AV2D: A Data-Driven Hydrological Forecasting Approach Based on Aggregate Variables

Abstract

In this study, we address the difficult problem of flash flood prediction in Caribbean watersheds for which current hydrological system are not well suited. These basins have small surface areas with steep slopes due to their volcanic origin, and they are subjected to tropical rainfall conditions such as massive and localized precipitations. We propose a data-driven solution whose main originality is two-folded: (1) the predictive model is defined as a set of aggregate variables that act as classifiers, (2) an evolutionary algorithm is implemented to find best juries of such classifiers. The design of this solution was guided by the necessity to reach three main objectives: precision, readability, and flexibility. Indeed, a flood forecasting solution should not only provide accurate prediction performances, but it should also give clear explanations about how and why an alert is triggered or not on one hand, and be easily adaptable on similar catchments on the other hand. The concept of aggregate variables allow to reach the objective of readability by using simple rules based on threshold over-passing of aggregated values, while the data-driven nature of the solution and the use of combinations of aggregate variables allow to reach the objective of flexibility. The results obtained for the case study of a typical Caribbean river, for which runoff data are available at three locations, demonstrate the efficiency of the solution.