Real-time flow forecasting in the absence of quantitative precipitation forecasts: A multi-model approach

Abstract

Summary In situations where Quantitative Precipitation Forecasts (QPF) over lead-times are not readily available, but lead-time flow forecasts are required, models based either solely on the recorded flow data series, or on the recorded series of flow and past rainfall are needed. For the first input scenario considered, involving the flow series only, five lumped models which are based on the simple autoregressive process and artificial neural networks are used. For the second input scenario, involving two recorded series, i.e. rainfall and runoff, two linear transfer function type models and two models based on artificial neural networks are employed. Using a multi-model approach, the outputs of the models in each scenario are combined using three techniques of combination, namely, the simple and the weighted average methods and the neural network method. The Nash–Sutcliffe model efficiency index, the Root Mean Square Error and the Mean Relative Error are used to assess the relative performance of the models and combinations. Daily recorded flow and rainfall data of two rivers, one in France and the other in Ireland, are used to generate flow forecasts up to six day lead-times. For the French catchment, it is observed that, among the individual forecasting models utilising only the observed flows, the best forecast efficiency in verification for lead-times of one and two days is achieved by the simple autoregressive model whereas for the higher lead-times, the neural network model performs best. In contrast, among the models employing both observed flow and rainfall, the linear transfer function model utilising departures from the seasonal means produces the highest efficiency. For the Irish catchment, the neural network models employing the departures from the seasonal means are consistently best in verification for both scenarios, namely, using observed flow only and using both the observed flow and rainfall series. For both scenarios, the performance is better when both rainfall and flow data are used rather than the flow data only. In both scenarios of application, all three methods of model output combination produce very similar efficiency values which are generally better than the efficiencies of the individual models used in the combinations.