Performance assessment of Bayesian Causal Modelling for runoff temporal behaviour through a novel stability framework

Abstract

A strong innovative tendency is nowadays emerging that largely comprises new hydrological modelling approaches, based on Causal Reasoning through Probabilistic Graphical Modelling (PGM), because its ability to support probabilistic reasoning from data with uncertainty. These novel modelling frameworks are quite diverse and disperse not only in terms of techniques but also regarding its aims. It seems necessary to find a general and robust methodology for assessing its performance. This paper aims to provide a novel general methodology for assessing the performance of PGM based on Bayesian Causality for modelling and analysing the riverś runoff behaviour. For it, a structured four-step approach is developed and showed throughout the paper. The proposed methodology begins with the identification of the two main factors that condition the Bayesian Causal (BC) Modelling: the number of synthetic series (data amount) and the number of intervals for probability distributions for training and learning processes. The developed analysis comprises the definition of three levels for the first factor and seven levels for the second one, as well as the design of an innovative stability framework that assesses the level of BC Modelling performance. Furthermore, it has been necessary to create-define two novel indexes, named “Similarity and Stability Indexes” from 21 scenarios arising from the combination of the levels of the both factors. The optimal combination of factors is identified through a bi-objectives recursive approach based on previous indexes. Main results drawn successfully show a high relationship between the level of modelling performance, measured in terms of stability, and the river runoff temporal behaviour, measured in terms of temporal dependence. This research may help water managers and engineers to develop more rigorous and robust hydrological causal modelling implementations.