Similaire à Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment - KDIR 2012
Design & implementation of phasor data concentrator compliant to ieee c37.118...Nitesh Pandit
Similaire à Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment - KDIR 2012 (20)
Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment - KDIR 2012
1. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Some empirical evaluations of a temperature
forecasting module based on Artificial Neural
Networks for a domotic home environment
F. Zamora-Mart´nez, P. Romeu, J. Pardo, D. Tormo
ı
Embedded Systems and Artificial Intelligence group
´
Departamento de ciencias f´sicas, matematicas y de la computacion
ı ´
˜ ´
Escuela Superior de Ensenanzas Tecnicas (ESET)
Universidad CEU Cardenal Herrera, 46115 Alfara del Patriarca, Valencia (Spain)
KDIR – October 6, 2012
2. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Index
1 Introduction
2 Domotic home environment setup
3 Data preprocessing
4 Neural Network description
5 Experimentation
6 Conclusions and future work
3. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Introduction
Index
1 Introduction
2 Domotic home environment setup
3 Data preprocessing
4 Neural Network description
5 Experimentation
6 Conclusions and future work
4. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Introduction
SMLhouse
5. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Introduction
Introduction and motivation
SMLhouse is a domotic solar house project presented at the
SolarDecathlon 2010.
The Computer Aided Energy Saving (CAES) system is being
developed to decrease power consumption, increasing energy
efficiency, keeping comfort parameters.
Indoor temperature is related with comfort and power
consumption.
Artificial Neural Networks (ANNs) are a powerful tool for pattern
classification and forecasting.
This work is an empirical experimentation to set the best ANN
parameters in a real forecasting task.
6. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Domotic home environment setup
Index
1 Introduction
2 Domotic home environment setup
3 Data preprocessing
4 Neural Network description
5 Experimentation
6 Conclusions and future work
7. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Domotic home environment setup
Hardware architecture
Lights,
roller-shutters,
HVAC, . . .
Temperature, air
⇒ ⇒ Ethernet
quality, humidity,
...
Light Switches,
dimmers, . . .
8. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Domotic home environment setup
Software architecture
First layer: data is acquired from the KNX bus by iOS interface ANN Modules
the Open Home Automation Bus (openHAB).
Persistence
Second layer: data persistence module collect (REST interface)
sensor and actuator values every minute. KNX-IP Bridge → openHAB ⇐
9. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Domotic home environment setup
Software architecture
First layer: data is acquired from the KNX bus by iOS interface ANN Modules
the Open Home Automation Bus (openHAB).
Persistence ⇐
Second layer: data persistence module collect (REST interface)
sensor and actuator values every minute. KNX-IP Bridge → openHAB
Timestamp Name Value
... ... ...
2011-03-30 10:51 Dinning Room Temperature 30.0
2011-03-30 10:52 Dinning Room Humidity 52.0
... ... ...
10. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Domotic home environment setup
Software architecture
iOS interface ANN Modules ⇐
Third layer: two applications that could
communicate between themselves. A native iOS Persistence
application for manual control. A couple of
(REST interface)
modules that can actuate autonomously. KNX-IP Bridge → openHAB
11. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Domotic home environment setup
Software architecture
iOS interface ANN Modules ⇐
Third layer: two applications that could
communicate between themselves. A native iOS Persistence
application for manual control. A couple of
(REST interface)
modules that can actuate autonomously. KNX-IP Bridge → openHAB
12. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Data preprocessing
Index
1 Introduction
2 Domotic home environment setup
3 Data preprocessing
4 Neural Network description
5 Experimentation
6 Conclusions and future work
13. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Data preprocessing
Data details
Acquisition
The data temperature signal is a sequence s1 s2 . . . sN of values,
sampled with a period of 1 minute.
Preprocessing
1 Low-pass filter (mean with 5 samples): s1 s2 . . . sN where
si = (si + si−1 + si−2 + si−3 + si−4 )/5
2 Data normalized subtracting mean and dividing by the standard
deviation: s1 s2 . . . sN where
si − s
¯
si =
σ(s )
14. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Data preprocessing
Dataset size
Partition Number of patterns Days
Training 30 240 21
Validation 10 080 7
Test 10 080 7
Validation partition is sequential with training partition.
Test partition is one week ahead from last validation point.
Mean and standard deviation normalization values were
computed over the training plus validation.
15. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Data preprocessing
Plot of the dinning room temperature for validation partition
26
25
24
23
22
21
ºC
20
19
18
17
16
15
0 2000 4000 6000 8000 10000
Time (minutes)
16. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Neural Network description
Index
1 Introduction
2 Domotic home environment setup
3 Data preprocessing
4 Neural Network description
5 Experimentation
6 Conclusions and future work
17. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Neural Network description
Neural Network description
At time step i:
18. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Neural Network description
Neural Network description
At time step i:
the ANN input receives:
the hour component of the current time (locally encoded) and
a window of the previous temperature values (α is step, and M is
number of steps):
si si−α si−2α . . . si−(M−1)·α
19. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Neural Network description
Neural Network description
At time step i:
and computes a window with the next predicted temperature
values (L is forecast horizon):
si+1 si+2 si+3 . . . si+L
Known as multi-step-ahead direct forecasting.
20. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Neural Network description
Multi-step-ahead forecasting approaches
Multi-step-ahead iterative forecasting was very extended in
literature. Only one future value is predicted and reused to predict
iteratively the whole window. Better for small future horizons.
Multi-step-ahead direct forecasting approach is based on the
computation of the future window in one step. Better for large
future horizons.
21. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Neural Network description
Training details
Error back-propagation algorithm with momentum term.
The ANN learn to map predicted output values (oi ) with
corresponding true values ( pi ),
minimizing the MSE function
MSE
1
E = ∑ (oi − pi )2
2L i
22. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Neural Network description
Training details
Error back-propagation algorithm with momentum term.
The ANN learn to map predicted output values (oi ) with
corresponding true values ( pi ),
minimizing the MSE function, adding weight decay L2
regularization
MSE weight decay
1 w2
E = ∑ (oi − pi )2 + ε ∑
2L i w∈{W HO W IH }
2
23. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Experimentation
Index
1 Introduction
2 Domotic home environment setup
3 Data preprocessing
4 Neural Network description
5 Experimentation
6 Conclusions and future work
24. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Experimentation
Experimentation: training parameters
An exhaustive exploration leads to this parameters:
learning rate of 0.001,
momentum of 0.0005,
weight decay of 1 × 10−7 ,
input window step of α = 2,
input window size of M = 30,
one hidden layer with 8 neurons and logistic activation function.
output window horizon L experiments will be shown in detail.
The ANN best topology was (15 + 24) × 8 × L.
25. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Experimentation
Experimentation (II): evaluation
ANNs were trained modifying the output window horizon focusing
results only on L = 60, 120, 180 (denoted by NN–060, NN–120,
NN–180).
Evaluation measures
Mean Absolute Error (MAE):
1
MAE = |pi − pi |
N∑i
Normalized Root Mean Square Error (NRMSE):
∑ (pi − pi )2
i
NRMSE =
∑ ( pi − pi )2
¯
i
26. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Experimentation
Experimentation (III): forecasting mean temperatures
In order to focus the temperature forecasting measured errors on
their future use on an automatic control system, we will compute
the mean (or max/min) temperature forecasted by the model in
the selected forecasting window.
Then we could measure the MAE value between this mean and
the ground truth mean on the same window.
27. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Experimentation
Experimentation (IV): individual models plot
0.14
NN−060
NN−120
NN−180
0.12
0.10
0.08
MAE
0.06
0.04
0.02
0.00
20 40 60 80 100 120 140 160 180
Window upper bound
Plot of the MAE error computed over the mean of forecasting windows
0–20, 0–40, 0–60, 0–80, . . . , 0–180, using ANN models trained with
L = 60, 120, 180.
28. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Experimentation
Experimentation (V): ensemble of models
An ensemble of NN–060 and NN–180 model would ensure good
performance in all cases.
A linear combination of ANN outputs was performed, following:
NN–060 NN–180
os + ol
i i
, for 0 ≤ i < 60 ;
oi = 2
NN–180
l
oi , for 60 ≤ i < 180 .
29. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Experimentation
Experimentation (VI): ensemble vs individual models plot
0.14 0.14
NN−060 NN−060
NN−120 NN−120
NN−180 NN−MIX
0.12 0.12
0.10 0.10
0.08 0.08
MAE
MAE
0.06 0.06
0.04 0.04
0.02 0.02
0.00 0.00
20 40 60 80 100 120 140 160 180 20 40 60 80 100 120 140 160 180
Window upper bound Window upper bound
Plot of the MAE error computed over the mean of forecasting windows
0–20, 0–40, 0–60, 0–80, . . . , 0–180, using NN–060, NN–120, and
NN–MIX models (right).
30. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Experimentation
Experimentation (VII): validation and test set final results
NN–MIX model results for validation set
Window Min Max Mean
0–60 0.029/0.050 0.047/0.061 0.027/0.043
60–120 0.068/0.115 0.099/0.135 0.079/0.122
120–180 0.129/0.214 0.165/0.233 0.143/0.223
NN–MIX model results for test set
Window Min Max Mean
0–60 0.139/0.188 0.173/0.254 0.150/0.205
60–120 0.255/0.371 0.239/0.360 0.270/0.394
120–180 0.334/0.539 0.381/0.603 0.352/0.566
NRMSE/MAE on minimum, maximum, and mean temperature
forecasting for validation and test sets.
31. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Experimentation
Experimentation (VIII): validation set forecasting plot
26
NN−MIX
25 Ground Truth
24
23
22
21
ºC
20
19
18
17
16
15
0 2000 4000 6000 8000 10000
Time (minutes)
Plot of validation set forecasted mean temperature versus ground truth
mean temperature using a forecasting window of 0–60 with NN–MIX
model.
32. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Experimentation
Experimentation (IX): test set forecasting plot
30
NN−MIX
Ground Truth
28
26
ºC
24
22
20
18
0 2000 4000 6000 8000 10000
Time (minutes)
Plot of test set forecasted mean temperature versus ground truth mean
temperature using a forecasting window of 0–60 with NN–MIX model.
33. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Conclusions and future work
Index
1 Introduction
2 Domotic home environment setup
3 Data preprocessing
4 Neural Network description
5 Experimentation
6 Conclusions and future work
34. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Conclusions and future work
Conclusions
A real hardware/software architecture was introduced for domotic
home environments: SMLhouse.
Preliminary data was used for model testing and validation.
Monitoring and manual control systems are running.
Intelligent control modules are being developed: dinning room
temperature forecast module.
Promising results: little MAE error was achieved (0.6◦ C for three
hours forecast).
It motivates the integration of this ideas into an automatic control
system.
35. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Conclusions and future work
Future work
Covariate forecasting.
Extend forecasting module to air quality, humidity, power
consumption, insolation, . . .
Introduce confidence on the prediction, based on prediction
intervals.
Replace feedforward ANN with a recurrent neural network:
Long-Short Term Memory.
36. Some empirical evaluations of a temperature forecasting module based on Artificial Neural Networks for a domotic home environment
Conclusions and future work
Questions?
Thanks for your attention!