Assessing the performance of eight real-time updating models and procedures for the Brosna River

摘要

The flow forecasting performance of eight updating models, incorporated inthe Galway River Flow Modelling and Forecasting System (GFMFS), was assessedusing daily data (rainfall, evaporation and discharge) of the Irish Brosnacatchment (1207 km2), considering their one to six days lead-time dischargeforecasts. The Perfect Forecast of Input over the Forecast Lead-time scenariowas adopted, where required, in place of actual rainfall forecasts. The eightupdating models were: (i) the standard linear Auto-Regressive (AR) model,applied to the forecast errors (residuals) of a simulation (non-updating)rainfall-runoff model; (ii) the Neural Network Updating (NNU) model, alsousing such residuals as input; (iii) the Linear Transfer Function (LTF)model, applied to the simulated and the recently observed discharges; (iv)the Non-linear Auto-Regressive eXogenous-Input Model (NARXM), also a neuralnetwork-type structure, but having wide options of using recently observedvalues of one or more of the three data series, together with non-updatedsimulated outflows, as inputs; (v) the Parametric Simple Linear Model (PSLM),of LTF-type, using recent rainfall and observed discharge data; (vi) theParametric Linear perturbation Model (PLPM), also of LTF-type, using recentrainfall and observed discharge data, (vii) n-AR, an AR model applied to theobserved discharge series only, as a na?ve updating model; and (viii)n-NARXM, a naive form of the NARXM, using only the observed discharge data,excluding exogenous inputs. The five GFMFS simulation (non-updating) modelsused were the non-parametric and parametric forms of the Simple Linear Modeland of the Linear Perturbation Model, the Linearly-Varying Gain Factor Model,the Artificial Neural Network Model, and the conceptual Soil MoistureAccounting and Routing (SMAR) model. As the SMAR model performance was foundto be the best among these models, in terms of the Nash-Sutcliffe R2 value,both in calibration and in verification, the simulated outflows of this modelonly were selected for the subsequent exercise of producing updated dischargeforecasts. All the eight forms of updating models for producing lead-timedischarge forecasts were found to be capable of producing relatively goodlead-1 (1-day ahead) forecasts, with R2 values almost 90% or above. However,for higher lead time forecasts, only three updating models, viz., NARXM, LTF,and NNU, were found to be suitable, with lead-6 values of R2 about 90% orhigher. Graphical comparisons were made of the lead-time forecasts for thetwo largest floods, one in the calibration period and the other in theverification period.