Machine learning vs. ADM1: Reliable biogas prediction with minimal data requirements in full-scale plants

TL;DR

This study compares machine learning models and a simplified biochemical model for predicting biogas production in full-scale plants, showing they can be reliable with minimal data.

Contribution

The study introduces a practical comparison of simplified ADM1 and machine learning models for biogas prediction with minimal data requirements.

Findings

All models achieved reliable performance with Nash-Sutcliffe efficiencies above 0.78.

LSTM provided high accuracy using only daily feedstock mass, reducing chemical analysis needs.

LSTM training time was 141 times longer than ADM1, highlighting a computational trade-off.

Abstract

Anaerobic digestion harnesses microbial processes to convert organic wastes into renewable biogas, offering a sustainable pathway for energy production. In agricultural settings, biogas plants often co-digest livestock manure with crop residues, yet seasonal variations in feedstock quality introduce fluctuations that challenge process stability and yield optimization. Mechanistic models such as the Anaerobic Digestion Model No. 1 (ADM1) provide detailed biochemical simulations but require extensive substrate characterization, limiting their practicality for full-scale operations. Here we show that a simplified ADM1, alongside machine learning approaches—random forest and long short-term memory (LSTM) networks—achieves comparable accuracy in predicting daily biogas and methane production from a full-scale plant over 2023–2024. All models yielded Nash-Sutcliffe efficiencies above 0.78, with random forest excelling when incorporating feedstock quantities and maize silage volatile solids. While LSTM proved effective even with minimal inputs, it incurred a training time 141 times greater than ADM1, highlighting critical trade-offs in computational efficiency. These findings advance hybrid modelling strategies for real-time monitoring, enabling operators to balance predictive precision with data requirements to enhance renewable energy integration and agricultural sustainability.

Image 1

•We compared a simplified ADM1 with RF and LSTM models to simulate biogas production at a full-scale plant.•All three models achieved reliable predictive performance with Nash-Sutcliffe efficiency values above 0.78.•LSTM enabled high-accuracy forecasting using only daily feedstock mass, reducing the need for chemical analyses.•Maize silage was identified as the most critical input feature influencing biogas and methane production across models.•LSTM training time was 141 times higher than ADM1, revealing a significant computational-accuracy trade-off.

We compared a simplified ADM1 with RF and LSTM models to simulate biogas production at a full-scale plant.

All three models achieved reliable predictive performance with Nash-Sutcliffe efficiency values above 0.78.

LSTM enabled high-accuracy forecasting using only daily feedstock mass, reducing the need for chemical analyses.

Maize silage was identified as the most critical input feature influencing biogas and methane production across models.

LSTM training time was 141 times higher than ADM1, revealing a significant computational-accuracy trade-off.

Figures