Combination of Kalman Filtering and Recursive Gaussian Process Regression With Application to Vapor Compression Cycle Control

This paper presents the combination of recursive Gaussian process regression with a Kalman Filter to learn data-based models for non-measurable outputs in real time. The algorithm utilizes a Kalman Filter with an auxiliary parameter for a disturbance estimate. In parallel, the estimated disturbance estimate is employed in a recursive Gaussian process regression for the identification of a corresponding disturbance model. This disturbance model is in turn leveraged to improve the Kalman Filter prediction step. The proposed algorithm is compared in simulations with an alternative baseline implementation. Exemplarily, an experimental validation of the algorithm is presented for the learning of a heat-transfer coefficient in a vapor compression cycle as used in refrigeration systems and heat pumps.