Real-time prediction of thermal load and indoor temperature for buildings with a recursive identification method in noise conditions

Accurate prediction of building thermal load and indoor temperature serves as a critical prerequisite for energy-efficient and comfortable HVAC system operation. Prediction via RC thermal network model represents a key methodology, with thermal parameter identification being the cornerstone of this approach. Conventional intelligent search methods suffer from prohibitive computational demands, hindering their practical engineering applications. Furthermore, the noise introduced by flawed sensor and transmission interference significantly compromises the reliability of the model-based prediction method. This study proposes an online identification method integrating Polynomial Kalman Smoother (PKS) and Recursive Generalized Total Least Squares (RGTLS) to achieve robust thermal parameter identification in noisy conditions and enable real-time thermal load and indoor temperature predictions. PKS first estimates noise error covariance matrix, followed by RGTLS’s parameter identification. Experimental validation demonstrates that the method accomplishes high-resolution parameter identification within 2 s while eliminating extensive historical data storage requirements, substantially reducing hardware configuration demands. In noisy conditions, the average identification error of the parameters decreased from 59.78% to 6.35%, achieving the MAE of 0.48 °C for indoor temperature and 6.90 kW for cooling power prediction. Noise tests and hyperparameter sensitivity analyses reveal that the method achieves high performance and robustness during identification. The proposed reliable prediction method enables direct deployment on legacy HVAC systems and simultaneously establishes a foundation for embedded MPC controllers, enhancing operational efficiency to accelerate building decarbonization.