Transformer-based multi-step time series forecasting of methane yield in full-scale anaerobic digestion

Abstract

Accurate multi-step forecasting of methane yield is crucial for optimizing feedstock planning, detecting gradual process changes, and mitigating operational risks in full-scale anaerobic digestion (AD). However, previous studies have primarily focused on one-step prediction repeatedly, limiting their ability to capture long-term trends and process dynamics. To address this challenge, a Transformer-based model was employed and modified to handle the characteristics of AD datasets. Unlike autoregressive models, the Transformer utilizes parallel computation and a self-attention mechanism, enabling it to retain long-term dependencies, making it well-suited for multi-step forecasting tasks. The dataset from full-scale AD system was analyzed to assess stationarity and temporal dependencies, and the proposed model was compared with autoregressive sequential models. Results demonstrated that the proposed model outperformed others by up to 67 %, particularly in multi-step ahead forecasting. Data analysis confirmed that fluctuation in feedstock organic loading and digester temperature seasonality contributed to non-stationarity. Additionally, the model analysis revealed that the multi-head self-attention mechanism effectively captured different aspects of the time series, improving long-term dependency retention. Generalization and robustness of the model was confirmed by validating model on the three additional full-scale AD datasets. These findings highlight the potential of the Transformer model for enhancing methane yield forecasting in anaerobic digestion systems.