2.png
Model predictive controle

Introduction 

Optimal control is at the core of what we do at DeltaQ. More concretely, an extension of optimal control Model Predictive Control. In the MPC problem what we aim to do is to pose the optimal control problem in a way such that the optimization problem accounts for the evolution (I.e., the dynamics) of the system (in our case the thermal evolution of the building, and therefore its expected energy consumption) and respects the constraints of the system in question (e.g., thermostatic setpoints).  

City buildings_Monochromatic.png
image.0.png

Figure 1 shows the idea. Given a reference trajectory (or in our case thermostatic setpoints agreed with the customer) we want to keep the predicted output within boundaries (in the case below the cost function tracks the trajectory). As a result of the optimisation problem you have a prediction of control actions, not only for k+1 but for k+p (usually p, the control horizon, is denoted in the literature as N, or NH). In our case we use a sampling time of 15 minutes.  

While a system could be informed of the previous control actions, in our current configuration we use merely the past actions and estimate the state of the system (i.e., the evolution of internal variables besides the measurable ‘output’ variables). Our approach is what it is most common in the literature (measure/estimate intitial state, compute optimal problem, implement the first output value (e.g., send control signals to the building). The process is repeated at the next time-step and so on. Hence, if well implemented, the MPC model is self-corrective (due to state estimation and the corrective nature at the next time step.  

Project Definition 

Introducing blackbox models (not only NN) is at the core of what we want to do next when  it comes modelling. NN plus heuristic is not exaclty the newest of ideas, but which is  applicable nonetheless. We have the technical abilities and know-how to do this and we have started work on that direction. We are hoping for some fresh input on directions that  we have not considered to solve the blackbox, data-driven MPC problem.

Data 

  •  Weather data (file d.p): 

    • Outdoor temperature 

    •  Facade_irradiance on directions north (0003), south (0002), east (0000) and  west (0001).  

    • Wind speed (often not required, feel free to use it)    

  • Space temperature measurements (file d.p)

    • Space_temperature_air (room temp measurement) for each space_00XX. 

    • In this context a space_00XX is a room in the building 

  • Control inputs (file u.p):  

    • df_control_input (cooling_mode and heating_mode for each vrv_00XX)

    • In this context a vrv_00XX is a VRV end-unit, with two modes (heating/cooling). 

    • Normally there are other modes (automatic, failure etc) but we have cleaned  those out and made all conversions for what we think are the  activation/deactivation of the system on a heating/cooling mode.

Project members​

Project Responsible (Board):

Mathijs Carlu

Project Leader:

Boris Dragnev

Project Members:

Divin Kirenga

Louise Koopman

Michiel Van Nijverseel

Tommy Saeys