Report

Overview of Model Predictive Control in Buildings Tony Kelman MPC Lab, Berkeley Mechanical Engineering Email: [email protected] Tony Kelman MPC in Buildings January 11th – Slide 1 Outline Model predictive control Basic idea and elements Advantages, disadvantages Modeling and MPC in buildings What works, what doesn’t Generating models using historical data Advanced control behavior Experimental projects Success stories Increased scope and capabilities over time Tony Kelman MPC in Buildings January 11th – Slide 2 Model Predictive Control Use predictive knowledge for control Basic components System model – state evolution vs inputs and disturbances Constraints on inputs or states – requirements, actuator limits Cost function – reference tracking, energy, comfort Forecast trajectories of future disturbance inputs – weather, occupancy, utility rates Optimization algorithm – fast enough to solve in real time Advantages: multivariable, model based, nonlinear, constraint satisfaction, incorporates predictions Disadvantages: computational complexity, design effort of accurate modeling Tony Kelman MPC in Buildings January 11th – Slide 3 Model Predictive Control Initialize with current measurements at time t Predict response over horizon of p steps Solve for best input sequence, apply first element u*(t) Repeat at time t+1 with new measurements (feedback) Tony Kelman MPC in Buildings January 11th – Slide 4 Optimization Formulation Predicted states xk, inputs uk, disturbances wk At each time step, solve: Constrained finite time optimal control problem Optimization much faster if explicit structure of J, f, g (and derivatives) can be provided Tony Kelman MPC in Buildings January 11th – Slide 5 Outline Model predictive control Basic idea and elements Advantages, disadvantages Modeling and MPC in buildings What works, what doesn’t Generating models using historical data Advanced control behavior Experimental projects Success stories Increased scope and capabilities over time Tony Kelman MPC in Buildings January 11th – Slide 6 Modeling for Building Energy Systems Common practice is black-box simulation DOE2, EnergyPlus, TRNSYS, etc Useful for design, very difficult to use for control Derivative-free optimization not very efficient or scalable Need model structure for optimization and control Simpler approach: reduced order modeling Physics based model structure Data driven parameter identification Can adjust accuracy vs complexity tradeoff Large scale real time optimization tractable Tony Kelman MPC in Buildings January 11th – Slide 7 HVAC Example System HVAC good target for energy savings by better control Common configuration for commercial buildings: VAV with reheat Control inputs: supply fan, cooling coil, heating coils, zone dampers, air handling unit dampers States: zone temperatures Tony Kelman MPC in Buildings January 11th – Slide 8 Thermal Zone Model Tony Kelman MPC in Buildings January 11th – Slide 9 Simplest Useful Model Abstraction Network of bilinear systems Thermal zone model u1 u2 Q_2 Q_1 A (simple extension to multiple states per zone, RC network analog) Q_n un Static nonlinearities Equipment performance maps (chillers, cooling towers, pumps, fans, coils) Equality and inequality constraints Comfort range Dynamic coupling: thermal zones, supply air & return air Uncertain load predictions Human: occupancy, thermal comfort, … Environment: ambient temperature, solar radiation, … Tony Kelman MPC in Buildings January 11th – Slide 10 How to Generate Reduced Models Several options to create model data Direct physics based lumped parameters Model reduction from high fidelity design tools Use historical data for model identification Identification results vs measured data, Bancroft library Tony Kelman MPC in Buildings January 11th – Slide 11 Using Data to Quantify Uncertainty SMPC Prediction model Historical load realization Load Ambient temperature Tony Kelman MPC in Buildings January 11th – Slide 12 Advanced Control Behavior A. Kelman, Y. Ma, A. Daly, F. Borrelli, Predictive Control for Energy Efficient Buildings with Thermal Storage: Modeling, Stimulation, and Experiments, IEEE Control System Magazine, 32(1), page 44-64, February 2012. Time-varying price Penalize peak power MPC is able to incorporate time-varying energy price and reduce peak power consumption Tony Kelman MPC in Buildings January 11th – Slide 13 Outline Model predictive control Basic idea and elements Advantages, disadvantages Modeling and MPC in buildings What works, what doesn’t Generating models using historical data Advanced control behavior Experimental projects Success stories Increased scope and capabilities over time Tony Kelman MPC in Buildings January 11th – Slide 14 Experimental Projects UC Merced –, Merced, CA 4% Improvement LBNL+UTRC- Storage, Chiller Optimization Horizon 24hrs, Sampling 30min Problem Size: ~300 variables , ~1440 constraints CERL Engineering Research Laboratory, Champaign, IL 15% improvement. UTRC- HVAC distribution – 5 zones Horizon 4hrs, Sampling 20 min, Problem Size: ~1600 variables , ~1400 constraints Naval Station Great Lakes, North Chicago, Illinois UTRC- Conversion + Storage – 250 zones Problem Size: ~~20k variables , ~?? constraints CITRIS Building (UC Berkeley) – Major issues Siemens - Generation + HVAC distribution -135 Zones Horizon 4hrs, Sampling 20 min, Problem Size: ~~10k variables , ~?? constraints Brower Center (Slab Radiant), Berkeley, CA Architecture Department Models based on step tests experiments White Oak, Silver Spring, MD Simplified models and BLOM tool critical for realtime implementation of large MPC experiments Honeywell Microgrid Optimization Tony Kelman MPC in Buildings January 11th – Slide 15 Distributed Implementation Tony Kelman MPC in Buildings January 11th – Slide 16 Distributed Implementation Coordinator Dual variables Supply Fan Cooling coil damper Tony Kelman Heating coil VAV damper Zone temperature MPC in Buildings January 11th – Slide 17