Multi-model data fusion for river flow forecasting: an evaluation of six alternative methods based on two contrasting catchments

摘要

This paper evaluates six published datafusion strategies for hydrological forecasting based on two contrastingcatchments: the River Ouse and the Upper River Wye. The input level anddischarge estimates for each river comprised a mixed set of single modelforecasts. Data fusion was performed using: arithmetic-averaging, aprobabilistic method in which the best model from the last time step is used togenerate the current forecast, two different neural network operations and twodifferent soft computing methodologies. The results from this investigation arecompared and contrasted using statistical and graphical evaluation. Eachlocation demonstrated several options and potential advantages for using datafusion tools to construct superior estimates of hydrological forecast. Fusionoperations were better in overall terms in comparison to their individualmodelling counterparts and two clear winners emerged. Indeed, the six differentmechanisms on test revealed unequal aptitudes for fixing different categories ofproblematic catchment behaviour and, in such cases, the best method(s) were agood deal better than their closest rival(s). Neural network fusion ofdifferenced data provided the best solution for a stable regime (with neuralnetwork fusion of original data being somewhat similar) — whereas a fuzzifiedprobabilistic mechanism produced a superior output in a more volatileenvironment. The need for a data fusion research agenda within the hydrologicalsciences is discussed and some initial suggestions are presented. style="line-height: 20px;">Keywords: data fusion, fuzzy logic, neural network, hydrological modelling