Performance assessment of tree-based model predictive control

Performance assessment of tree-based model predictive control

Stive, P.M.

Van de Giesen, N.C. (mentor)

Civil Engineering and Geosciences

Watermanagement

Water Resources

2011-06-22

This research focuses on polder-belt canal systems. More is demanded from these systems every day. Man induced changes, like increasing population density and increasing land value on one hand and climate change in the form of longer dry spells and more extreme precipitation events on the other hand are the main sources. The operation of the structures in these systems plays a critical role in successfully dealing with these challenges. To get the most out of the current system and its structures, operation by humans alone is not enough, they need to be aided by computers. A promising technique is Model Predictive Control (MPC). A control algorithm that uses a model of the system and forecasts of the future disturbances to determine the control actions for the structures, whilst adhering to the constraints of the system. Forecasts are uncertain and are therefore provided in the form of ensemble forecasts that consist of multiple scenarios. MPC uses only one scenario and is thus vulnerable to these uncertainties. Treebased Model Predictive Control (TBMPC) considers the complete ensemble to determine control actions. TBMPC has, however, only been tested in theory. Only open loop simulations have been carried out, no continuous closed loop simulations have been done. TBMPC uses the complete ensemble, but to save calculation time reduces it to a tree-shaped representative ensemble with fewer nodes. This means aggregating nodes and scenarios on various points in the ensemble. There are multiple rules that determine which nodes and scenarios are aggregated, however, their optimal setting is not known. Also if TBMPC has more added benefit over MPC on certain system configurations (e.g. configurations with higher discharge or storage capacity) is not known either. A model is developed to simulate the performance of MPC and TBMPC. It can deal with different precipitation series, forecasts, system configurations and control algorithm parameters. All simulations have a duration of one year and a one hour time step. The rules that determine which nodes and scenarios are aggregated are investigated first. Transforming the inflow forecast scenarios to cumulative inflow scenarios before determining which nodes and scenarios to aggregate yields better results. The threshold value is also important as it determines whether or not two scenarios are close enough to each other to be aggregated. Nothing was known about the right value for this parameter. One of the objectives was to be able to fine tune this value to the system con guration. Setting this value as a percentage of the maximum pump capacity of the system works well across di erent system configurations. The optimal value is 100% of the maximum pump capacity. The scenario reduction algorithm (i.e. the algorithm that creates the tree from the original ensemble) has two parts. First it reduces the number of scenarios in the ensemble to a predefined number and secondly it creates a tree out of the reduced ensemble. No information was available about the right amount of scenarios for the first part. The simulations show that using more than eight to 10 scenarios does not yield any better performance, but only increases calculation time. To determine if TBMPC is more beneficial on certain system configurations 81 configurations are examined. The performance of MPC with a perfect forecast (i.e. equal to the inflow), MPC and TBMPC is simulated for these configurations. For all configurations TBMPC shows a considerable added benefit, however, there are different reasons for different configurations. For configurations with a high storage and pump capacity the increased performance can be attributed to a more stable water level and a slight improvement in the deviation from set point. For configurations with a low storage and pump capacity the added benefit of TBMPC is seen in a better peak event anticipation. Overall significant improvements to TBMPC have been realized. It is shown that TBMPC not only works in theory, but provides benefits over MPC in practice for a multitude of configurations as well. TBMPC can now be tuned to the water system configuration it is used on and it can be set to reduce calculation time as much as possible without decreasing the performance.

tree-based model predictive control
performance assessment

http://resolver.tudelft.nl/uuid:79a718a3-cd99-4f44-8391-3de5e93c3dcd

2011-07-21

Student theses

master thesis

(c) 2011 Stive, P.M.