The third knowledge exchange meeting of the Hydroinformatics Platform on 25 June focused on uncertainty and uncertainty propagation. More and more decisions are based on measurements and models, which in turn draw on measurements and system knowledge. It is known that these measurements include errors and that system knowledge is not perfect either. Nevertheless, in general, decisions are based firmly on the assumption that the available results are correct. It is not yet common practice to take errors and uncertainties into account in decisions based on the measurements and models. The effects of uncertainties, and how they can be taken into account in decision-making, including their added value, were on the agenda during the afternoon.
Uncertainty is everywhere. Weather forecasts are a familiar example: the behaviour of a complex system – the atmosphere – can be predicted to only a limited extent. Due to the imperfection of models and the incompleteness and limited accuracy of measurements, a reasonably accurate weather forecast can be given only a few days in advance. That is usually not possible for longer-term predictions and inaccuracy increases as the forecasting horizon becomes more distant.
That also applies to other approaches involving measurements and models. However small the uncertainty, it has an effect on the results. We call this ‘uncertainty propagation’. Generally, uncertainties are not taken into account in the further use of the data. The fact that measurement results and predictions are considered correct is known as the ‘deterministic approach’.
There has been a change in thinking in a number of areas. Instead of a deterministic approach focused on expected events, people are switching to thinking in terms of possible events and probabilities: the probabilistic approach. This involves using current knowledge to determine the probability of a particular scenario, A or B etc, as well as the explicit consideration of uncertainties. Examples include predictions of extreme rainfall due to climate change and requirements for the safety of dikes.
Uncertainties in drinking water systems
Kicking off the meeting, KWR researcher Peter van Thienen gave examples of sources of uncertainty in drinking water on the basis of hydraulic models. For example, there may be uncertainty in parameter values such as the diameter of a slowly closing pipe. Or there is unpredictability, as with demand for water in the long term, because a number of factors, none of which can be determined completely, are involved: demographics, policy, technology and conduct.
If these uncertainties are included in the modelling, the results are no longer singular. The models will generate a range of solutions with a probability distribution. As a result, the performance of a network may differ significantly from the modelled deterministic situation.
Figure 1: Probabilistic pressure distribution on a node in a network model. The uncertainties in pipe diameters, wall roughness, valve positions, etc. make it highly likely that the actual pressure will not be the same as the pressure found deterministically.
It is important for the probabilistic approach to provide an indication of the probability of given outcomes. This makes it possible, for example, to produce better designs but also to intervene better when there are emergencies.
Predictive model supports gas network operation
The second contribution of the afternoon came from Stephan Talle and Onno Wesselink of Accenture. They explained how Gasunie uses a decision-supporting simulation model to manage the national gas network. This Gas Transport Management System was built on the basis of a number of coupled models. A real-time simulation model (a digital twin of the transport network) calculates the most likely current network status on the basis of measurement data and an up-to-date network model. This is worked up with predictions of future energy consumption using the Flow Forecaster, a 4-layer neural network model that predicts future gas demand for 150 gas take-off stations. A simulation model can then calculate the future network status for up to the next 48 hours on the basis of the forecast demand for gas.
There are various uncertainties in this system, as became clear during the presentation. For example, only a limited number of locations are fitted with sensors. The situation in the rest of the network is predicted on the basis of a model. In addition, the Flow Forecaster uses historical data about the weather and gas take-off to derive associations for expected gas consumption. Although the models result in the most probable process values in the network, they do not produce ranges or probability distributions. The advantage of a probabilistic approach is that it sees low probabilities with major consequences, including the impact of a management action. However, care should be taken not to overload operators with information about highly improbable scenarios.
Figure 2: Schematic representation of the links between the models in Gasunie’s Gas Transport Management System.
Optimisation of backwashing for sand filters
Lucy Ng’ang’a, a trainee at Brabant Water, explained how forecasting is used to manage the washing of sand filters. This process produces water containing large amounts of sediment that can only be discharged after being left for some time in sedimentation basins. In traditional management approaches, there is a possibility that the number of filter washings may exceed the daily capacity, as a result of which the water discharged from the sedimentation basin will not comply with the permits obtained by the water authority. A model has been developed to calculate the time required before the next rinsing of the filter. These predictions are made using measurement data and a long short-term memory (LSTM), recurrent neural network (RNN) approach. On the basis of a number of criteria, such as a minimum interval of six hours between rinses and a maximum runtime for each filter, the model determines control rules for optimising the intervals between filter rinses.
Uncertainties in water levels
At the end of the afternoon, Ferdinand Diermanse of Deltares talked about how to work with uncertainties relating to flood risk management , for example in the design and assessment of flood defences. On the one hand, it is impossible, for example, to map out every detail of a dike body and to determine its actual strength with complete certainty. At the same time, dike capacity to withstand extreme loads (such as sea level rise and wave impact) is assessed with numerical models. In addition to the uncertainty in these models, there are also uncertainties relating to the loads themselves. Sea level rise, for example, is highly uncertain. In the case of flood defences, those uncertainties are addressed using safety factors. Furthermore, an adaptive strategy of the kind used in the Delta Programme uses a step-by-step plan of measures based on probabilities and vulnerabilities. Short-term interventions and long-term options are also included in the plan. Furthermore, adaptation tipping points (ATPs) – points at which the current strategy will no longer work – have been defined. The long-term adaptation strategy is adjusted on the basis of the continuous monitoring of the implemented actions and developments in the system as a whole. Over the next twenty years, this will involve reserving space, building in flexibility, experimentation and using opportunities created by other developments.
Certainty about uncertainty
On the basis of the knowledge exchange meeting, the conclusion is that decisions can be placed on a firmer footing if uncertainties are taken into account in data and models. A probabilistic method is the appropriate way to do this. A deterministic approach is less appropriate in terms of finding the right solution when there are uncertainties at different scale levels. At the same time, it became clear that the probabilistic approach is far from being in widespread use and also that communications about uncertainty are not easy. That is an extra incentive to maintain our discussions about this area in order not to create a world full of false certainties.
Would you like to read more about uncertainty and uncertainty propagation? There is a good review article in H2O.