Learning from examples of manual control of a central plant refrigerated cabinet

The normal development cycle for algorithms for controlling temperature in central plant refrigerated cabinets is laborious, since each program that is tested must be imprinted on a read-only memory chip. In order to shorten the development cycle a computer based system was set up to test sets of control rules in software rather than hardware. A computer was linked to a central plant refrigerated cabinet allowing software monitoring of the various sensor readings and both manual and rule-based control of the actuating valve. Input and output data was collected during three separate periods while the system was under the manual control of a refrigeration engineer. From this data the fields containing the temperature on and off the cabinet and the evaporator, and the setting of the actuating valve, were selected. Experiments were conducted using machine learning algorithms to induce decision trees and sets of control rules from this data. In each experiment one set of data was used as a training set and all three sets of data were used as testing sets. This was done with each of the algorithms. The machine learning algorithms achieved accuracies of between 12.2% and 98.8% on this task.