Transfer Learning for Time-series Forecasting of Greenhouse Microclimate

This paper presents the adaption of a data-driven model of the microclimate of a first greenhouse (of 877 m2) to a second, different greenhouse (of 1900 m2) by using transfer learning. The temporal evolution of temperature and humidity inside the first greenhouse were modeled using a transformer-based model. The transformer was trained and validated with an 81-day dataset from the first greenhouse, for multi-step prediction of the microclimate. Subsequently, the transformer was used with a 48-day dataset from the second greenhouse. Transfer learning was used to fine-tune the weights and biases of the transformer. Results show that, by using transfer learning, only two days of data were necessary to train the transferred model. This also shows that the initial transfomer generalized well and learned basic greenhouse dynamics that only need to be adapted slightly when greenhouse specifications change. Therefore, transfer learning is presented as a method that can facilitate the reuse or adaptation of neural-based predictive controllers between different greenhouses.