SunCast : Fine-grained Prediction of Natural Sunlight Levels for Improved Daylight Harvesting

SunCast: Fine-grained Prediction of NaturalSunlight Levels for Improved Daylight

Harvesting

Jiakang Lu and Kamin WhitehouseDepartment of Computer Science, University of

VirginiaIPSN 2012

Outline

• Introduction• SunCast• Related work• Experiment• Evaluation• Limitation and future work• Conclusion

Introduction

• Artificial lighting consumes 26% energy in commercial building

• Daylight harvesting is the approach of using natural sunlight– Reduce lighting energy by up to 40%– Smart glass– Not stable– Caused glare(刺眼 ) and discomfort

Daylight harvesting

• Nature sunlight changed rapidly– 50% existed systems are disables by users– Window transparency changed slowly

• Window change speed v.s. daylight change speed– Glare – Energy waste

• Problem – How to minimize both glare and energy usage

Objective

• SunCast– Prediction natural sunlight level• Fine grained

– Control the window transparency• Adjust in advance

– Purely data-driven approach to create distribution– Instead of making an explicit environment model

Related work

• Predict average sunlight over time period• Weather forecast : only predict cloudiness in

the sky, can not predict the effect of shadow at particular locations

• Control system need more fine-grained information instead of forecast websites

SunCast

• Predicting sunlight values :3 steps– calculates the similarity between the real-time

data stream and historical data traces– uses a regression analysis to map the trends in the

historical traces to more closely match patterns of the current day

– combines the weighted historical traces to predict the distribution of sunlight in the near future

Step1: Similarity

• Difference d between two days data

• Similarity(weight)

Step2: regression

• Linear Regression• Y : current data, X:historical data, find a,b• Y* : predicted data, X:historical data

Step3: creating distribution

• Apply h historical traces • Produce prediction distribution x

Window transparency

• Wt : percentage of window transparency– 0% : closed, 100%:fully open

• Objective function :

• wSpeed: window switching speed• Maximum prediction window len

Prediction and reaction

• Prediction algorithm is ideal for rapid sunlight changes

• Stable sunlight, window transparency control has better performance based on current sunlight condition

• Hybrid scheme : switch smoothly between prediction and reaction according β

• β is light error threshold

Experiment

• Two test bed : residential house and campus• House 4 weeks, campus 12 weeks

Setup

• Hobo data logger• Sensor node– Light– Temperature– Humidity– Sample/min

Other methods• Reactive

– periodically measures the current daylight and sets window transparency to come as close to the target setpoint as possible

• Weather– Select the same cloudiness level from historical data as

• Oracle– Using the actual future light values instead of predicted values

• Optimal– Control window transparency directly

Setpoint= 2000 lux

• Energy : artificial lighting maintains the setpoint• Glare: harvested light above the target setpoint,

Evaluation analysis

• Impact of – Window switching speeds – window orientations– cloudiness levels

Window switching speeds

• Vary from 10~100 min

window orientations

cloudiness levels

Improvement over reactive

• SunCast has the largest effect on lighting stability

• Experiment on four predictive feature window• Light stability improvement over reactive

scheme

Improvement over reactive

Improvement over reactive

limitation

• Unpredictable – Sunrise – Sunset– Trees– Clouds – Nearby buildings– Environmental factors

Future works

• Merge data traces from multiple light sensors• Group estimation • Solar power system• Predict sunlight more opportunities for energy

harvesting

Conclusion

• SunCast– Continuous prediction over time– Distributions of prediction

• Predictive window control scheme– Reducing glare 59%– Saving more energy by artificial lighting

• Applied to other applications– Highway traffic prediction– City pollution levels– Building occupancy

My Question

• How many of historical data are enough?• Weather method v.s. predictive ?