Prediction of electric charge with deep learning models

Abstract

Electricity forecasting is a strategic asset for optimizing the management of power grids, promoting the integration of renewable energies, reducing operating costs, and enhancing the efficient and sustainable use of electrical energy. In order to satisfy demand in real time, energy network operators need to plan the production and distribution of electricity. This efficient operational planning is made possible by electricity load forecasting, which predicts the evolution of demand on different time scales (daily, weekly, and seasonal). The thesis aims to improve intelligent energy management by providing an in-depth study and notable advances in the field of electrical load modeling. New approaches to the study form its foundation, and the application of deep learning techniques to reliably predict changes in electrical load is given particular attention. The research work revolves around the evaluation and improvement of existing models, implementing deep neural network architectures such as convolutional neural networks (CNNs), and models with embedded attention mechanisms (Transformers). The use of complex datasets with rich temporal information content has allowed to capture the temporal dependencies inherent in electrical charge profiles. Firstly, one of the contributions of this research consists of a rigorous methodical approach aimed at acquiring a thorough understanding of the database provided to represent by the Algerian electricity production company. An in-depth statistical study of over 10 years of electricity consumption was undertaken, analyzing in detail the characteristics and trends inherent in the dataset. This provided crucial insights into the distribution of core variables, potential correlations, and temporal patterns, laying the foundations for a comprehensive understanding of the underlying context. Subsequently, the research focused on exploring different machine learning and deep learning architectures. Several approaches were tested and evaluated, with the emphasis on selecting the most appropriate models for a specific challenge. This step involved the exploration of various architectures, such as classical machine learning models like Support Vector Machine and Linear Regression as well as more sophisticated deep learning architectures, including convolutional neural networks (CNNs), Transformers, and other models. The final contribution of this research is the design and implementation of an innovative system capable of dynamically switching to renewable energy sources, thus helping to reduce CO2 emissions during power generation. This system intelligently integrates the use of gas, offering a cleaner, more ecological alternative while protecting human health. This approach is based on the implementation of an automatic switching mechanism between different energy sources, exploiting the advantages of renewable energies when CO2 emissions have reached a critical threshold during electricity production. This strategy aims to minimize the overall carbon emissions footprint of power generation, responding to environmental imperatives and growing concerns about climate change. These contributions converge to create a more efficient energy system. On the socio-economic plan, rigorous planning helps reduce costs, while minimizing energy losses. Enhanced planning gives power producers’ greater control over electricity distribution and promotes more efficient management. In addition, in environmental terms, these advances contribute to a greener future by reducing pollutant emissions, thus working towards an atmosphere less impacted by pollution.