Elucidating the performance of hybrid models for predicting extreme water flow events through variography and wavelet analyses

Accurate prediction of extreme flow events is important for mitigating natural disasters such as flooding. We explore and refine two modelling approaches (both separately and in combination) that have been demonstrated to improve the prediction of daily peak flow events. These two approaches are firstly, models that aggregate fine resolution (sub-daily) simulated flow from a process-based model (PBM) to daily, and secondly, hybrid models that combine PBMs with statistical and machine learning methods. We propose the use of variography and wavelet analyses to evaluate these models across temporal scales. These exploratory methods are applied to both measured and modelled data in order to assess the performance of the latter in capturing variation, at different scales, of the former. We compare change points detected by the wavelet analysis (measured and modelled) with the extreme flow events identified in the measured data. We found that combining the two modelling approaches improves prediction at finer scales, but at coarser scales advantages are less pronounced. Although aggregating fine-scale model outputs improved the partition of wavelet variation across scales, the autocorrelation in the signal is less well represented as demonstrated by variography. We demonstrate that exploratory time-series analyses, using variograms and wavelets, provides a useful assessment of existing and newly proposed models, with respect to how they capture changes in flow variance at different scales and also how this correlates with measured flow data – all in the context of extreme flow events.