Timing error correction procedure applied to neural network rainfall-runoff modelling

Several studies have observed that neural network models will often produce phase-shift errors or timing tags in their output results. This paper investigates a potential solution to the timing error problem through the application of a procedure first applied in sunspot prediction. This procedure was applied to two neural network hydrological forecasting models for the River Ouse, in northern England, using a neuro-evolution toolbox. Models were optimised on a combination of root mean squared error and a timing correction factor. The application of this correction procedure produced timing improvements of up to about six hours on average over shorter forecasting horizons, whereas longer horizons showed little or no overall improvement in timing. The correction procedure also produced improved lower-magnitude estimates at the expense of higher-magnitude events over shorter forecasting horizons and, more significantly, improved higher-magnitude estimates at the expense of lower-magnitude events over longer forecasting horizons.