Only fragments of genomes can be extracted from wastewater samples, so models are needed to estimate proportions of Variants of Concern (VOCs). This relies on known lists of mutations to define a VOC. I extend existing models with structural time series models which assume the proportion of a VOC in each location is a deviation from a city-wide proportion, using either an autoregressive structure or a Gaussian process for the temporal trend. An extension of this model estimates the proportions simultaneously with the definitions of the VOCs using unsupervised machine learning. This model takes information from multiple samples to find patterns of mutations (basis functions) that appear together, with proportions (coefficients) constrained to be smooth over time and space. I find concordance between temporal trends of estimated VOCs and known VOCs. This allows for tracking variants without clinical sequencing (which is costly and often biased).