Aiming to address the issues of poor generalization of data-driven models and insufficient representation of spatial heterogeneity in current mushroom house model predictive control, a data mechanism collaborative modeling and global temperature field collaborative control method for heterogeneous mushroom houses was proposed, with the goal of energy saving and consumption reduction in industrial mushroom room production. Firstly, an environmental sensor placement optimization model was developed to capture complex thermal dynamics by using only a minimal number of sensors. Secondly, a simplified physics-informed neural networks (PINNs) architecture based on fully connected layers was designed, incorporating simplified Navier - Stokes and energy conservation equations as physical constraints to enhance temperature field prediction accuracy. Finally, an MPC framework was formulated with a joint optimization objective of temperature tracking error and energy consumption, and systematically validated in two representative production settings: a prefabricated modular Pleurotus citrinopileatus cultivation chamber in Fangshan District and a standardized Hypsizygus marmoreus factory facility in Tongzhou District of Beijing. Results showed that under sparse monitoring with only six sensors, the proposed method achieved a root mean square error (RMSE) of 0.267℃ in reconstructing the full three-dimensional temperature field in the modular chamber. The PINNs - MPC strategy enabled precise whole-space temperature control in the standardized facility, maintaining temperatures stably within 14.2℃ to 15.2℃, whereas in the modular chamber, localized temperature violations occurred due to strong environmental disturbances and non-standardized agronomic operations. In terms of energy efficiency, PINNs - MPC reduced energy consumption by 10.7% compared with conventional threshold-based control; however, its conservative control behavior limited further savings, as a thermally constrained MPC approach based on an energy-balance equation achieved a higher reduction of 13.1%. By effectively integrating physical laws with sparse observational data, this work significantly enhanced modeling fidelity and spatial awareness of thermal fields in heterogeneous agricultural environments, offering a theoretically grounded and technically viable pathway toward intelligent, energy-efficient, and robust environmental control in modern mushroom cultivation.