ABSTRACT The growing reliance on fossil fuels for energy generation has raised concerns about their significant contribution to global warming and the associated risks of supply instability. Anaerobic Digestion (AD) within Wastewater Treatment Plants (WWTPs) offers a renewable alternative by producing biogas, while effective operational optimisation requires accurate forecasting of biogas yields under varying conditions. This study addresses this challenge by developing, tuning and evaluating five Deep Learning (DL) architectures for biogas production prediction: one‐dimensional Convolutional Neural Network (1D‐CNN), Long Short‐Term Memory (LSTM), Gated Recurrent Unit (GRU), Transformers and Residual Encoding. Among these, the GRU model demonstrated superior performance, achieving a Root Mean Square Error (RMSE) of 139.1 m 3 /day and a Mean Absolute Error (MAE) of 135.9 m 3 /day. The adaptability of the GRU model to different datasets was examined through Transfer Learning (TL), revealing a clear difference in performance depending on the TL approach used: the retrained model achieved a RMSE of 230.1 m 3 /day and a MAE of 229.9 m 3 /day, whereas the model without retraining exhibited higher errors of 358.7 and 358.8 m 3 /day, respectively. A key contribution of this work lies in its comprehensive Explainable Artificial Intelligence (XAI) analysis, which applied both ante hoc attention mechanisms and post hoc interpretability techniques such as SHAP and LIME. The XAI methods consistently identified biogas production, the study's target variable, as the most influential feature in the model's predictions. Among the remaining features, some changes were observed in their impact on model predictions. Moreover, the study highlighted how TL affects prediction performance and the stability and consistency of feature importance, thereby improving the transparency and trustworthiness of the forecasting models.

Abstract In recent years, we have seen a growing need for energy, which has had environmental consequences through the use of fossil fuels. Some of the sectors of our society make intensive use of energy, as is the case with wastewater treatment plants (WWTPs). Through anaerobic digestion, these infrastructures can produce energy, therefore improve energy efficiency and decrease the environmental footprint. This study aims to design, tune and evaluate a hybrid deep learning (DL) model, called incremental dual path network (IDPN), to forecast energy production in an anaerobic reactor for the next two days. The hybrid model’s performance was compared against five DL models conceived: long short-term memory (LSTMs), multi-layer perception (MLP), gated recurrent units (GRUs), Transformers and convolutional neural networks (CNNs). Furthermore, two data processing strategies were applied due to system failures and missing values. Four model evaluation metrics and the obtained results show that the hybrid model, which combines LSTMs and CNNs, presented the best performance in both approaches of data processing, with the best candidate model presenting a mean absolute error (MAE) of 312.1 kWh, root mean squared error (RMSE) of 341.6 kWh, mean absolute percentage error (MAPE) of 15.9% and R $$^{2}$$ 2 of 0.95. Following this, an ablation study was conducted, demonstrating that across several variations, the baseline IDPN consistently achieved the best results. Moreover, in both approaches, the removal or modification of the CNN led to a severe decline in performance, surpassing the impact of altering the LSTM, reinforcing its importance in the model’s architecture.

Environmental sustainability is one of the biggest concerns nowadays. With increasingly latent negative impacts, it is substantiated that future generations may be compromised. The research here presented addresses this topic, focusing on air quality and atmospheric pollution, in particular the Ultraviolet index and Carbon Monoxide air concentration, as well as water issues regarding Wastewater Treatment Plants, in particular the pH of water. A set of Machine Learning regressors and classifiers are conceived, tuned, and evaluated in regard to their ability to forecast several parameters of interest. The experimented models include Decision Trees, Random Forests, Multilayer Perceptrons, and Long Short-Term Memory networks. The obtained results assert the strong ability of LSTMs to forecast air pollutants, with all models presenting similar results when the subject was the pH of water.