Progress Report Presenter : Min-chia Chang Advisor : Prof. Jane Hsu Date : 2011 - 03 - 01.

Progress ReportPresenter : Min-chia ChangAdvisor : Prof. Jane HsuDate : 2011 - 03 - 01

OutlinePrediction of AC State (revised)Definition of AC Waste AnalysisResult of AC Waste AnalysisThesis – Chapter 2

2011/03/01 2NTU CSIE iAgent Lab



Data- label: (OFF, ONno green, ONgreen) define⟶ y={0,1,2}

- OFF : close- ONno green : Tindoor < TuserSetting , valve = OFF- ONgreen : Tindoor > TuserSetting , valve = ON

- feature : define⟶ ｘ , which is a vector- (Tindoor, Hindoor, Tvent, Hvent, Toutdoor, Houtdoor)- Context data


Place- R104_1, R104_2, R104_3, R104_4 (classroom) - R204_1, R204_2, R204_3, R204_4, R204_5, R204_6 (computer classroom)- R318_1(professor room)- R324_1, R324_2 (seminar room)- R336_1, R336_2 (lab)- R439_1 (seminar room)- R521_1, R521_2 (seminar room)


zone

note: each zone contains only one AC controller

Dataset D={( ｘ n,yn)}, where n=1 to N

- each minute of labeled period (original : intersection of vent and indoor) - labeled by camera (original : controlled on purpose by duck) - size = 77,439


DatasetTime- R336 : 2010-12-18 ~ 2011-01-06- R204 : 2011-01-06 ~ 2011-01-17- R324 : 2011-01-20 ~ 2011-01-30


Execution environment Weka Function: SVM

- Kernel: RBF Cross Validation: 3-fold

- In each iteration :


DatasetTrainingDataTestingData note : NEVER use testing data before

you predict.

AccuracyBaseline- ｘ =(Tvent)

- accuracy = 72.66%- time =

- ｘ =(Tindoor , Hindoor , Tvent , Hvent , Toutdoor , Houtdoor)- accuracy = 93.21% - time =

- ｘ =(Tindoor , Hindoor , Tvent , Hvent , Toutdoor , Houtdoor , ……)


Bagging (bootstrap aggregation)


DatasetTrainingDataTestingData

K=?, S=?• K fixed - If S decreases, then time decreases.

• S fixed - If K increases, then the result of the vote is more convinced.

……

K training datasize = Ssize = S size = S

re-sampling

Bagging


……

K training datasize = Ssize = S size = S

TrainingData re-samplingy=0 y=1 y=2

S/3 S/3 S/3

• K fixed - If S decreases, then time decreases.

K=?, S=?


x=6dimensions K=1 K=10 K=30 K=50

S=300 63.14%16s

74.78%2m4s

80.4%6m10s

79.69%10m18s

S=1500 85.37%30s

87.90%4m20s

88.96%13m16s

89.20%22m8s

S=3000 88.27%47s

90.48%7m7s

91.03%21m36s

91.08%38m07s

S=15000 91.93%6m29s

92.52%31m11s

S=30000 92.55%19m18s

baseline:93.21%45m12s

time && accuracy => trade-off

Generate featurex=(Tindoor , Hindoor , Tvent , Hvent , Toutdoor , Houtdoor , some context data)


context data dimensions valuechilled water host 3 {0,1}chilled water temperature 1 integerrotation speed of pump 1 floatnew or old (building) 2 {0,1}floor 5 {0,1}room type 6 {0,1}area 1 floatday of the week 7 {0,1}weekday or weekend 2 {0,1}semester or vacation 2 {0,1}hour of the day 24 {0,1}


x=60dimensions

K=1 K=5 K=10 K=30 K=50

S=300 58.32%18s

78.11%1m23s

76.03%2m41s

79.38%7m31s

83.26%13m12s

S=1500 84.82%39s

89.12%3m12s

88.84%6m18s

91.27%18m12s

91.46%31m42s

S=3000 90.41%1m34s

92.33%7m23s

93.21%13m18s

93.60%40m14s

93.59%1h03m50s

S=15000 94.72%14m30s

95.38%1h12m55s

S=30000 95.43%39m40s

baseline:96.11%1h32m50s

K=?, S=?

Process the missing value missing value : Tindoor , Hindoor , Tvent , Hvent processing method :

method 1 : encoding e.g. : (?, ?, 15.2, ?) => (0, ?, 0, ?, 1, 15.2, 0, ?)

method 2 : interpolation 　 (linear) e.g. : 2011-01-31 23:50 : (20, 45, 10, 70) 2011-01-31 23:51 : (?, 45.2, 9.9, 70.2)…… 2011-02-01 00:00 : (20.1,45.5,10.2,69.2)=> ? = 20.01

method 3 : encoding + interpolation


Result


baseline bagging(K=30, S=3000)(Tvent) 72.66%11m34s

70.43%

6 dimentions 93.21%45m12s

91.03%21m36s+ generate features 96.11%

1h32m50s93.40%40m14s+ missing value (encode) 96.07%

1h13m43s93.43%

+ missing value(interpolation) 99.79%54m41s

98.58%

+ missing value(encode + interpolation) 99.60%1h33m40s

96.10%

+ normalize 97.55% 94.05%



Problem definition : energy(AC) waste analysis Input :

- state of motion sensor, mn={no, yes} - AC information

Output :- proportion of AC waste


System overview


motion sensorstate

AC statepredictorthermal comfort questionnaire

Tindoor Toutdoorthermal comfort equationand offset

AC information input

inputoutputproportion of AC wasteAC wasteanalysis

Condition of AC waste state of motion sensormn state of ACyn Tindoor ? TcomfortableRange waste or notN 0 higher NN 0 among NN 0 lower NN 1 higher Y N 1 among YN 1 lower YN 2 higher YN 2 among YN 2 lower YY 0 higher NY 0 among NY 0 lower NY 1 higher N?? Y 1 among NY 1 Lower N??Y 2 Higher Y Y 2 Among NY 2 lower Y




Currently, I encountered a problem here ……


Problem


AC statepredictorAC informationcollectorDataset Anothertestingdata

• Some place we did not have in the dataset yet.• Some patterns of the feature’s combination (dependent on time) in another testing data haven’t seen in the dataset .

error prediction

2011-01-01~2011-01-31all2010-12-17~2011-01-30R336 R324R204

Current condition feature :

- Tindoor, Hvent- new or old (building)- floor- room type- area

bagging - S=30000- K=5


Condition of AC waste waste situation 1. mn = no and (yn = 1 or yn = 2) 2. mn = yes and yn = 2 and Tindoor < TcomfortableRange 3. mn = yes and yn = 2 and Tindoor > TcomfortableRange


Proportion of AC waste waste situation 1. mn = no and (yn = 1 or yn = 2) 2. mn = yes and yn = 2 and Tindoor < TcomfortableRange 3. mn = yes and yn = 2 and Tindoor > TcomfortableRange


place mn=no mn=yes yn=0 yn=1 yn=2 waste 1 waste 2 waste 3336_2 58% 42% 26% 61% 13% 36.7% 9.5% 0%

204_1 47% 53% 23% 70% 7% 33.8% 4.0% 0%

204_2 43% 57% 19% 39% 42% 35.8% 17.8% 0%

204_3 48% 52% 53% 44% 3% 23.9% 16.7% 0%

204_4 57% 43% 53% 41% 6% 28.2% 4.0% 0%

204_5 65% 35% 15% 13% 72% 57.2% 20.8% 0%

204_6 65% 35% 55% 41% 5% 31.9% 2.3% 0%

204 33% 67% 7% - - 31.0% - -

2011.01



backup


3. 問題定義 - 第一部分

2011/01/0829

NTU CSIE iAgent Lab

空調狀態預測機

空調資訊收集器

出風口溫度收集器

室內溫度收集器

室外溫度收集器

舒適比例分析中心

熱舒適溫度公式

空調設定收集器

有 / 無人資訊收集器

Progress Report Presenter : Min-chia Chang Advisor : Prof. Jane Hsu Date : 2011 - 03 - 01.

Documents

ntu csie iagent labx

ntu csie iagent labdatalabel

ntu csie iagent labplacer104

ntu csie iagent lab6

ntu csie iagent labk

ntu csie iagent lab

s decreases

s fixed