Multi-step Flow Routing Using Artificial Neural Networks for Decision Support

Abstract

The use of surrogate and data-driven models has the potential to decrease the computational effort of streamflow predictions in time-critical model applications such as Data Assimilation (DA) or Model Predictive Control (MPC). In the present work, it was evaluated the use of Artificial Neural Network (ANN) as replacement of a physical model in a MPC implementation for the multi-objective optimization of a reservoir system. The presented application covers the flow routing of a reservoir release from Tres Marias dam in Brazil and downstream tributaries to Pirapora gauge for lead times between 1 and 360 h (15 days). The ANNs were trained using Levenberg–Marquadt algorithm, and three different transfer functions were evaluated. It was also tested two output correction techniques, namely an ARX model for error prediction and bias correction for lead time until 15 days ahead. The ANNs model shows good capability of handling with minor disturbances. Best performance was found using the purelin transfer function. For the error correction, the ARX models showed better error reduction when compared with the bias correction technique, which could not reduce the error at the end of lead time.