Predictive heuristic control: Inferring risks from heterogeneous nowcast accuracy

Urban areas are expanding globally, concentrating waste streams and negatively impacting the natural environment. Overloading of existing urban drainage systems (UDS) through densification of existing areas, urbanisation of surrounding areas and climate change can increasingly lead to combined sewer overflows (CSOs) and urban flooding (Mahaut & Andrieu 2019), which negatively impact the receiving water bodies and public health (Ten Veldhuis et al. 2010; Botturi et al. 2020; Owolabi et al. 2022). To mitigate this problem, (blue-green) infrastructure expansion is often used as a solution but can be too expensive or impossible to implement due to space constraints. Real-time control (RTC) aims to use real-time information about the system-state to optimally operate the UDS (Schütze et al. 2002) and thereby reduce the impact on the receiving water bodies (Langeveld et al. 2013) without the need for large-scale infrastructure investments. Implementation of these operative optimisation practices is becoming increasingly widespread, aided by their inclusion in the new Urban Wastewater Treatment Directive of the European Union (European Commission 2022).

The way in which the set-points (the target setting of each controlled actuator) of the actuators are determined (the procedure as per the nomenclature set by Schütze et al. (2002)) can be classified as either heuristic or real-time optimisation control (García et al. 2015). Although there is some theoretical evidence that real-time optimisation can outperform heuristic control (Lund et al. 2018), the additional efforts required for this type of control can outweigh the operational performance of the UDS (Beeneken et al. 2013; Mollerup et al. 2017). Consequentially, despite the numerous publications on real-time optimisation methods, heuristic control strategies remain the most ubiquitously applied version of RTC for UDS, as published examples on implemented real-time optimisation procedures, in practice, remain scarce (van der Werf et al. 2022). The lack of implemented examples and therefore uncertainty about the practical performance of predictive-based control procedures remains one of the main barriers against real-time optimisation control implementation (Naughton et al. 2021). Integrating forecasts within existing heuristic control procedures, under the principle of nowcast-informed RTC, could provide the missing link needed towards the widespread implementation of more advanced forms of RTC.

The influence of nowcast uncertainty has been shown to be negligible in real-time optimisation procedures (Fiorelli et al. 2013; van der Werf et al. 2023). This is attributed to the continuous updating of the model (when using model-predictive control), forecasts and a discount factor. Rule-based RTC (RB-RTC) strategies, as applied in practice, typically function with pre-defined set-points and thresholds embedded in suitable if-then operation rules (García et al. 2015). These RB-RTC methods can be augmented with prediction, showing an increase in the efficacy of RB-RTC potential. Before expanding this form of control with nowcast or long-term forecasts, the prediction accuracy and potential associated risks should be considered. In previous work, the efficacy of this nowcast-informed RB-RTC strategy has been shown to work well for single detention ponds or storm drains (Table 1). The application of nowcast-informed RB-RTC for CSO reduction in combined sewer systems, however, has not been studied to the same extent compared to detention ponds and storm drains.

Importantly, nowcast accuracy (the ability to correctly predict the rainfall over a given horizon) is not homogeneous and depends on the rainfall characteristics (Lin et al. 2005; Fabry & Seed 2009; Imhoff et al. 2020). The effects of these heterogeneities in nowcast accuracy on UDS modelling and consequently the RTC efficacy have not been explicitly considered in existing forecast-induced uncertainty studies (Achleitner et al. 2009; Schellart et al. 2014; Löwe et al. 2016) where uncertainty is commonly modelled using the homogeneous Gaussian noise (Pleau et al. 2001; Svensen et al. 2021). This can lead to inaccurate conclusions being drawn on the influence of nowcast uncertainty on the performance of RTC strategies and should therefore be avoided.

Therefore, this paper aims to develop efficient, nowcast-informed RB-RTC strategies and to assess if there are risks associated with heterogeneous nowcast accuracy. Furthermore, it aims to assess if these risks can be directly inferred from nowcast and rainfall data.

Firstly, available nowcast and observed rainfall datasets have to be obtained. Here, the observed rainfall dataset is a rain-gauge-adjusted radar product, following a data assimilation procedure detailed by Overeem et al. (2009). The nowcast data is an open-source product obtained using an advection model based on real-time observation of two C-bands operated by the Royal Netherlands Meteorological Institute (KNMI, see dataplatform.knmi.nl). The rainfall datasets obtained this way then have to be discretised into individual rainfall events to allow for an event-based risk assessment (using the methodology set out in Section 2.3). Here, the definition of a rainfall event uses a minimum inter-event time (MIT) of 12 h based on the rain-gauge-adjusted radar product, recommended as this is the approximate emptying time of the system (Joo et al. 2013).

Before the RB-RTC procedure through the inclusion of nowcast is considered, an analysis of the potential of RTC-based operational improvement has to be done, relative to existing UDS operation without nowcast. Using the previously established rainfall events, the model-based performance evaluation methodology presented by van Daal et al. (2017) is followed to determine the current functioning of the RTC. For the same events, the adjusted absolute central basin approach (aCBA), previously used in the computation of performance indicators (van der Werf et al. 2021), is used to compute the maximum performance potential of the UDS through RTC, while the former indicates the current performance of the UDS. The adjusted aCBA computes the maximum achievable performance of the current UDS configuration by the iteratively lumping pumped section into a single basin with the aim to compute the minimum CSO volume theoretically possible (see Supplementary Material section 1 (Extended Methodology and Figure S1) for a more detailed breakdown of this previously developed method).

Rainfall events are then ranked based on the difference between the predicted performance and the maximum achievable performance. Events with the largest difference are analysed by first identifying the location and timing of the CSO events. Then, the settings of the actuators influencing the CSO location are analysed before, during and after the CSO event. Potential improvements are noted by using the expert judgement of the dynamics of the system and by assessing the filling degrees in the UDS up- and downstream of an activated in combination with actuator settings to see if local improvements could have been realised.

Based on this performance improvement potential, updated rules expanding on the existing heuristic RTC procedure are developed. Here, the existing RB-RTC is updated into a nowcast-informed RB-RTC through two types of rule augmentations: (1) nowcast-based bifurcations at the existing rules (where if-then logic is added to the current if-then rule) and (2) replacement of the existing rules by outputs of a nowcast-forced model. To assess the effect of potential nowcast accuracy heterogeneity, the nowcast-informed RB-RTC should include augmentations at different phases (beginning, during and at the end of the rainfall event) and based on different rainfall properties (total depth or intensity).

The implementation of the control rules can either be (1) nowcast-based heuristic control or (2) model-informed heuristic control. Nowcast-based heuristic control directly uses the nowcast data within the RB-RTC procedure (e.g. if predicted rainfall depth <0.5 mm, then pump-off). Conversely, the model-informed heuristic control uses the nowcast as rainfall forcing in the (simplified) model of the UDS, making the RB-RTC procedure dependent on the filling degree predicted through the UDS model. In the model-informed heuristic control procedure, the monitoring data available is used to generate the initial conditions of the model, and the available nowcast data is thereafter used as the forcing of the model. Every set of updated nowcast-informed rules is implemented and simulated separately (see Sections 2.2.1 and 2.2.3 for the assessment methodologies used).

The rules for perfect nowcast data (using the observed rainfall) are developed following a heuristic approach, relying on UDS understanding rather than formal optimisation to develop and calibrate the rules. This form of rule generation was deemed sufficient as long as the operational performance with respect to the control objective of the nowcast-informed RB-RTC using perfect prediction was higher compared to the baseline RB-RTC.

Two types of risks associated with various forms of uncertainties in a real-time optimisation context were previously identified: (1) risk of UDS performance loss and (2) risk of operational deterioration of UDS. The former indicates a decreased improvement compared to the perfect prediction scenario, but a net increase in UDS operational performance compared to a baseline, and the latter means a loss of performance compared to the baseline, induced by operational uncertainties.

To assess these potential risks, two methods are compared: (1) assessing the accuracy of the nowcast dependent on features (related to, for example, the depth or stage of the rainfall) used in the RB-RTC procedure and (2) comparing the model-informed performance using real nowcast data as the input of the RB-RTC procedure. The first method does not require UDS nor RTC modelling and is, therefore, the preferred method. The validity of this method, however, can only be established if the risks are first calculated through a model-based risk assessment, the methodology which is further expanded in Section 2.2.2. Therefore, the two methods are compared here to assess if the same conclusions can be drawn in which case the forecast-based risk assessment would be preferred; otherwise, the model-based risk assessment has to be used.

To assess if there is heterogeneity in the nowcast accuracy, several key properties are considered, keeping the nature of the use of the nowcast within the RB-RTC procedure in mind. The key properties assessed for heterogeneity in accuracy are (1) accuracies dependent on the stage of the rainfall event (start, middle or end), (2) accuracy dependent on forecast horizon, (3) forecast consistency and its relation to accuracy and (4) accuracy dependent on total rainfall depth. These key properties are further explained below.

As the prediction horizon used in this work (2 h) is significantly smaller than the MIT (12 h), sub-events are discretised from the larger events, with the sub-MIT set to 3 h (thereby keeping no-rain time periods within the assessed data).

The risks associated with the inclusion of different nowcast attributes cannot be directly inferred from the statistical methods described previously, given the non-linear nature of RTC performance to rainfall characteristics (Vezzaro 2021). For this reason, a model-based risk assessment method is needed to quantify the risk associated with the nowcast accuracy properties assessed.

Two previously defined risks, the risks of performance loss and operational deterioration, related to the various analysed aspects of the nowcast rainfall data, need to be computed. To achieve this, the updated procedures shown in Section 2.1 are implemented in the EPA StormWater Management Model (SWMM) 5.1 software (Rossman 2015) through the PySWMM python interface (McDonnell et al. 2020). To gain a better understanding of the performance and its dynamics, reading of the SWMM output files is done through the python SWMM API (Pichler 2022). Both full hydrodynamic (FH) and simplified models can be used for the model-based risk assessment, providing adequate model calibration has been conducted. Here, a simplified model of the case study UDS was used, given the too-inhibitive computational inefficiency associated with FH models.

With the new rules implemented, -rainfall events representative of the spread of rainfall characteristics typical for the used case study are simulated using the aforementioned SWMM model. The number of rainfall events used depends on the available data and computational efficiency of the used SWMM model, noting that it is recommended to use multiple years of data where possible (van Daal et al. 2017). Each simulation is done using the observed rainfall as the nowcast data (perfect forecast) and another using real nowcast data available in real time through the KNMI for RTC purposes. The perfect and real forecast outputs are then compared to the established baseline (the RB-RTC procedure set-up prior to the inclusion of nowcast data). The aforementioned risks of performance loss and operative deterioration associated with each set of nowcast-informed RB-RTC procedures are then calculated by comparing the performances in two simulations with the baseline RB-RTC performance.

The a priori implications determined through the nowcast-based risk assessment (Section 2.1) are compared to the model-based risk assessment method. As the risk of operational deterioration and performance loss are directly quantified through the model-based risk assessment and the nowcast-based risk assessment has to infer these based on various nowcast metrics, the comparison between the two methods cannot be directly quantitative. The operational risks (both operational deterioration and performance loss) related to each set of nowcast-informed rules, determined through the model-based risk assessment, are therefore ranked first (from highest to lowest risk). This ranking is then compared to the relative performance of the nowcast property associated with the nowcast-informed rules to see if a high risk is associated with a relatively poor performance in nowcast accuracy.

The set-points for each of the pumping stations are derived in a central computer after receiving the relevant information from the water level and flow sensors from the UDS and the WWTP. These set-points are then sent to the programmable logic controllers at each pumping station, which use the new set-point as a target achieved following a local control algorithm. This allows for a distributed implementation of the central control, overcoming difficulties with complicated, unreadable codes.

The control strategy was set up as a multi-objective problem, aiming to reduce the frequency and volume of CSO discharges to the urban canals first, then the total CSO volume and finally the reduction of overloading the WWTP. As no formal real-time optimisation algorithm was used, a hierarchical form of control objective, namely the reduction of CSO volumes discharged to the urban canals, was used as opposed to a weighted CSO volume synthesis (as proposed by Vezzaro & Grum (2014)). When reporting the results, the total CSO volume is used as means to indicate the improvement of the system performance, unless an increase in the CSO volume towards the urban canals has been caused.

Due to physical constraints of the pumping station configurations, the minimum sampling interval of the pumps (the time between adjusting the pumping station set-points) is 15 min. This sampling interval is, therefore, used within the model-based risk assessment.

The shown schematisation of the UDS (Figure 4) is used to build the simplified simulation model, with each sewer district modelled as a reservoir and the storage curve and hydrological data (for each sewer district) based on a previously calibrated model used for the development of the CAS2.0 RB-RTC procedure (Langeveld et al. 2022). The storage curve is derived from the full hydrodynamic model of the sewer districts. CSOs in each district are implemented and the depth of the flow curve is derived from full hydrodynamic model data. The existing RTC rules are augmented using six different sets of new rules (i.e. new RTC procedures), utilising nowcast data in different ways (Table 2). Given the presence of the CSO pumps, some of the nowcast-informed rules apply only to these actuators (RTC procedure 1). The other RTC procedures make the operation of all pumping stations in the original RB-RTC using different total depths for activation (RTC Procedures 3 and 5), event phases (RTC Procedures 1 and 2), spatially aggregated data (RTC Procedure 6) or fully relying on the nowcast and current data within the established RB-RTC (Procedure 4). More details including examples of the developed procedures can be found in the Supplementary Materials (Section 2 and Figures S2–S7).

The potential for performance improvement compared to the baseline RB-RTC was assessed for the 95 events previously described. Based on the results obtained using the aCBA method, a total CSO volume reduction of 37.2% was possible through optimal operation of the UDS when compared to the baseline. For rainfall events with a smaller total rainfall depth, complete preventions of CSOs have been achieved. For larger rainfall events, the maximum CSO volume reduction was 5.8%. This is consistent with previous results, showing the highest potential of RTC impact in the events with low- to mid-rainfall depths (van der Werf et al. 2022).

The predictive ability to correctly predict larger rainfall events is significantly lower compared to smaller rainfall depths (Figure 7). The POD value drops from 0.93 for predictive ability above 0.25 mm to around 0.5 when the threshold is set to 1 mm of cumulative predicted rainfall. This drops further to a POD value of 0.1 for any threshold above 5 mm. The SPC, on the other hand, increases during the same interval, ending in a maximum value of 0.98 for the larger rainfall events. The abundance of data points is significantly higher for the lower fractions partially due to the discretisation method of the rainfall events (using an MIT of 12 h).

These results suggest a negative bias within the nowcast dataset, as an underestimation of the real rainfall would see a rise in SPC and a drop in POD with an increasing threshold used. This is in line with previous research showing a decreased ability to predict large (often convective) rainfall events (e.g. Foresti et al. 2016; Pulkkinen et al. 2020). This has also been shown to lead to decreased discharge forecast abilities (Imhoff et al. 2022) for hydrological models, conclusions which can largely be extrapolated to the UDS.

The RMSE in the cumulative rainfall prediction (see Figure 6(b)), however, increases exponentially when considering the individual pixel predictions (i.e. the input to any UDS model used in a model-predictive sense). This is in line with the previous exponential decrease in Pearson's correlation as reported by Imhoff et al. (2020).

A stronger link between the mean correlation coefficient and the depth of the rainfall event was found. The consistency of the nowcast for 0–0.5 mm of rainfall in the forecasted hour was significantly higher (⁠⁠) compared to other intervals. This relationship explains the difference between the normalised RMSE and the total RMSE shown in Figure 9, as the higher end of the mean correlation coefficient is dominated by low rainfall data (thus having a relatively low RMSE but no difference in normalised RMSE).

From a control perspective, the time-based discretisation method gives a more accurate representation of the end of the event, as the volume-based method forces the inclusion of rainfall data. The implementation of end-of-event logic (i.e. switching off CSO pumps when no more rain is expected) will follow the time-based implementation, meaning that this form of control is unlikely to be strongly affected by nowcast-induced errors.

However, when the real nowcast data is used, a considerable risk of both performance loss and operational deterioration is noted for all updated RB-RTC procedures (Figure 11). Only two RTC procedures (Procedures 1 and 2) reduced the total CSO volume of the operation under the real nowcast scenarios but significantly less compared to the theoretical potential (Table 3). Procedures 4 and 6 show a minor increase in the CSO reduction compared to the baseline (both 0.8%). However, most notably, Procedures 3 and 5 show a net increase in the total CSO volume of 17.7 and 15.7%, respectively. In the case of Procedure 3, this increase in volume is predominantly caused by the pumped CSOs being activated unnecessarily, as shown in Figure 11 as the column on the low part of the baseline performance. As a result, 40.8% of the events within the dataset had a worse performance compared to the baseline RB-RTC. Procedure 5 also resulted in a number of events with deteriorated UDS operation (25.0%), but the relative CSO increase was caused less by the unnecessary CSO pumping and more by a decreased performance throughout all the rainfall events.

RTC of UDS aims to utilise the existing UDS optimally by ensuring the optimal distribution of runoff throughout the system. Consequentially, if there is a failure throughout the system (i.e. nowcast inaccuracies), the possibility of adverse effects increases. The two control systems which had the highest performance decrease (Procedures 3 and 5) were both designed to (1) use the numerical output of the nowcast algorithms directly and (2) operate closer to the total filling of the UDS. No relation between the apparent risk and the RTC potential (i.e. the performance using the perfect data) was found.

Finally, in the shown case study, it was observed that the augmentation of the RB-RTC with nowcast data leads to a significant decrease in operational performance compared to both the perfect forecast case and the baseline RB-RTC. This is despite the potential of nowcast to decrease the total CSO volume for all analysed procedures.

Of the six RTC procedures presented here, Procedure 1 (relying on the correct prediction at the end of the event to switch off the pumped CSOs) was able to reduce the CSO volume in both the perfect and real nowcast cases. This is consistent with the ability of the nowcast model to predict the end of a rainfall event (as previously shown in Figure 10). Although this accuracy is not perfect (mean SPC of 0.81), the effect of incorrect predictions appears to have a small impact on the UDS operational performance (3% CSO volume increase compared to the perfect baseline).

Conversely, a higher risk of operational deterioration was observed when the RB-RTC procedure sets the CSO pump settings directly dependent on the forcing of the forecasted rainfall depth in a modelled setting. This was further exacerbated by switching on the pumped CSOs if a higher threshold was predicted (using a FD of 0.9 as opposed to 0.8). As these procedures used the entire forecast horizon to predict the depth (with an activation only occurring at the higher end of the rainfall prediction), both rely on two of the weakest points identified in the nowcast data. These are using the direct numerical prediction over the full horizon (120 min) and using a high threshold for the activation of the rules. Procedure 5 relied on a consistent forecast in order to activate, though, as also shown through the nowcast performance assessment, this did not increase the robustness of the control procedure.

The results from the nowcast accuracy assessment yielded similar trends compared to the model-based risk assessment methodology. Although it is not possible to quantify the risks directly from the nowcast data, due to aforementioned non-linearities and the nature of the risks identified here, the statistical analysis of nowcast data gives a strong indication on how the nowcast can be included in future RB-RTC strategies. This, in turn, allows bypassing the extensive model-based analyses, as presented here, and allows operators to make informed decisions based on less data and model outputs.

The maximum potential CSO reduction reported here (14.6% for Procedure 4) is smaller when compared to the corresponding reduction reported by other studies shown in Table 1. However, those studies focussed on the evaluation of either a single detention pond or a stormwater system, i.e. drainage (sub-)systems with a relatively high level of controllability (in terms of the level of influence of actuators on the behaviour of the system). As controllability is one of the key factors influencing RTC efficacy potential (van der Werf et al. 2022), it is expected that the optimised functioning improved the system more in the previous studies compared to those reported here. A lack of information on the full RTC potential of the previously reported systems, furthermore, makes the direct comparison impossible.

When comparing the performance loss related to the nowcast accuracy with real-time RTC optimisation procedures, the self-adjusting mechanism reported to reduce the impact (Fiorelli et al. 2013; Courdent et al. 2015) was not observed in the model-based risk assessment results. This is likely to be a result of the relatively long sampling interval and the design of the updated RB-RTC procedures. However, it is possible that an increased level of risk could be an inherent drawback of RB-RTC including nowcast compared to real-time optimisation procedures. Additional evidence, based on different case studies following the presented methodology here, is necessary to ascertain the latter and attribute and quantify the relative importance of the former two.

Furthermore, studies which investigate the impact of nowcast uncertainty through synthetic errors (e.g. Svensen et al. 2021) are likely to misreport the potential impacts of nowcast uncertainties, given the dynamics observed here. The use of real nowcast algorithms in combination with real observed data is therefore considered to be a must for applicable results.

Although the methodology presented here applies to other case studies, the main findings depend on the available nowcast data (both events within the dataset and the forecast model used to generate the nowcast data) and hence the results obtained in this work may, therefore, not be directly transferable (He et al. 2013; Imhoff et al. 2020). The nowcast data accuracy assessment, as presented here, should, therefore, be done prior to the implementation of nowcast-included RB-RTC in other case studies to ensure similar dynamics are observed.

With the rise of machine learning (ML) applications in the field of rainfall nowcasting (Ravuri et al. 2021; Amini et al. 2022), new nowcasting methods which focus on specific rainfall properties (i.e. timing or magnitude of the start of an event, total sum of the horizon or prediction of the ending of the event) could improve the functionality of RB-RTC augmented with nowcast. Similarly, the use of ensemble forecasts has previously been shown to be beneficial in an RTC context (He et al. 2013; Courdent et al. 2015; Balla et al. 2020). Understanding the differences within the ensemble forecast can further decrease potential uncertainties and their effect on the operation of UDS.

It should be noted that the updated RB-RTC procedures were not formally optimised nor designed to be robust against nowcast uncertainties, as the aim of this research was to investigate the effects of nowcast uncertainties on RB-RTC procedures. Furthermore, the prediction of the end-of-the-event was particularly useful in the used case study due to the presence of CSO pumps, enabling the direct switching off of these pumps when no rainfall was expected. The applicability of the end-of-the-event prediction to other case studies might be more limited.

When implementing nowcast-informed RB-RTC, a detailed trade-off between the added benefits of a longer horizon and the induced uncertainties should be done. Increased horizons have been shown to plateau in their potential for system operative improvement (Jafari et al. 2020) and therefore inclusion of the entire horizon, as was done in some of the procedures developed here, might lead to sub-optimal operation.

This research aims to understand how heterogeneity in nowcast accuracy, dependent on varying rainfall properties, can affect and be used in a RB-RTC procedure for the improvement of UDS operation. Using both perfect and real nowcast rainfall data in the Rotterdam UDS, a model-based study was performed to assess the risk of six different RB-RTC procedures potentially resulting in performance loss or operational deterioration of the analysed UDS. Based on the results obtained, the following conclusions could be drawn:

• Multiple forms of nowcast-informed RB-RTC can significantly decrease the pollution load from UDS when using perfect nowcast data. However, significant risks are associated with using real nowcast data to control the UDS resulting, in some cases, in significantly increasing the CSO frequency and volume.

• Nowcast accuracy is heterogeneous and depends on the stage and depth of the rainfall event. The nowcast data indicated a strong predictability of the end of a rainfall event but poor performance at the start and in the middle of an event, particularly when larger rainfall depths were observed.

• This heterogeneity relates directly to risks of operative deterioration and performance loss, as the impact of the uncertainty can be inferred directly from the nowcast data. Using these inferred risks, decisions can be made about the development of nowcast-informed RB-RTC.

• The dynamics related to heuristic control versus real-time optimisation control in the context of uncertain forecasts differ, as the heuristic control does not appear to have a strong self-correcting mechanism. A real-time optimisation strategy might, therefore, be a more robust implementation of nowcast-informed forms of control.

Based on these conclusions, additional research is needed to understand if the results and conclusions drawn from this case study are transferable to other UDSs. Furthermore, the ability to formally optimise the nowcast-informed RB-RTC using real nowcast data was not assessed here and a comparative assessment of a nowcast-informed RB-RTC calibrated on real and perfect data should be done. Additional measures to ensure the resilience of heuristic control to uncertain rainfall data should be investigated.

The authors would like to thank the Royal Dutch Meteorological Institute (KNMI) for providing the nowcast data used in this paper. Furthermore, the authors also thank Imber Advies for preparing the models used for the design of the control rules and used in the assessment of the framework and the City of Rotterdam for providing the necessary data. This research was carried out through funding of the Kennisprogramma Urban Drainage, which is sponsored through ARCADIS, Deltares, Evides, Gemeente Almere, Gemeente Arnhem, Gemeente Breda, Gemeente ‘s-Gravehage, Gemeentewerken Rotterdam, Gemeente Utrecht, GMB Rioleringstechniek, KWR Watercycle Research Institute, Royal HaskoningDHV, Stichting Rioned, STOWA, Sweco, Tauw, Vandervalk&deGroot, Waterboard de Dommel, Waternet and Witteveen&Bos. We would like to thank all the partners for their continued support.

J.A.v.d.W. carried out the analysis of the nowcast and schematised the inclusion of forecast in the control strategy. J.G.L. designed the original control strategy and acquired funding for the research. J.A.v.d.W. wrote the initial draft of the manuscript. All the research was carried out under the active supervision of Z.K. and J.G.L. All authors co-edited the final manuscript.

Data cannot be made publicly available; readers should contact the corresponding author for details.

The authors declare there is no conflict.