This package provides the raw data supplied alongside the article: Florian Wenk, Carlo Albert, Eberhard Morgenroth, Andreas Froemelt: "Hybrid Models to Close Knowledge-Gaps in Water Systems: Which Architecture to Choose?", TO BE COMPLETED: DOI, Journal Info etc. will follow when published Please cite this article when using the data. The dataset encompasses operating data of lane 2 of ARA Neugut (ARA = Water Resource Recovery Facility) for one and a half years (July 2022 to December 2023). Further it contains all model results presented in the article, needed for reproducibility due to the stochasticity included in the training process. The abstract of the article is: Hybrid modelling, the combination of data-driven and mechanistic models, is a promising technique for modelling water resource recovery facilities (WRRFs) allowing the integration of unknown or hardly known processes with established mechanistic models. The combination can be made using delta and serial architectures as well as universal differential equations. While hybrid models of all types have been developed, it remains unclear how the architectures compare and what the benefits and limitations of each one are. To address this, we used long-term operational data from a full-scale WRRF to build hybrid models of each architecture for the same problem of identifying a process. This provided a realistic setting to systematically compare the models to each other and to an established mechanistic model, revealing different use-cases for the architectures: Delta and serial architectures are faster to develop and provide good predictions, making them ideal for improving model accuracy for plant operation and digital twins. The universal differential equation is harder to build but allows for better interpretation, which renders it more promising to extend process knowledge. However, the well-calibrated mechanistic model, extended by established understanding, excels the hybrid models in accuracy and is also easier to interpret. This highlights the strength of purely mechanistic models in modelling known processes and indicates the value of mechanistic parts in hybrid models. Establishing these use-cases provides guidance on how to choose a hybrid model architecture and is an important step towards good modelling practice for hybrid models.