Electricity Demand Forecasting during COVID-19 Lockdowns

The coronavirus disease 2019 (COVID-19) pandemic has caused significant changes in electricity consumption patterns due to the strict lockdown measures implemented by many governments worldwide.

Project Description

This project aims to reproduce a solution developed in this paper to the issue of poor performances exhibited by traditional electricity load forecasting models since the beginning of the pandemic of COVID-19. These models are trained on historical data and rely on calendar or meteorological information.

We are only focusing on fine-tuning which allows the model to quickly adapt to new consumption patterns without requiring exogenous information.

The developed models are applied to forecast the electricity demand during the French lockdown period, and expert aggregation is used to leverage the specificities of each prediction and enhance results even further.

Data

The data used in this project is publicly available and can be found in the data directory.

• Température : temperature in Celcius
• Temp95 and Temp99 : exponentially smoothed temperatures of factors .95 and .99
• TempMin99 and TempMax99 : minimal and maximal value of Temp99 at the current day
• Consommation : electricity consumption in MW
• Consommation1 and Consommation7 : consumptions of the day before and the week before
• DateN : number of the day since the beginning of the dataset
• TimeOfYear : time of year (0 = 1st of January at 00:00, 1 = 31st of December at 23:30)
• DayType : categorical variable indicating the type of the day (0 = Monday, 6 = Sunday)
• DLS : binary variable indicating whether it is summertime or not

Getting started

We chose to develop this project in both Python and R to get the best out of the already developed libraries.

Prerequisites

Before you start, make sure you have the following packages installed on Python :

• numpy
• pandas
• torch
• tqdm

and on R :

• opera
• mgcv
• caret
• riem

Cloning the repository

To clone the repository, run the following command in your terminal

git clone [email protected]:Exion35/load-forecasting.git

or

git clone https://github.com/Exion35/load-forecasting.git

Navigating through the repository

1. Get the Italian weather data with get_it_weather.ipynb
2. Process both French and Italian data with preprocessing.ipynb
3. Build the experts (GAM, GAM Saturday, GBM) with build_experts.ipynb
4. Fine-tune the GAM and display the results with fine_tuning_gam.ipynb
5. Aggregate the experts with aggregate_experts.ipynb (you can come back to fine_tuning_gam.ipynb to display the aggregation plot)

Results

Numerical Performance In MAPE (%) and RMSE (MW).

Method	Test 1	Test 2
GAM	5.40%, 3076 MW	3.77%, 2030 MW
GBM	6.34%, 3483 MW	5.04%, 2607 MW
Fine-tuned	3.96%, 2417 MW	3.78%, 2024 MW
GAM $\delta$	10.96%, 6063 MW (!)	4.40%, 2313 MW
GAM $\delta$ - Fine-tuned	-	-
GAM Saturday	4.02%, 2520 MW	5.78%, 3227 MW
Aggregation without GAM Saturday	2.67%, 1625 MW	3.35%, 1891 MW
Aggregation with GAM Saturday	2.64%, 1553 MW	3.06%, 1727 MW

References

• D. Obst, J. de Vilmarest and Y. Goude, "Adaptive Methods for Short-Term Electricity Load Forecasting During COVID-19 Lockdown in France," in IEEE Transactions on Power Systems, vol. 36, no. 5, pp. 4754-4763, Sept. 2021, doi: 10.1109/TPWRS.2021.3067551.