ZEMCH 2015 - International Conference Proceedings | Page 251

This method consequently reflects that there is a relationship between all the firings happening when an activity in occurring as a whole . TA approach computed each firing as an independent sample and based its classification on just only timestep sensor event . Conversely , CDA assumes that a single firing doesn ’ t contain enough information standalone , and has to be considered as a part of a bigger event , which we called ‘ chunk ’. In this case , the features include the sum of all the firings for each sensor , the time the activity started , the time the activity ended and the duration .
In Table II we can observe the different processing of 5 timeslices of data using TA and CDA approaches . While TA represents the data in 5 samples of 4 sensors as inputs and 5 labels as outputs , CDA will group the data into just two samples including the entire sensor readings occurred while one activity was continued in time . CDA also includes start time , end time and duration of the ‘ chunk ’.
Table II : Differences between processing the same information using TA and CDA approaches .
Timeslice Approach
t = 1
t = 2
t = 3
t = 4
t = 5
Sensor1
1
1
0
0
1
Sensor2
1
1
1
0
0
Sensor3
1
0
0
1
1
Sensor4
0
0
0
1
1
Label
4
4
4
3
3
Sample
1
2
3
4
5
Chunk Data Approach
Time Start
3pm
5pm
Time End
4pm
8pm
Duration
20 min
190 min
Sensor1
2
1
Sensor2
3
0
Sensor3
1
2
Sensor4
0
2
Label
4
3
Sample
1
2
4 Experimental results
Dataset 1 has been evaluated using only the timeslice approach using a variation of the timeslice duration to study how the different values affect the final model performance . For the dataset 2 we also included the CDA approach since the initial values of the timeslice approach method were alarmingly lower compared to those reported in the publications by the team who made this dataset publicly available .
4.1 Model Performance Based On Timeslice Variation We evaluated the models to each scenario and observed how the accuracy changed for different timeslices . Despite the loss of information due to the increase of the TS duration , the SVM model showed to be a robust solution since it barely changed its performance while increasing the timeslice . The other models also maintained the values within certain levels , yet showing some
Modelling occupant activity patterns for energy saving in buidings 249