Discussion of ‘Reservoir Computing approach to Great Lakes water level forecasting’ by P. Coulibaly [J. Hydrol. 381(2010) 76–88]

This paper extends published findings on the use of standard and recurrent neural network solutions for dynamical system modelling/monthly water level forecasting of the Great Lakes, North America. Earlier results are visualised and benchmarked using multiple linear regression. Feedforward solutions are observed to perform either linear or quasi-linear modelling operations. The superior performance of recurrent solutions is attributed to their highly dynamic, non-linear structure, influencing the manner in which feedback loops are incorporated. The echo state network was very powerful in such respects. Of particular interest, our research has also demonstrated that the inclusion of a recursive term in a linear regression model will have no impact on its predicted output.