Demand-side Energy Management in Smart Buildings Prashant Shenoy University of Massachusetts GreenMetrics 2013 Keynote School of Computer Science
Motivation: Buildings and their Energy Usage q Buildings are significant energy consumers § 76% of electricity and 48% of total energy in US University of Massachusetts Amherst - School of Computer Science 2
Building Energy Usage Breakdown q Residential: lighting: 11%, HVAC: 55% q Office: Lighting: 26%, HVAC: 50% University of Massachusetts Amherst - School of Computer Science 3
Green Net-Zero Buildings q Net-zero buildings: zero overall footprint § “Green” design: balance generation and consumption § Many new green buildings are net zero q What about existing buildings? § Fact: 80-90% of buildings we will encounter already built University of Massachusetts Amherst - School of Computer Science 4
How to Smarten and Green Buildings? q Demand-side Energy Management (DSEM) § Manage energy usage by regulating demand q Use on-site renewables : solar, wind, geo-thermal § Fall back to grid only when needed § Reduces carbon footprint and grid load q Manual : reduce usage, conserve energy § e.g., turn off some lights when not needed q Automated DSEM § Sense-Analyze-Control Approach § Grid provides signals, smart-building responds University of Massachusetts Amherst - School of Computer Science 5
Intelligently Reducing Energy Use q Intelligent DSEM q Automatically defer elastic loads § Control charging of Electric Vehicles § Schedule appliances q Identify and eliminate waste § Align AC thermostat schedules to occupancy patterns q Manage demand during peak periods University of Massachusetts Amherst - School of Computer Science 6
Peak Load Reduction: Why and How? q Peak grid load: disproportionate marginal and environmental costs § Peaking power plants: inefficient and “dirty” § Lower peaks: reduced brown-outs q Peak load reduction techniques § Time-shift supply: Use energy storage • charge at off-peak and use at peak § Time-shift demand: schedule loads, shed loads University of Massachusetts Amherst - School of Computer Science 7
Modeling and Prediction Challenges q Modeling and prediction is key § Need to understand before we can optimize q Sense-analyze-control approach to smart buildings § Modeling, prediction key to analysis and control q Analyze, model and predict building energy usage q Model individual loads as well as aggregate loads § How to model electrical loads in a home? University of Massachusetts Amherst - School of Computer Science 8
Talk Outline q Motivation q Background on Electrical Loads q Modeling Electrical Loads q Using the Models q Conclusions University of Massachusetts Amherst - School of Computer Science 9
Smart Meters and Energy Monitors q Smart meters can meter fine-grain real-time usage § Utility-grade: 1-5 min resolution § Consumer-grade: 1sec resolution q Meters and sensors can monitor individual loads § Per circuit-breaker or per outlet q Data : 3 homes for 2+ years, >80 breakers, 100+ devices University of Massachusetts Amherst - School of Computer Science 10
Modeling Electrical Loads q Prior work: modeling aggregate demand profiles [ISO] q Individual loads: simple on-off models [Hart’89,Kim’12] § Device/load can either be on or off § Draws a fixed power when on § Example: light bulb light 250 Power (W) 200 150 100 50 0 0 1 2 3 4 Time (min) q Simple extension: multiple discrete ‘on’ states [REDD] University of Massachusetts Amherst - School of Computer Science 11
Today’s Electrical Loads q Today’s devices are significantly more complex § exhibit rich, complex variations in power usage HRV 1000 180 LCD TV 160 Power (W) 800 140 Power (W) 120 600 100 80 400 60 40 200 20 0 0 0 1 2 3 4 5 0 5 10 15 20 Time (min) Time (min) q Question: How can we design better models to capture this behavior? University of Massachusetts Amherst - School of Computer Science 12
Electrical Loads: A Primer q Loads: resistive, inductive, capacitive, nonlinear q Resistive : AC current and voltage waveforms align q Devices with heating (pure resistive) elements § Lights, toaster, coffee maker, oven, space heater University of Massachusetts Amherst - School of Computer Science 13
Electrical Loads q Inductive loads: current waveform lags voltage § Devices with AC Motors: AC, vacuum cleaner, fridge q Non-linear loads: non-sinusoidal current draw § Electronic devices with switch-mode power supplies • LCD TV, music system, computer, battery chargers University of Massachusetts Amherst - School of Computer Science 14
Talk Outline q Motivation q Background on Electrical Loads q Modeling Electrical Loads q Using the Models q Conclusions University of Massachusetts Amherst - School of Computer Science 15
On-Off Model q On-Off Model: two state model § states: q Captures behavior of small resistive loads q Example: Lights light 250 Power (W) 200 150 100 50 0 0 1 2 3 4 Time (min) University of Massachusetts Amherst - School of Computer Science 16
On-Off Decay Model q Models inductive and large resistive loads § rush of current at startup, then settles to steady power q On-off decay: on surge, decay to stable, off § active, inactive, peak power § rate of exponential decay ⇢ p active + ( p peak − p active ) e − λ t , 0 ≤ t < t active p ( t ) = , t ≥ t active p off 800 refrigerator coffee maker 1000 1600 700 vacuum cleaner 1400 Power (W) 600 800 Power (W) Power (W) 1200 500 600 1000 400 800 400 300 600 200 400 200 100 200 0 0 0 0 1 2 3 4 5 6 7 8 9 10 0 20 40 60 80 100 120 0 10 20 30 40 50 60 Time (min) Time (sec) Time (min) University of Massachusetts Amherst - School of Computer Science 17
Stable Min-Max Model q Deviations from a stable min or max power q Captures behavior of non-linear loads q Model parameters: § Active power § Maximum ‘spike’ deviation (uniformly distributed) § Mean inter-arrival time (exponentially distributed) 45 HRV Mac Mini 1000 180 LCD TV 40 160 35 Power (W) 800 Power (W) 140 Power (W) 30 120 600 25 100 80 20 400 60 15 40 10 200 20 5 0 0 0 0 1 2 3 4 5 0 5 10 15 20 0 5 10 15 20 Time (min) Time (min) Time (min) University of Massachusetts Amherst - School of Computer Science 18
Range-bound Random Model q Some non-linear loads lack a stable min or max, but are range-bound § E.g., microwave q Model these devices as a random walk § Bounded by a max, min power 1500 microwave p max = 1480 1480 1460 Power (W) 1440 1420 p min = 1400 1400 1380 0 0.5 1 1.5 2 2.5 3 Time (min) University of Massachusetts Amherst - School of Computer Science 19
Basic Model Types Model Load Examples On-Off Pure Resistive Lights motors, large On-off decay Inductive, large resistive heating elements Stable min-max Non-linear LCD TV Range-bound random Non-linear Microwave University of Massachusetts Amherst - School of Computer Science 20
Model Accuracy q Compare models to actual device power signature q Model as on-off decay versus simple on-off § Entire device cycle and first 30 seconds only 150 Root Mean Square Error 125 100 On-o ff Decay 75 On-o ff Decay (first 30 secs) On-o ff 50 On-o ff (first 30 secs) 25 0 Coffee maker Toaster Dryer q Models more accurate at capturing device behavior University of Massachusetts Amherst - School of Computer Science 21
Composite Loads q Many devices consist of multiple basic load types § Fridge: compressor, light, water dispenser § Dishwasher: motor, pump, heating element § Central AC: compressor, fan, humidifier q Composite model: composition of basic models § Parallel composition § Serial composition § Cyclic/Periodic : repeating sequence on-off decay, cyclic 600 washing machine random range, cyclic on-off decay 500 (growth) random 400 Power (W) stable min range 300 on-off decay, 200 cyclic 100 0 0 5 10 15 20 25 30 35 40 45 Time (min) University of Massachusetts Amherst - School of Computer Science 22
Talk Outline q Motivation q Background on Electrical Loads q Modeling Electrical Loads q Using the Models q Conclusions University of Massachusetts Amherst - School of Computer Science 23
Using the Models: NILM q NILM: Non-intrusive Load Monitoring § Disaggregate individual loads from aggregate § Premise: individual loads have discernible “features” q Prior work: edge detection or HMM-based § Assume loads are on-off devices q Better load models => Better NILM University of Massachusetts Amherst - School of Computer Science 24
Waste Identification q HVAC: > 40-50% of total usage § Controlled by a thermostat q Thermostats are hard to program “correct” 7 6 5 power 4 3 2 1 0 500 1000 1500 2000 2500 3000 Time [minutes] q Correlate occupancy with thermostat schedules § Identify waste § Derive “optimal” schedules for each building/home University of Massachusetts Amherst - School of Computer Science 25
Recommend
More recommend
