Prediction of event-based stormwater runoff quantity and quality by ANNs developed using PMI-based input selection

Event-based stormwater runoff quantity and quality modeling remains a challenge since the processes of rainfall induced pollutant discharge are not completely understood. The complexity of physically-based models often limits the practical use of quality models in practice. Artificial neural networks (ANNs) are a data driven modeling approach that can avoid the necessity of fully understanding complex physical processes. In this paper, feed-forward multi-layer perceptron (MLP) network, a popular type of ANN, was applied to predict stormwater runoff quantity and quality including turbidity, specific conductance, water temperature, pH, and dissolved oxygen (DO) in storm events. A recently proposed input selection algorithm based on partial mutual information (PMI), which identifies input variables in a stepwise manner, was employed to select input variable sets for the development of ANNs. The ANNs developed via this approach could produce satisfactory prediction of event-based stormwater runoff quantity and quality. In particular, this approach demonstrated a superior performance over the approach involving ANNs fed by inputs selected using partial correlation and all potential inputs in flow modeling. This result suggests the applicability of PMI in developing ANN models. In addition, the ANN for flow outperformed conventional multiple linear regression (MLR) and multiple nonlinear regression (MNLR) models. For an ANN development of turbidity (multiplied by flow rate) and specific conductance, significant improvement was achieved by including a previous 3-week total rainfall amount into their input variable sets. This antecedent rainfall variable is considered a factor in the availability of land surface pollutants for wash-off. A sensitivity analysis demonstrated the potential role of this rainfall variable on modeling particulate solids and dissolved matters in stormwater runoff.