Combining mechanistic and data-driven techniques for predictive modelling of wastewater treatment plants

Mechanistic models are widely used for modelling of wastewater treatment plants. However, as they are based on simplified and incomplete domain knowledge, they often lack accurate predictive capabilities. In contrast, data-driven models are able to make accurate predictions, but only in the operational regions that are sufficiently described by the dataset used. We investigate an alternative hybrid model, combining mechanistic and data-driven techniques. We show that the hybrid approach combines the strengths of both modelling paradigms. It allows for accurate predictions out of the training dataset without the need for complete domain knowledge. Moreover, this approach is not limited to wastewater treatment plants and can potentially be applied wherever mechanistic models are used.