WaterPACT: Water Prediction and Control Technology

It is one thing to know how your water system will behave in the days to come. How proud we were when we were able to migrate our, for those days, enormous numerical models from the confined environment of our research institutes to a real-time environment. It is a second thing to extend these deterministic forecasts with certainty ranges, which is something that ensemble forecasts allow you to do. Knowing what will happen, with what probability, is a very valuable piece of information for the decision maker facing an extreme event. Yet he or she is left with still one question to be answered: ‘What to do now?’.

Here we enter the field of operational water management, real-time control, as described by a more recent initiative Water Prediction and Control Technology (www.waterpact.org and the 2013 special issue of Journal of Hydroinformatics on WaterPaCT). The ‘what to do now’-question can be answered in a rational manner by means of real-time optimization using the present state of the water system, the predictions and the system constraints, both physical and operational. The uncertainties in the predictions can be incorporated in the optimization, but past implementations (e.g. robust model predictive control) have shown to lead to too conservative control actions, for example have led to always evacuate, even if only one ensemble member causes exceeding a maximum water level threshold. A more recent optimization scheme that incorporates an ensemble system in a more balanced manner is the Tree-based Model Predictive Control. This method constructs a tree from the ensemble system and considers a branching point to be a location where the uncertainty, of which ensemble member (or branch of very similar ensemble members) is actually occurring, is resolved. Hence, control actions for the now-moment are computed, given the effect of the entire spectrum of possible future trajectories and their probabilities.

Result of Tree-based Model Predictive Control maintaining a water level below a maximum threshold compared to Model Predictive Control using deterministic actual forecasts and MPC using perfect forecasts (upper bound to performance)

Result of Tree-based Model Predictive Control maintaining a water level below a maximum threshold compared to Model Predictive Control using deterministic actual forecasts and MPC using perfect forecasts (upper bound to performance)

These rational optimal control actions can be communicated to the decision-maker as one of the alternatives in his hectic search for the right decision anticipating a disaster. When the objective function of the optimization is right, as is the system model, it will be the right choice given the ensemble forecast that is available at that moment.

1 comments

  1. Sorry, question to your figure: the maximum threshold is at -0.60? What doe set point mean? How do you derive a perfect forecast?

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.