Accelerating hydrodynamic simulations of urban drainage systems with physics-guided machine learning.

We propose and demonstrate a new approach for fast and accurate surrogate modelling of urban drainage system hydraulics based on physics-guided machine learning. The surrogates are trained against a limited set of simulation results from a hydrodynamic (HiFi) model. Our approach reduces simulation times by one to two orders of magnitude compared to a HiFi model. It is thus slower than e.g. conceptual hydrological models, but it enables simulations of water levels, flows and surcharges in all nodes and links of a drainage network and thus largely preserves the level of detail provided by HiFi models. Comparing time series simulated by the surrogate and the HiFi model, R2 values in the order of 0.9 are achieved. Surrogate training times are currently in the order of one hour. However, they can likely be reduced through the application of transfer learning and graph neural networks. Our surrogate approach will be useful for interactive workshops in initial design phases of urban drainage systems, as well as for real time applications. In addition, our model formulation is generic and future research should investigate its application for simulating other water systems.

Robust model for estimating pumping station characteristics and sewer flows from standard pumping station data.

Flow data represent crucial input for reliable diagnostics of sewer functions and identification of potential problems such as unwanted inflow and infiltration. Flow estimates from pumping stations, which are an integral part of most separate sewer systems, might help in this regard. A robust model and an associated optimization procedure is proposed for estimating inflow to a pumping station using only registered water levels in the pump sump and power consumption. The model was successfully tested on one month of data from a single upstream station. The model is suitable for identification of pump capacity and volume thresholds for switching the pump on and off. These are parameters which are required for flow estimation during periods with high inflows or during periods with flow conditions triggering pump switching on and off at frequencies close to the temporal resolution of monitored data. The model is, however, sensitive within the transition states between emptying and filling to observation errors in volume and on inflow/outflow variability.

Comparing the impact of time displaced and biased precipitation estimates for online updated urban runoff models.

When an online runoff model is updated from system measurements, the requirements of the precipitation input change. Using rain gauge data as precipitation input there will be a displacement between the time when the rain hits the gauge and the time where the rain hits the actual catchment, due to the time it takes for the rain cell to travel from the rain gauge to the catchment. Since this time displacement is not present for system measurements the data assimilation scheme might already have updated the model to include the impact from the particular rain cell when the rain data is forced upon the model, which therefore will end up including the same rain twice in the model run. This paper compares forecast accuracy of updated models when using time displaced rain input to that of rain input with constant biases. This is done using a simple time-area model and historic rain series that are either displaced in time or affected with a bias. The results show that for a 10 minute forecast, time displacements of 5 and 10 minutes compare to biases of 60 and 100%, respectively, independent of the catchments time of concentration.

Comparison of short-term rainfall forecasts for model-based flow prediction in urban drainage systems.

Forecast-based flow prediction in drainage systems can be used to implement real-time control of drainage systems. This study compares two different types of rainfall forecast - a radar rainfall extrapolation-based nowcast model and a numerical weather prediction model. The models are applied as input to an urban runoff model predicting the inlet flow to a waste water treatment plant. The modelled flows are auto-calibrated against real-time flow observations in order to certify the best possible forecast. Results show that it is possible to forecast flows with a lead time of 24 h. The best performance of the system is found using the radar nowcast for the short lead times and the weather model for larger lead times.