Implementing a digital twin for optimized real-time control of Gothenburg's regional sewage system

The impacts of climate change, increased urbanization and urban growth, aging infrastructure, community values, etc., are affecting cities and their utilities in the water industry. Tools supporting optimal decision-making when planning and operating collection systems and water resource recovery facilities (WRRF) with respect to these impacts were highlighted in the International Water Association's White Paper on digital twins in the urban sector (Valverde-Pérez et al. 2021). Garrido-Baserba et al. (2020) describe the potential of using available instrumentation and online data for tackling the problems of aging infrastructure and the need to extend infrastructure service life at low cost. The concept of a digital twin (DT) is generalized as incorporating a model of an object, an evolving set of data related to the object and a means of dynamically updating or adjusting the model in accordance with the data (Wright & Davidson 2020).

Modelling of collection systems and wastewater treatment processes is well established and several nordic utilities, technology providers and universities banded together to further develop digital tools suitable for actual infrastructure operations and planning. The Future City Flow (FCF) consortium was created in 2015 to develop such enhanced decision support tools for collection systems, the consortium consists of six utilities, three technology providers and two universities/research institutes (Hagman et al. 2018). Partial financing was provided by Vinnova, the Swedish governmental agency for innovation systems. The FCF development work has resulted in three digital tools for collection systems: Data – for advanced examination of data and automatic report generation; RTC (real-time control) – for enhancing operation optimization in the immediate future using RTC and model-based predictive control of existing infrastructure; and Planning – for the long-term optimization of collection systems. The Planning tool uses simplified, surrogate catchment models that speed up computations to allow for long-term simulations (several years) while retaining the behaviour of the catchment. By easily being able to test different scenarios (e.g. increasing/decreasing population, increasing/decreasing infiltration/inflow, wet/dry year), various proposed future measures can be evaluated, and optimal solutions identified. This paper focuses on the FCF RTC optimization tool and presents a case study from Gryaab AB, one of the Swedish utilities in the FCF consortium.

Generally, a catchment model is formulated and verified for each subcatchment using a deterministic model concept found in MIKE+. The hydrological processes are described with a general hydrological model (MIKE+ RDI) that, in addition to runoff from impervious areas (FRC – fast response component), also considers infiltration into the sewer system from the surrounding soil (SRC – slow response component) as illustrated in Figure 1(c). The latter strongly depends on the hydrological history (i.e. preceding events), and surface, root zone and groundwater storages, as well as snow accumulation are taken into account. Each subcatchment model is a simplified conceptualization of the real subcatchment and is designed to reflect the salient characteristics of the subcatchment, such as sewage flows, accumulated overflows, transport time within the subcatchment, controllable structures, etc., without having too much detail that would adversely affect computational time for RTC. This surrogate technique (Wright & Davidson 2020) overcomes some of the inherent limitations of high-fidelity models as noted by Meneses et al. (2018). The hydraulics of the main transport system, e.g. transport, storage, controllable structures, overflows, etc. are described using the St Venant's equations in a hydrodynamic model (MIKE1D Pipeflow).

Data assimilation (DA), detailed weather forecasts and control strategies need to be developed and implemented. DA (Hutton et al. 2014; Lund et al. 2019) is applied to the calibrated collection system model to deal with residual anomalies and uses weighting function algorithms to optimize the online model with current and historical observations and system status.

An FCF DT continuously analyses the deviations between measured and calculated discharges and compensates for them up to the time of the forecast. The deviations are interpreted and can then be used during the forecast calculation via an ‘error forecast’ based on a first-order auto-regressive model. In this way, the DT model calculation is corrected both in the historical part (hindcast) and the forecast. The DA is based on volume corrections at each time step during a predefined time period based on observed data (hindcast) and thereafter a simpler decay function, which successively reduces the correction farther into the prognosis based on a defined ‘half time’. For example, after 6 h the correction can be about 50% and after 12 h it can be about 25% of the original correction at the time of forecast. Missing measurement data, or data with anomalies, is handled by a preprocessor that fills in gaps and repairs anomalies based on pattern recognition through a combination of multi-variable statistical methods and machine learning algorithms on historical data sets. The DA weighting can also be tailored for specific conditions or scenarios. Although other, more advanced DA alternatives are available, for this type of application, the level chosen is deemed sufficient as a good compromise between simplicity/complexity and computational time. Forecast calculations for the coming 48 h are typically performed hourly, although this can be tailored with respect to the response time and size of the connected catchment.

Control strategies are then developed for the FCF DT to mirror the necessary operating conditions or constraints in the catchment. Secure data connections are created between the FCF DT and the catchment's control system, other metres and gauges, as well as any weather forecast suppliers.

A ‘Forecast on Demand’ (FOD) feature was developed to allow the operator to modify automatically suggested strategies and rerun the simulation with these changes. With FOD, the operator can explore different strategies when, for example, faced with high hydraulic loads to an inlet pumping station. FOD can also analyse historical events for training purposes.

The FCF DT has a cloud-based GUI which provides the operator with information such as:

• a)

  state variables and current conditions (e.g. measured flows and levels).

• b)

  calculated flows and levels at selected points in the catchment.

• c)

  current and calculated sewer overflow volumes.

• d)

  water volumes in the network and basins (both hindcast and forecast).

• e)

  estimated runoff from each subcatchment for each type of flow component (i.e. inflow compared with infiltration).

• f)

  performance indicators for the catchment (e.g. maximum levels, flows and storage in the catchment).

• g)

  visual animation of the rain forecast and of the control strategy.

Gryaab's gravity-fed tunnel system has a total length of about 130 km (as stylized in Figure 3(b)) with cross-sectional areas varying from around 20 m² (with flat bottoms) closer to the WRRF to about 6 m² (with V-bottoms) in some of the tunnels farthest out in the system. About 25% of the catchment is combined sewers and there is considerable infiltration/inflow in many of the separate sewer systems. Up to about 250,000 m³ of flow attenuation is possible in Gryaab's tunnel system in two main sections, one in the north tunnel branch nearest the Rya WRRF where levels are controlled by the plant's main pumping station and the second, in the south branch, where control valves can regulate the flow through inverted siphons connecting the south branch to the north branch and thus regulate the attenuation in the south branch. The catchment experiences large flow variations during storm events and one objective was to utilize the available attenuation capacity to minimize CSO's and maintain a reasonable loading on the WRRF. Some facts about the catchment are given in Table 1.

To help meet Gryaab's goals and objectives, a FCF DT was developed to provide online simulation and forecasting capabilities with weather adaptive control. An existing MIKE+ hydrologic and hydraulic model for the tunnel system and associated catchments was updated and calibrated with flow and level measurements, both in terms of inflow components related to rain events and to infiltration/inflow from groundwater. This major system recalibration was performed in 2015 with data from available flow meters and comparison of yearly volumes. Some of the 47 subcatchments have no flow measurements, so the strategy adopted was to calibrate as many subcatchments with flow measurements as possible and then use the inflow to the WRRF as the flow measurement for the whole catchment and adjusting the remaining subcatchments to get the best possible fit.

A well calibrated model with good rain data should produce an accurate description of the flows to a WRRF. Results from the calibration performed in 2015 are shown in Table 2. It was also noted at that time the challenge of comparing measured yearly volumes with modelled results which cannot reasonably reflect all the local control exercised in the catchment that is not modelled in the model. Blockages, sediment, screenings and pump failures are examples of transient phenomena that are not taken into account in a model, but the actions required to alleviate these conditions in actual operations can change the distribution of water volumes that are led to the WRRF or directly to the recipients, thus making the comparison with modelled water volumes challenging.

For the Gryaab DT, weather forecasts are provided by SMHI (Swedish Meteorological and Hydrology Institute) and are based on the KNEP product (Korta Nederbörds (short precipitation) Ensemble Prognosis), which is a hybrid between C-band radar and several weather models, where the most likely forecast is selected (Lumley et al. 2019). New forecasts are produced hourly and have a forecast horizon of 36 h with a time resolution of 1 h and a spatial resolution of 2 km. KNEP features a gradual transition from primarily C-band radar data forecasts in the first few hours to primarily weather model-based forecasts towards the end of the forecast horizon. The probable movement of the precipitation is estimated using wind speeds and directions from the meteorological prognose models and, by using temperature, the likelihood of snow can be estimated. Each subcatchment thus gets its own customized rain series in the forecast horizon as the rain data grid is interpolated into the subcatchments and these are combined with measured rainfall data from available rain gauges in the catchments. The measured rain data are processed for quality assurance prior to being applied in a simulation by comparing rain gauges in close proximity to each other. Should data from a rain gauge be considered invalid a surrogate rain series is applied.

Large variations in rain forecasts from hour to hour can cause actions to be initiated in a DT that may require significant compensation in coming simulations. Albeit the same situation is also true for the operator who is reliant on the same information from the catchment, from weather forecasts, etc., to plan catchment and/or WRRF operations. In practice, the operator may also follow a strategy that may underreact or overreact to a given situation due to unreliable information from, for example, rain forecasts. To help the operator gauge rain forecast uncertainty, a second type of ensemble rain forecast was implemented. The SMHI product MEPS (MetCoOp Ensemble Prediction System) uses a 10-member ensemble of numerical weather forecasts with perturbed initial conditions (Lumley et al. 2019). The perturbations allow the chaotic development of the weather to give a measure of uncertainty in the forecast. This uncertainty is illustrated by the forecast band seen in Figure 5(d) and provides the operator with an indication of the variation that can be expected and thus an estimate of the uncertainty in the prognosis. With this information, the operator could, for example, test other FOD scenarios to explore possible implications for the rain event.

The reactive control constraints imbedded in the distributed controller system (DCS) for the Rya WRRF inlet pumping station as well as the control valves at the inverted siphons were mirrored in the FCF DT as the DT is subject to the same constraints. The control strategies developed utilize the simulator's flow forecasts for optimizing the pumped flow to the WRRF, tunnel and pipe volume utilization and CSOs at controllable structures in the network.

For the inlet pumping station, the operating range was divided into eight conditions, from typical dry weather flow to extreme events. In comparison to reactive systems that have no information about conditions after the time of the forecast, here the flow forecasts are used to judge the future state point for each condition and suitable setpoints are suggested to fulfil the given objectives. The principal strategies used in the Gothenburg catchment along with examples of rules for each condition are given in Table 3.

While developing the FCF RTC DT for the Gothenburg catchment, much testing was done with the operators to gain their confidence in the DT, and to help bridge over the transition from an advisory mode (i.e. recommending setpoints) to a test phase where the FCF RTC DT setpoints were directly implemented in the DCS (after plausibility control) for controlling the WRRF inlet pumping station during small and medium rain events. As this large WRRF is only part-time manned, the automatic implementation of the FCF RTC setpoints can result in, for example, less need for active operator intervention at night making standby duty easier. In more extreme events, limitations in the inlet pumping station and plant safety procedures require manual operator intervention and the FCF DT returns to an advisory mode. This experience reinforces the human factor recommendations in Torfs et al. (2022).

A comparison of key performance indicators before and after implementation of this implementation of FCF RTC is difficult since the initial development of this DT began in the early 1990s (Gustafsson et al. 1993) and many of the subcatchments have changed considerably over the years. This combined with the stochastic nature of weather and precipitation has made it difficult to quantify improvements. Working with FCF RTC has significantly improved the understanding of the dynamics of the tunnel system and many of the larger subcatchments, and their behaviour in varying situations, for example how bottlenecks in the piping or tunnels can affect system hydrodynamics in extreme events.

Once established, the usefulness of the FCF RTC DT increased as the simulator could be used in specific studies to gain insight into the behaviour of the catchment as a whole and for specific subcatchments (Gustafsson et al. 2017). One study, for example, examined the potential for CSO reduction and results indicated that yearly CSOs could be reduced by up to 65% and, simultaneously, bypass volumes at the WRRF could be reduced by 85% through the coordinated dynamic operation of the major CSO sites and increased utilization of the available tunnel volumes by installing additional sluice gates in the tunnels to create more storage. The City of Gothenburg has implemented an online CSO optimization tool, a spin-off from FCF RTC, with the objective to minimize pollutant loading on the recipients in their local catchments. As mentioned in section 1, FCF Planning was developed to assist utilities, city planners and others with optimizing mitigation plans, budget allocation and decision-making and thus target resources where they will have the greatest impact in improving a catchment. The simplified, surrogate models used in FCF Planning allow simulation times in the order of minutes where the more detailed models used in FCF RTC require considerably longer simulation times.

The FCF DT provides a sound base for modelling, simulating, forecasting and controlling wastewater catchments. Valuable insight was gained into catchment dynamics while creating the DT and this insight could be used to quantify and optimize CSO reductions, flows to the WRRF, etc. according to the defined objectives. The predictive control imbedded in FCF recommends setpoints for controlling the catchment that can provide decision support for operations or even be directly applied to the DCS as model-based predictive control. The FCF forecasts together with the possibility to test scenarios with FOD help the operators be better prepared for large rain events. Besides the need for accurate catchment data when building a DT, a DT needs good quality online data from the catchment and, perhaps most importantly, high-quality rain forecasts, especially when prognosing long time horizons with the DT.

The Future City Flow consortium was partially financed by Vinnova, the Swedish governmental agency for innovation systems, and which is hereby gratefully acknowledged.

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

The authors declare there is no conflict.