Table of Contents
The February 2017 an energy consumption foretasting challenge was launched on the website driven data (Power Laws: Forecasting Energy Consumption). We use the energy consumption data provided for this challenge to build a forecast algorithm based on boosted trees and auto regressive model.
1 Download the data
We use data provided by Schneider Electric for a forecasting challenge here.
2 Prepare the data
The challenge provide consumption and outside temperature for several building.
The following graph represent the times series for the consumption and temperature for the site 8.
3 Model1: xgboost
We firstly we build a model without tuning the parameters.
This first model get an average prediction error of 3877 kW, which represent around 10 percent of the average consumption.
| root mean average error | mean absolute error | mean consumption | median consumption | |
|---|---|---|---|---|
| 1 | 5820.37 | 3877.06 | 33487.80 | 27928.97 |
The following graph show the forecast and actual consumption.
4 INPROGRESS Model2: xgboost + arima
We plot the seasonality
Let's say that the time series is stationary.