Online Learning Framework for Nonlinear Dynamic Thermal Modeling

In order to obtain highly-efficient management schemes for thermal components and systems, the thermal dynamic modeling must be robust and accurate. Resistance-capacitance (R-C) thermal models involving unknown coefficients have been widely employed to realize the dynamic modeling following the gray-box methodology. In most of cases, R-C thermal models are amused to be linear, neglecting all the possible dependence of heat transfer coefficients on the node temperatures. Those linear models may become inapplicable, when nonlinearity effects are non-negligible. In addition, the training process of R-C models to determine the unknown coefficients is usually offline currently, which indicates the users have to re-train the models once the heat transfer features change during the operation process. Here, an ensemble Kalman filter (EnKF)-based framework integrated with Gaussian process (GP) is adopted to train nonlinear R-C thermal models to realize the self-adaptive dynamic modeling. The machine learning method, GP, guarantees the high-degree self-adaption of the framework, while the EnKF achieves the online training and the integration of nonlinear R-C thermal model and GP. Furthermore, the performance of our algorithm is evaluated on the synthetic datasets, which well proves its feasibility.