A Review of AI Methods for the Prediction of High-Flow Extremal Hydrology

Forecasting systems for foreseeing water levels and flow rates have become necessary to mitigate climate change negative impacts. Most of these systems are based on powerful tools such as Artificial Intelligence (AI) methods. This paper presents a comprehensive review of AI methods for high-flow extremes prediction. The review starts with an overview of the state-of-the-art AI techniques and examples of their application, followed by a SWOT analysis to benchmark their predictive capability based on set of criteria. Finally, the most suitable AI methods for short-term and/or long-term prediction, based on a rigorous suitability assessment are proposed. As a result, Fourteen AI methods have been identified. Their evaluation revealed that the methods that averagely behave the best for achieving high-flow extremes prediction are ANNs, SVMs, wavelets and Bayesian methods, at all-time scales. The latter, as stochastic methods, have the privilege by their cheap computation cost, their reliability and ability to handle hydrological uncertainty, and their capacity to perform causal relationships between features. This study also urges researchers to further explore the predictive potential of decision trees, ensembles, CNNs, MARS, GP and agent-based methods for high-flow extremes.

Automated summary

This paper provides a comprehensive review of AI methods for predicting high-flow extremes. It identifies ANNs, SVMs, wavelets, and Bayesian methods as the most suitable methods at all-time scales, and suggests exploring the potential of other methods.