GECCO 2013 Industrial Competition Computer Engineering Lab, School of Electrical and IT Engineering Farzad Noorian.

GECCO 2013 Industrial Competition

Computer Engineering Lab, School of Electrical and IT EngineeringFarzad Noorian

2

GECCO 2013

› Genetic and Evolutionary Computation Conference

- Organized by ACM SIGEVO

› GECCO Industrial challenge:

- http://www.spotseven.de/gecco-challenge/

- sponsored by GreenPocket GmbH

3

Introduction

› About the Competition

› Pre-processing

› Features

› Training and Cross-validation

› Results

4

The Competition

› Real room climate time series

- Outside temperature as an additional input

- Irregular time-series

- Very noisy

5

Preprocessing

› From original data

6

Preprocessing

› Outliers were removed

7

Preprocessing

› A weighted moving average with a small window

8

Preprocessing

› Regularized using linear approximation

9

Preprocessing

› Only values at hourly boundaries were used.

10

Features

› Only the outside temperature was given.

› No outside humidity.

› Human perception based on both.

11

Features

› Publicly available data from Weather Underground™ for Köln

- Temperature

- Humidity

- Dew Point

12

Features for Temperature Forecasting

› Weekday seasonality → Only weekdays used

- Seasonality removed only from indoor temperature

› A window of last n hours room temperatures

› A window of previous m and next m dew points from Wunderground

13

Features for Humidity Forecasting

› A window of last n hours

› m previous and m next external humidity from Wunderground

- Open, Low, High and Close of that days humidity

› No seasonality or data filtering

14

Learner

› Support Vector Machines

- With Radial Kernel

› Advantages of SVMs

- Efficiently trained

- Unique global optima

15

Cross-validation

› Using R package caret

› Cross validation for features and parameters

- Using from a 4-day window to 15-day window to train

- Validating using next 3 available days

› Final training on all data

16

Final Results

› Prediction in hourly, linearly approximated to 10 minutes

17

Questions?

› Feel free to email: [email protected]