The microbial communities in biological wastewater treatment plants (WWTP) are complex and consist of thousands of different microorganisms. Many species are process-critical, while others may cause severe problems with foaming and bulking. Being able to not only control, but also predict the microbial community in the near future is of great interest and here modern deep learning models may have a largely unexplored potential. Here we trained a deep learning model based on a 5-year time series data set of activated sludge from a full-scale WWTP to try to predict the microbial community in the near future. The prediction accuracy of an LSTM model combined with IDEC for pre-clustering ASV's was approximated using the Bray-Curtis (BC) dissimilarity measure and averaged at 0.157 for 5 different IDEC clusters. The best prediction had an average BC value of 0.110 indicating good accuracy. The predictions generally seemed to capture the seasonal fluctuations well a full year into the future, but did not precisely capture the magnitude of the fluctuations. This shows that modern deep learning models can be a very useful way to predict the community structure in full-scale WWTPs. Further development could lead to an invaluable tool for management of full-scale WWTPs and optimize their performance.
