hydrological models, particularly in regions with limited data availability.

These AI-powered models can learn complex, nonlinear relationships between various hydrometeorological variables and flood characteristics, without the need for explicit physical parameterizations. Additionally, they can effectively assimilate a wide range of data sources, including satellite observations, weather forecasts, and real-time sensor measurements, to improve the accuracy and timeliness of flood predictions.

One notable example is the Google Flood Forecasting Model, which uses machine learning to generate operational, public-facing forecasts of water levels up to 6 days in advance, based on real-time weather forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF).

Benefits of Hybrid and AI-Driven Flood Forecasting

The integration of ensemble hydrological modelling and AI-powered approaches in hybrid flood forecasting systems offers several key benefits:

1. Improved Predictive Accuracy: By combining the physical understanding of hydrological processes from physics-based models with the data-driven learning capabilities of AI, hybrid models can outperform standalone models in terms of forecasting skill, particularly for extreme events.

2. Enhanced Uncertainty Quantification: Ensemble-based approaches provide a more comprehensive representation of the uncertainties associated with flood predictions, enabling better risk assessment and decision-making.

3. Reduced Computational Demands: Data-driven models, such as those based on deep learning, can be significantly faster to train and run compared to traditional physics-based hydrological models, which is crucial for operational applications.

4. Improved Operational Convenience: The ease of use and adaptability of AI-powered models can facilitate the integration of flood forecasting systems into existing operational workflows, making them more accessible to end-users and decision-makers.

5. Ability to Leverage Big Data: AI-based models can effectively assimilate and learn from large, diverse datasets, including satellite observations, weather forecasts, and real-time sensor measurements, to enhance the predictive capabilities of flood forecasting systems.

Challenges and Future Directions

While the advancements in hybrid and AI-driven flood forecasting are promising, there are still several challenges and areas for further research and development:

1. Physically Plausible and Explainable Predictions: Ensuring that the AI-powered models produce physically realistic and interpretable results, especially for extreme events, is an ongoing challenge that requires further investigation.

2. Incorporation of Human Influences: Integrating the impacts of human activities, such as land use changes, water management, and infrastructure development, into flood forecasting models is crucial but often hindered by data limitations.

3. Seamless Prediction Across Timescales: Developing hybrid systems that can seamlessly link short-term weather-driven forecasts with longer-term climate-driven predictions is an important area for improving the utility of flood forecasting for decision-makers.

4. Spatial Variability and Scaling: Addressing the challenges of modelling the spatial variability of hydrological processes and scaling up from local to regional or global forecasting systems is a key research priority.

5. Operational Uptake and Trust-Building: Fostering closer collaboration between the hydrological, forecasting, and machine learning communities, and demonstrating the added value of hybrid and AI-driven approaches, is crucial for driving the operational adoption of these new technologies.

As the impacts of climate change continue to intensify, the need for robust and adaptive flood forecasting systems has never been greater. By embracing the power of ensemble hydrological modelling and artificial intelligence, the flood control and water management community can modernize their approaches and better prepare communities for the growing threat of floods.

Visit Flood Control 2015 to learn more about the latest advancements in flood forecasting and other flood control technologies.

Tip: Implement real-time monitoring to swiftly respond to flood risks