We present a learning-based robust observer design for thermal-fluid systems, pursuing an application to efficient energy management in buildings. The model is originally described by Boussinesq equations which is given by a system of two coupled partial differential equations (PDEs) for the velocity field and temperature profile constrained to incompressible flow. Using proper orthogonal decomposition (POD), the PDEs are reduced to a set of nonlinear ordinary differential equations (ODEs). Given a set of temperature and velocity point measurements, a nonlinear state observer is designed to reconstruct the entire state under the error of initial states, and model parametric uncertainties. We prove that the closed loop system for the observer error state satisfies an estimate of L2 norm in a sense of locally input-to-state stability (LISS) with respect to parameter uncertainties. Moreover, the uncertain parameters estimate used in the designed observer are optimized through iterations of a data-driven extremum seeking (ES) algorithm. Numerical simulation of a 2D Boussinesq PDE illustrates the performance of the proposed adaptive estimation method.