Using ensemble weather forecast in a risk based real time optimization of urban drainage systems

Abstract

ABSTRACT. -Global Real Time Control (RTC) of urban drainage system is increasingly seen as cost-effective solution in order to respond to increasing performance demand (e.g. reduction of Combined Sewer Overflow, protection of sensitive areas as bathing water etc.). The Dynamic Overflow Risk Assessment (DORA) strategy was developed to operate Urban Drainage Systems (UDS) in order to minimize the expected overflow risk by considering the water volume presently stored in the drainage network, the expected runoff volume based on a 2-hours radar forecast model and an estimated uncertainty of the runoff forecast. However, such temporal horizon (1-2 hours) is relatively short when used for the operation of large storage facilities, which may require a few days to be emptied. This limits the performance of the optimization and control in reducing combined sewer overflow and in preparing for possible flooding. Based on DORA's approach, this study investigated the implementation of long forecast horizon using an ensemble forecast from a Numerical Weather Prediction (NWP) model. The uncertainty of the prediction is characterized by an ensemble of 25 forecast scenarios. According to the status of the UDS and the forecasted runoff volumes, the objectives for the control strategies might vary from optimization of water volumes to reduction of CSO risk. Thus different modes are implemented in DORA-LF (Long Forecast) in order to adjust the control strategies to the situations. In order to handle the long forecast, the horizon is divided into multiple and variable time step. This new approach was tested on selected rain events and shows an improvement in the protection of sensitive areas during long or/and coupled events by allowing anticipated CSO in low sensitivity areas.

Model predictive control of urban drainage systems: A review and perspective towards smart real-time water management

“…This means that only one control action value is calculated for the entire control horizon instead of a time-varying trajectory of control actions. Courdent et al (2015) used a varying setting duration that became coarser with time. The setting duration is often not reported explicitly in the reviewed literature, probably because the authors use the modeling time step (which is often not given either) or the sampling interval as the setting duration, but omit this information.…”

“…These lengths coincide with the forecast ability of many radar nowcast products (Thorndahl et al, 2017), which is probably why longer forecasts are normally not applied. Courdent et al (2015) used rain from a NWP model as input to the internal MPC model, allowing for a collective horizon of 13 hr. Fiorelli et al (2013), Pleau et al (2005), and Vezzaro and Grum (2014) are the only reviewed papers to report that MPC is either in or ready for operation; they used sampling intervals of 10, 5, and 2 min, respectively, setting durations of 10, 5, and 120 min, and all applied a collective horizon of 2 hr.…”

“…-Minimization of the monetary cost related to a CSO event. This can also include the uncertainty related to the forecasted input and thus the risk associated with a CSO event, which was applied in the Dynamic Overflow Risk Assessment (DORA) method (Courdent et al, 2015;L€ owe et al, 2016;Meneses et al, 2018;Grum, 2012, 2014;Vezzaro et al, 2013Vezzaro et al, , 2014aVezzaro et al, , 2014b). -Preservation of system behaviors that are not included in the internal MPC model, such as ensuring that overflows can only occur from filled basins and that the set-points of local controllers are respected (Pleau et al, 1996) or including transport time in pipes (Fiorelli and Schutz, 2009;Fiorelli et al, 2013).…”

“…Penalties are often chosen through a trial-and-error process and it is generally observed that flooding and CSO volumes are given the highest penalties. It is possible to give the same type of objective the same penalty value, but the penalties can also be used to distinguish between, for example, different WWTPs or different CSOs with regard to receiving water sensitivity (Courdent et al, 2015;Duchesne et al, 2004;Fiorelli and Schutz, 2009;Fiorelli et al, 2013;Joseph-Duran et al, 2014b;L€ owe et al, 2016;Patry, 1983;Pleau et al, 1996Pleau et al, , 2005Vezzaro and Grum, 2014).…”

“…The penalties can also be changed when something happens that changes the cost of deviating from the reference value for the remainder of an event; for example, at a specific CSO location after an overflow has already occurred or when the first flush volume has passed through the WWTP (Patry, 1983;Vezzaro et al, 2014a). Courdent et al (2015) not only changed the penalty values, but also constructed different "modes" with individual objective functions, which were applied based on the system state and the forecasted rain. Pleau et al (2005) also used input-dependent modes to shift between passive control (for dry weather periods) and MPC (for wet-weather conditions).…”