Development of a digital twin of the DLR EDEN ISS greenhouse for data enabled analysis

Kurzfassung

Digital Twins as a virtual entity of a physical system open up a variety of new opportunities. Their development on the other hand can be very challenging especially for complex systems. The EDEN ISS Mobile Test Facility with its Future Exploration Greenhouse belongs to those more sophisticated setups. During its operations since 2018 a great volume of data is collected which is including more than 200 sensors and a wide crew documentation. In this paper the development of a Digital Twin is approached by utilizing this data for machine learning. Specifically, a Neural Network is developed that recreates the same changes on the features of the dataset as the physical system did in the past. In the event of success the Neural Network is holding a general representation of the physical system and also future or virtual changes can be simulated. In particular, time series forecasting is chosen to perform the prediction making by taking the past 24 hours as input and returning the following 24 hours as an output, both in a 5 minute step resolution. The development starts by the preparation of the dataset, where individual data tables are merged, missing data is interpolated and constant or corrupted features are identified and removed. The general structure of the Digital Twin is divided into 3 sub-models to address categorical differences of the features. The first distinction is made according to the documentation style between the sensor data and the crew data, which is containing the information about the plants. A further differentiation is done inside the sensor data between time controlled variables and environment controlled variables whether the values are regulated by a time schedule or not. Those 3 sub-models are named time controlled model, environment controlled model and plant model. The identification of time controlled variables is carried out by sorting over the error term of a model that assumes a 24 hour periodic repetition. The remaining ones from this identification make up the environment controlled variables. The first sub-model, the time controlled model is developed by applying a whole group of models on this subset and selecting the model with the highest precision by the error term they created. The previous mentioned repetition model was selected which offered a very small error in prediction. The second sub-model, the environment controlled model is generally following the same approach. A smaller set of Neural Networks is applied on the new structure and again the best performing model is chosen for further investigation. The LSTM-model is offering the lowest error in prediction here and is expanded and regulated for a second training. Besides further improvement, this model is still ending up on a high level of error when averaged over all features. These errors in prediction vary greatly on the individual variables and range from decent precision up to a deterioration over the baseline. The development of the third and final sub-model, the plant model resulted to be unsuccessful. The insufficient amount and resolution of the plant data is missing the requirements for the intended objective of bringing the plants biomass in correlation with system values. In conclusion, the development of a Digital Twin for the EDEN ISS system is prototypically achieved, while the approach of building a Digital Twin without the system being designed for it encountered heavy limitations.