IPSN 2013

Matchstick: A Room-to-Room Thermal Model for Predicting Indoor Temperature from Wireless

Sensor Data

IPSN 2013Carl Ellis (School of Computing and Communications, Lancaster

University, UK), Mike Hazas (School of Computing and Communications, Lancaster University, UK), James Scott (Microsoft

Research, Cambridge, UK)

NSLab study group 2013/4/15Speaker : Chia-Chih,Lin

Outline

• Introduction• Deployment• Modeling• Evaluation• Conclusion

Introduction

• Home space heating accounted for 62% of total domestic energy consumed

• Typically equipped with a programmable thermostate• several methods has worked to improve the comfort

and energy saving• But these utilized simple heating models for their

houses• Complementary ,a heating model could allow future

temperature trends to be predicted

Introduction cont.

• The paper purposes that simple temperature sensor, combined with real-time algorithms

• Two features for model– Recognize different spaces heat and cool not only

due to insulation, but also thermal masses– Automatically identifies rooms which appear to

have a thermal relationship

Introduction cont.

• Contribution:– Predictive performance : two-hour lookahead

error 1.5 degree or better (90% confidence level)– Highlight the energy savings opportunities

Outline

• Introduction• Deployment• Modelling• Evaluation• Conclusion

Deployment

• 4 houses– 2 in US (US1,US2)– 2 in UK (UK1,UK2)

• Variety of sensors used– UK : .NET Gadgeteer (ref. [17])– US :iButton Thermochron sensors

• UK home data: each homes radiator could be actuated independently

• Over various winter periods in 2010-2011

Deployment

• In UK deployment– WSN with 802.15.4 radio network to a PC server

located in house.– (per room temperature data) logged 5 sec/time– Outdoor temperature gathered from a local weather

station– Whole house gas measurement– Thermostatic radiator valves were actuated by House

Heat FHT-8Vs(controlled by PC)– Reading were downsampled to one measure per 5 mins

FHT 8V Wireless Actuator

Deployment

• In US deployment– 20 iButton Thermochrons– At least one in each room– 2~3 in large room– Out door temperature get by putting one iButton

outside– One place on furnace directly to sense actuation

time– Sensor sampled 10 mins per time

Deployment

• Building Characteristic– UK1 : • two-floor building with a gas boiler, TRV-equipped

radiators• Underfloor heating in first floor, radiator in second floor

– UK2 : three floor 19th century house with wall-mounted convection radiators

– US1&US2 : • north-west USA• Air heating system(powered by a furnace)

Outline

• Introduction• Deployment• Modeling• Evaluation• Conclusion

Modeling

• Use a regression based optimization model• Consider room-to-room interaction, thermal

mass delay, and outside temperature• Use a non-linear transformation of gas use• Fits between the heating scheduler

Modeling

• Training by historical data, then using model parameters to predict the result and adjust the schedule, parameters involves:– Current sensor data– Heating schedule

Thermal mass delay

• Delay between thermal energy input, change of the heating element temperature, and ambient indoor air temperature

Recursive non-linear transformation function

Gt : gas usageσ: thermal energy(stored in room’s heating element,[0~1])RTn :empirically determined by search the solution space and finding value when traning the model with historical data

Internal interaction between rooms

• Need to determine the thermally significant neighbors automatically [18]

• Recursive likelihood test is performed• Initially fitted with no neighbors-> likelihood-

ratio test ->if null hypothesis is rejected->the most likely neighbor added

Fitting the Matchstick Model

The mathematical form of Matchstick’s system equations

N : the set of all roomTn : temperature of room nG : gas usedTO : outside temperatureαt : loss of heat from the roomαg :heat transfer from the heating spaceβnj :transfer of heat from thermally significant neighboring roomsϒo : heat transfer with the outside

Outline

• Introduction• Deployment• Modeling• Evaluation• Conclusion

Evaluation

• Characterize the predictive accuracy of the model

• Analyze how the predictive accuracy changes for different rooms in different houses

• Investigate the effect of the model’s training aspects

predictive accuracy of the model

• 3 weeks predict ,1 week as training data• Supply two types of future knowledge– Future gas input– Future outside temperature

• Train model -> for each time step t(0~24) can predict p hours -> modeling each time step until t+p reached -> stored and compare to ground truth -> make error distribution

• (p : 1.5hr~6hr)

predictive accuracy of the model

different rooms in different houses

Compare to other model

[15] : each room relied upon the predictions of others in their model[9] : could be because the model does not capture neighboring interaction

Model Tuning

• How training data affect the model– Length of training data– How to select initial neighboring rooms to be

passed to the model

Length of training data

Using different policies for neighboring rooms

Saving analysis

Results

• UK1 saved 3.3% of its total gas• UK2 saved 2.3%• Original study [7] improve 8-18%

Outline

• Introduction• Deployment• Modeling• Evaluation• Conclusion

Conclusion

• Matchstick, a data driven adaptive model • Relies on relatively sparse sensor deployments• Predicts across three weeks of data in four

houses in two different countries.• Can achieve gas savings by trimming down

furnace or boiler actuation schedules

Q&A

• Thanks for listening !