Learning-Based Predictive Control with Gaussian Processes: An Application to Urban Drainage Networks

Many traditional control solutions in urban drainage networks suffer from unmodelled nonlinear effects such as rain and wastewater infiltrating the system. These effects are challenging and often too complex to capture through physical modelling without using a high number of flow sensors. In this article, we use level sensors and design a stochastic model predictive controller by combining nominal dynamics (hydraulics) with unknown nonlinearities (hydrology) modelled as Gaussian processes. The Gaussian process model provides residual uncertainties trained via the level measurements and captures the effect of the hydrologic load and the transport dynamics in the network. To show the practical effectiveness of the approach, we present the improvement of the closed-loop control performance on an experimental laboratory setup using real rain and wastewater flow data.