Prédiction de la consommation électrique par un prototype de Machine Learning et Deep Learning : Comparaison entre les Méthodes uni-variées et multivariées, étude de l’impact des données météorologiques et la génération des énergies renouvelables.

Abstract #

Electric energy is a main cog in the economic circuit in Morocco, in which all activity is both

cause and consequence of development, especially with the development of Smart Grid hence

the interest of monitoring systems for control, good planning to forecast electrical loads and

reduction of excessive electricity generation. These kinds of systems also allow for realistic

energy price estimation, scheduling management and future capacity planning.

It is in this context that our end-of-study project, carried out within the scientific institute of

Rabat, has objective to design and implement a supervision system, which base on the forecast

of the consumption of electrical energy. To do this, we base on a comparison between several

methods used to forecast energy in order to understand the impact of meteorological data on

their forecast performance on the one hand and an analysis of the relationship between the total

load of electricity (consumption) and the generation of renewable solar and wind energy. We

consider both time series and linear statistical regression models as well as linear and nonlinear

machine learning and deep learning methods. We will also create hybrid methods by combining

these different models.

To achieve this work, we first started by specifying the body’s needs at the end of the monitoring

system for the prediction of electricity consumption, then we proceeded to its design and

implementation. In order to improve the performance of the prediction, we have developed a

benchmark between several models of time series, machine learning, and deep learning to deal

with the univariate and multivariate case. We have also implemented hybrid models such as

ARIMA&SVR, ARIMA&LSTM, and CNN&LSTM.

The results of the evaluation show that deep learning models and hybrid models perform

better for both modalities. The CNN&LSTM model outperforms all models developed, making

it the best model in our study. The system implementation with Python language, known for

both its simplicity and its robustness and which relies on more solid libraries for data

visualization and implementation of time series models, Machine Learning and Deep Learning

for Forecasting.

Keywords: Monitoring system, Prediction, Univariate, Deep Learning, Hybrid models,

electricity consumption, Smart Grid, economic, renewable Énergies, Multivariate, Machine

Learning, Python.