Author(s): Medeiros, Gonçalo ; Marcondes, Francisco Supino ; Oliveira, P. ; Machado, José Manuel ; Novais, Paulo
Date: 2025
Persistent ID: https://hdl.handle.net/1822/95114
Origin: RepositóriUM - Universidade do Minho
Subject(s): Biodigestion; Deep learning; Electricity production; Energy forecasting; Time series forecasting; Transformer; Wastewater treatment plants
Global energy demand has been growing over the last decades, having a significant impact on the environment. Alternatives that can mitigate these effects are crucial for the future of our planet. Wastewater Treatment Plants (WWTPs) are a vital infrastructure to manage residual waters, presenting an opportunity for energy production from the released biogas during the anaerobic digestion phase. The present study uses a multivariate recursive forecasting approach to evaluate the performance of Transformer-based models in forecasting electricity production in this context. Transformer-based candidate models were developed, and their hyperparameters were tuned using a grid search. The best Transformer candidate achieved the second-best Root Mean Square Error (RMSE) value of 359.4 kWh, outperforming the Gated Recurrent Unit (GRU) by 9%, although the Long Short-Term Memory (LSTM) model performed the best.
FCT - Fundação para a Ciência e a Tecnologia(2022.06822)
