SM2D: A Modular Knowledge Discovery Approach Applied to Hydrological Forecasting

In this paper, we address the problem of flood prediction in complex situations. We present an original solution in order to achieve the main goals of accuracy, flexibility and readability. We propose the SM2D modular data driven approach that provides predictive models for each sub-process of a global hydrological process. We show that this solution improves the predictive accuracy regarding a global approach. The originality of our proposition is threefold: (1) the predictive model is defined as a set of aggregate variables that act as classifiers, (2) an evolutionary technique is implemented to find best juries of such classifiers and (3) the flood process complexity problem is addressed by searching for sub-models on sub-processes identified partly by spatial criteria. The solution has proved to perform well on flash flood phenomena of tropical areas known to be hardly predictable. It was indeed successfully applied on a real caribbean river dataset after both preprocessing and preliminary analysis steps presented in the paper.