See Finished Assignments

Control methodology for domestic energy efficiency optimization

Master's assignment

Student: Haining Wu
Committee: Albert Molderink, Gerard Smit, Maurice Bosman and Bert Molenkamp     
Programme: Embedded Systems - University of Twente
Finished: September 2011


In the recent decades, the need of energy is increased a lot. Meanwhile, the energy shortage crisis is getting worse. To deal with this problem and thanks to development of ICT technology, the energy generation is moving from central to decentralized, more renewable resource (sunshine, wind, etc) appeared instead of conventional resource. Rather than passively purchasing energy from the grid, the user changes to play an active role in the grid. They not only can consume the energy effectively, but also can produce energy, communicate with other users and feedback to the grid. Thus, with smart devices, smart controllers and communication technology, the grid behaviors in a smart way —so called “smart grid”.

In order to keep electricity grid with those changes function properly, without changing a lot of infrastructures in the current grid, many studies concern about the research of the new control methodologies. Among those studies, University of Twente proposed a new methodology to schedule domestic appliances, called “three-step methodology”. As can be seen from the name, the methodology consists of three steps. At the first step, a prediction of the demand and supply in the second day is made. At the second step, a plan of scheduling each device in each time interval in the following day is made based on the prediction. At the final step, a real time controller tries to schedule every device’s behavior following the plan, and guard users’ comfort with real time deviation from the plan, e.g. demand, price, etc.

Up to now, the major scheduling algorithm used in the third step is based on an Integer Linear Programming (ILP) formulation in each time interval. Via defining coefficients of decision variables in the cost function and adding steering signals from the plan, the algorithm tends to give preference to follow the plan when the prediction deviation is not large. In the simulations, this method shows very undesirable behavior when there’s large prediction deviation – it cannot follow the plan all the time and shows an irregular behavior, e.g. frequently turn a device on or off. To solve this problem, Model Predictive Control (MPC) is introduced to the third step. By looking ahead a few time intervals and calculate the minimum (maximum) cost in those intervals, MPC makes decision at the current interval. Simulation results show MPC outperforms ILP when prediction errors occur. But still, there’s some problems existed: 1. MPC takes much more time to execute, even though it can give an improved result, the execution time prevents it to be used in time-critical situations; 2. Due to the limit of the prediction and the plan, it is not robust enough to deal with some large deviation.

For the problems raised up above, the thesis gives four ideas on the improvements. Firstly, the MPC can be made explicit to save time during execution; secondly, probability is introduced into MPC to make it more robust against prediction error; thirdly, for the whole procedure, a judgment whether to execute MPC is made first before execution, if not, execute ILP instead. In this way, executing MPC only when it is necessary speeds up the procedure a lot; lastly, when the prediction error is too big to be worked around by MPC, re-planning is made by the global controller. All explanations of these ideas refer to Chapter 4.

To test the effects of those ideas, to verify case models, and to find correct parameters of the model, those ideas are tested in three use cases. The first one is scheduling a group of cars. The second one is scheduling smart devices in a single house. The third one is scheduling a group of micoCHPs. The models and problem formulations are described in Chapter 5. The simulation results are shown in Chapter 6. The results again prove MPC gives better performance when there are prediction errors. Besides that, in the car case, explicit MPC works fast online but has a high requirement for the computer memory and takes a lot of time to compute on beforehand. It can bring benefit especially when there is one plan for a long time. Probabilistic MPC shows better robustness with higher confidence level. In the single house case, the combination of ILP and MPC is proved to work for each device but the criteria of the judgment is dependent on the different type of devices. In the third case, re-planning is performed when the local controller finds that it cannot follow the plan with MPC. It shows that with prediction error, the global controller can make a new plan for the following intervals in real-time, under the condition that the plan doesn’t need to take much time to make. After that, local controllers schedules based on the new plan.

From the simulation results we can conclude that those improvements can be implemented in the context of the current algorithm, i.e. without changing the current structure. They can improve both the speed and robustness aspects and they are scalable and adaptive to different use cases. This work provides an important reference to the further implementation of MPC in the three-step methodology in University of Twente.