Legitimising neural network river forecasting models: a new data-driven mechanistic modelling framework

Abstract

Abstract. In this paper we address the difficult problem of gaining an internal, mechanistic understanding of a neural network river forecasting (NNRF) model. Neural network models in hydrology have long been criticised for their black-box character, which prohibits adequate understanding of their modelling mechanisms and has limited their broad acceptance by hydrologists. In response, we here present a new, data-driven mechanistic modelling (DDMM) framework that incorporates an evaluation of the legitimacy of a neural network's internal modelling mechanism as a core element in the model development process. The framework is exemplified for two NNRF modelling scenarios, and uses a novel adaptation of first order, partial derivate, relative sensitivity analysis methods as the means by which each model's mechanistic legitimacy is explored. The results demonstrate the limitations of standard, goodness-of-fit validation procedures applied by NNRF modellers, by highlighting how the internal mechanisms of complex models that produce the best fit scores can have much lower legitimacy than simpler counterparts whose scores are only slightly inferior. The study emphasises the urgent need for better mechanistic understanding of neural network-based hydrological models and the further development of methods for elucidating their mechanisms.