This repository details the winning solution established by our team LPSM204 for the data challenge organized prior the 2018 JDS conference by the Young Statisticians and Probabilists group.

The LPSM204 team is composed of three PhD Student from Sorbonne University, LPSM:

• Nazih Benoumechiara,
• Nicolas Meyer,
• Taïeb Touati,

and sharing the same office (204).

Short-term forecasting is a major issue for the main french electricity utility "Electricité de France" (EDF). The purpose of forecasting is to predict future scenario and to adapt supply to demand. Energy producers mainly consider electrical consumption history to predict future demand. Another important information is the weather, which can greatly influence the consumption. The challenge proposed here is about forecasting the electricity consumption of a small island based on a one year history of electricity consumption and weather condition.

The method established is based on two approaches :

• a time-series forecasting using the electricity consumption,
• a machine learning approach using the electricity consumption and the weather conditions.

The final result is a smart mix between the prediction of both approaches.

The notebook of the solution is available here.

Specific python packages are necessary if you're willing to test proposed methodology. In addition to the very casual ones (e.g., scipy, pandas, seaborn and others), you'll need:

• statsmodels (for the ARMA modeling),
• lightgbm (for Gradient Boosting trees),
• hyperopt (for hyperparameter tuning),
• holidays (for calendar holidays),

All can be easily installed from Anaconda or pip.

• French calendar of school holidays available here.

A talk will be held Tuesday 29 of May at the 2018 JDS conference to shortly present our solution. The schedule is available here and the slides are available here.

Following this competition, a paper should be submitted at the CSBIGS (Case Studies in Business, Industry and Government Statistics) journal in collaboration with the 2nd winning team.

Our thanks go to the organizer of the challenge and surely to our respective PhD supervisors who let us take some time to work on the challenge.