Improving ANN model performance in runoff forecasting by adding soil moisture input and using data preprocessing techniques

This study attempts to improve the accuracy of runoff forecasting from two aspects: one is the inclusion of soil moisture time series simulated from the GR4J conceptual rainfall-runoff model as (ANN) input; the other is preprocessing original data series by singular spectrum analysis (SSA). Three watersheds in China were selected as case studies and the ANN1 model only with runoff and rainfall as inputs without data preprocessing was used to be the benchmark. The ANN2 model with soil moisture as an additional input, the SSA-ANN1 and SSA-ANN2 models with the same inputs as ANN1 and ANN2 using data preprocessing were studied. It is revealed that the degree of improvement by SSA is more significant than by the inclusion of soil moisture. Among the four studied models, the SSA-ANN2 model performs the best.