Testing of new stormwater pollution build-up algorithms informed by a genetic programming approach

Pollution build-up and wash-off processes are often included in urban stormwater quality models. However, these models are often unreliable and have poor performance at large scales and in complicated catchments. This study tried to improve stormwater quality models by adopting the genetic programming (GP) approach to generate new build-up algorithms for three different pollutants (total suspend solids - TSS, total phosphorus - TP and total nitrogen - TN). This was followed by testing of the new models (also traditional build-up and wash-off models as benchmark) using data collected from different catchments in Australia and the USA, The GP approach informed new sets of build-up algorithms with the inclusion of not just the typical antecedent dry weather period (ADWP), but also other less 'traditional' variables - previous rainfall depth for TSS and maximum air temperatures for TP and TN simulation. The traditional models had relatively poor performance (Nash-Sutcliffe coefficient, E < 0.0), except for TP at Gilby Road (GR) (E = 0.21 in calibration and 0.43 in validation). Improved performance was observed using the models with new build-up algorithms informed by GP. Taking TP at GR for example, the best performing model had E of 0.46 in calibration and 0.54 in validation. The best performing models for TSS, TP, and TN are often different, suggesting that specific models shall be used for different pollutants. Insights into further improvements possible for stormwater quality models were given. It is recommended that in addition to the typical build-up and wash-off process, new generations of stormwater quality models should be able to account for the non-conventional pollutant sources (e.g. cross-connections, septic tank leakage, illegal discharges) through stochastic approaches. Emission inventories with information like intensity-frequency-duration (IFD) of pollutant loads from each type of non-conventional source are suggested to be built for stochastic modelling.