Detection of conceptual model rainfall—runoff processes inside an artificial neural network

The internal behaviour of an artificial neural network rainfall-runoff model is examined and it is demonstrated that specific architectural features can be interpreted with respect to the quasi-physical dynamics of a parsimonious water balance model. Neural network solutions were developed for daily discharge series simulated by a conceptual rainfall-runoff model given observed daily precipitation totals and evaporation rates for the Test River basin in southern England. Neural outputs associated with each hidden node, produced from the output node after all other hidden nodes had been deleted, were then compared with state variables and internal fluxes of the conceptual model (including soil moisture, percolation, groundwater recharge and baseflow). Correlation analysis suggests that hidden nodes in the neural network correspond to dominant processes within the conceptual model. In particular, different hidden nodes are associated with distinct "quickflow" and "baseflow" components, as well as a threshold state in the soil moisture accounting. The results also demonstrate that, for this river basin, a neural network with seven inputs and three hidden nodes can emulate the gross behaviour of the conceptual model.