ESTG - Mestrado em Ciência de Dados
URI permanente para esta coleção:
Navegar
Percorrer ESTG - Mestrado em Ciência de Dados por assunto "Análise de séries temporais"
A mostrar 1 - 1 de 1
Resultados por página
Opções de ordenação
- Day-Ahead Energy Consumption Forecasting in Academic Campi with Deep Learning ModelsPublication . Horta, Simão Matos Conceição; Grilo, Carlos Fernando de Almeida; Távora, Luís Miguel de Oliveira Pegado de Noronha e; Sousa, João Miguel Charrua de; Marques, Pedro José FrancoElectricity is essential in today’s society, with global demand projected to increase 50% by 2050. However, there are significant inefficiencies in power systems throughout production, transmission and distribution. Therefore, to ensure an efficient energy supply it is important to accurately forecast the energy consumption since it allows to deliver only the necessary resources. Advances in data science have provided essential tools to address these challenges by improving the reliability of energy consumption forecasts. Based on these principles, the current project has developed a daily predictive model with a one-day horizon to support energy management decisions, with the goal to optimize the energy efficiency at Campus 2 of the Polytechnic Institute of Leiria (IPL). For this purpose, the historical energy consumption on the campus was used to optimize several models through the NeuralForecast framework. Models were developed using only the endogenous consumption variable, as well as models incorporating the exogenous variables Day Type (which distinguishes between weekdays, Saturdays, and Sundays/holidays) and Academic Calendar (which distinguishes between classes, evaluations, school breaks, and vacation periods). The performance of each model was then evaluated using the Mean Absolute Percentage Error (MAPE) and the computational performance of the model was measured by the Total Parameter Count (TPC). The results show that the best-performing model was N-BEATSx trained on both exogenous variables, achieving a MAPE of 4.92% and 806k total number of parameters. However, the model that best balances complexity and performance is the MultiLayer Perceptron (MLP) with both exogenous variables, with only around 51k parameters and a MAPE of 5.57%. In summary, this study developed robust neural networks for energy consumption forecasting, providing both theoretical and practical advances to support decision-making aimed at optimizing energy efficiency.