Probabilistic Prediction of Anaerobic Reactor Performance Using Bayesian Long Short-Term Memory Artificial Recurrent Neural Network Model

Abstract

Engineering structures are designed and managed for safe operations and to meet defined levels of service. The floating cover on the wastewater treatment lagoon at Melbourne Waters Western Treatment Plant (WTP), in Werribee, is a critical asset whose primary functions are to allow for the anaerobic breakdown of organic matter in the raw sewage and to collect the biogas that is released during this process. Whilst our work initially focused on the development of asset-based diagnostic capabilities to understand how the floating cover respond to the prevailing physical conditions [1-3], coupling these diagnostic capabilities with service-based performance requirements results in new and different imperatives for the structural health and performance monitoring of this engineering structure. This paper reports on the development of an element of this new monitoring strategy and a machine learning capability to predict the biogas collection rates based on past operational decisions and conditions. The study presents a Bayesian Long Short-Term Memory neural network model and investigates its effectiveness for the probabilistic prediction of biogas collection rates at WTP. The probabilistic approach is based on a Gaussian distribution output layer and Monte-Carlo dropout method to estimate the aleatoric and epistemic uncertainties, respectively. The data pre-processing and optimisation of the neural network model are reported. The findings also indicate using a dropout probability beyond 40% adversely prevents learning of complex patterns in the data and overly regularises the network model prediction. The study serves as a fundamental basis in implementing machine learning to transition high-value assets into smarter structures with diagnostic and prognostic capabilities by providing an example where service requirements are also considered.