Energy losses in transmission networks are a strategic issue for the efficiency and reliability of electrical systems, particularly in regions with extensive networks that are subject to extreme environmental conditions. In Quebec, a province of nearly 9 million inhabitants located in eastern Canada, the state-owned company Hydro-Quebec (HQ) operates the largest hydroelectric network in North America 1 , characterized by very high voltage lines, high seasonal variability, and complex operating conditions, making accurate prediction of energy losses particularly dificult. Conventional deterministic models, based on physical laws, struggle to represent the complexity of the network in real conditions, particularly in the face of load dynamics and the diversity of data sources. This study is based on a unique dataset covering forty years of transmission loss history, provided through collaboration between the CYNERGIA team and Hydro-Quebec, allowing for the evaluation of machine learning approaches in a real-world context. We compare classical statistical models (ARIMA, Prophet) and ensemble learning algorithms (Random Forest, XGBoost) to analyze their predictive performance and robustness. The data were preprocessed, normalized, and partitioned chronologically to respect temporal dependencies and ensure a realistic evaluation. The results show that ensemble learning-based models outperform classical statistical approaches in capturing complex temporal patterns, providing reliable forecasts under varying operational conditions. This work demonstrates the potential of machine learning to improve the estimation of electrical losses and support the operational planning of large-scale electrical networks.
Kumavi et al. (Thu,) studied this question.