Data Cleaning and Anomaly Detection for an Intelligent Greenhouse

The effectiveness of greenhouse control can be improved by the application of model based intelligent control. Such control requires a good model of the greenhouse. For a large variety of industrial or recreational greenhouses the derivation of an analytical model is not feasible therefore black-box modeling has to be applied. Identification of black-box models requires large amount of data from real greenhouse environments. Measurement errors or missing values are common and must be eliminated to use the collected data efficiently as training samples. Rare weather conditions can temporally lead to unusual thermal behavior around and within the greenhouse. Anomaly detection run on the measurement data can identify such unusual samples and by excluding those from the model building better models and higher validation accuracy can be achieved. This chapter discusses problems of cleaning the measurement data collected in a well instrumented greenhouse, and introduces solutions for various kinds of missing data and anomaly detection problems. 1 The Concept and the Requirements of Intelligent Control Greenhouses are built in various sizes and types all around the world to house plants needing special environmental conditions. Greenhouses are widely used both for vegetable and ornamental plant production. The basic operation of a greenhouse is as follows: The transparent walls and roofs allow the solar radiation to pass through, but the warmed up air is kept inside. To prevent extreme high or low temperatures inside the house several actuators can be utilized. Shading curtains, automatic windows and active cooling systems can slow down the temperature rise in the summer while heating appliances are used in the cold season.