The effects of model reduction and data assimilation on greenhouse climate predictions

We investigated the effect of model reduction and data assimilation on prediction accuracy of greenhouse climate. For this, a first-principle model was reduced, and calibrated with measurement data. Calibration data consisted of a time series of temperature, humidity, and carbon-dioxide concentration in a rose greenhouse, together with 15 variables related to outside climate and control actions. The results indicate that model reduction does not produce a crucial loss of prediction accuracy. In contrast, data assimilation decreases the size and variance of prediction errors drastically, making predictions much more reliable. A static linear model seems to predict the most essential input-output response for temperature and humidity, but the predictive power for carbon-dioxide concentration is limited. The prediction errors have standard deviations of typically 2°C, for temperature, 5-10% for relative humidity, and 200-300 ppm for CO2. The prediction errors have biases of the same order, which differ per period for which the predictions are made. We believe these results are promising for modelling climate via a static black box approach, in combination with data assimilation. The relatively low computational demand for uncertainty analysis and easy model building provide a suitable starting point for investigating augmented systems, such as plant development based on controlled climate.