Does ANN really acquire the physics of the system? A study using conceptual components from an established water balance model

Abstract

Artificial neural networks (ANNs) are labeled as black-box techniques which limit their operational uses in hydrology. Recently, researchers explored techniques that provide insight into the various elements of ANN and their relationship with the physical components of the system being modeled which are commonly known as knowledge extraction (KE) techniques. However, the physical components of rainfall-runoff (RR) process utilized in these KE techniques are obtained from primitive baseflow separation techniques without considering other components of RR process utilizing mostly base flow and surface flow till now. To identify if ANN acquires physical components of RR process, a well-established water balance model (Australian Water Balance Model) has been utilized first time in this study. For this purpose, correlation and visualization techniques have been used for the Kentucky River basin, USA. Results show that ANN architecture having a single hidden layer with four hidden neurons was the best in simulating RR process and each of the four hidden neurons corresponds to certain subprocesses of the overall RR process, i.e., two hidden neurons are capturing surface flow dynamics with lower and higher flows, one is capturing base flow dynamics, and last one is having good relations with past rainfalls showing that ANN captures physics of basin's RR process.