Urban Storm Water Pollution Forecasting Using Recurrent Neural Networks 1

To reduce the impact of stormwater pollution on the environment, operators in charge of urban drainage systems and waste water treatment plants need real-time predictions of pollutant concentrations during rainfall events. In this paper, this problem is addressed using nonlinear recurrent neural networks, which are used to simulate the rainfall-runo transformation, as well as the production and transfer of solids in drainage catchments and sewer pipes. Prior knowledge provided by existing conceptual models has been used to design speci c neural architectures with small numbers of parameters. Training of these networks can be performed using the epochwise or on-line backsweep algorithms recently formalized by Pich e [9]. Experimental results demonstrate the e ciency of this approach for predicting suspended solid concentration at the outlet of urban drainage catchments.