Dynamic Heating, Ventilation, and Air-Conditioning (HVAC) system models are used for the purpose of control design, fault detection and diagnosis, system analysis, design and optimization. Therefore, ensuring the accuracy and reliability of the dynamic models is important before their application. Parameter tuning and model validation is a crucial way to improve the accuracy and reliability of the dynamic models. Traditional parameter tuning and validation methods are generally time-consuming, inaccurate and can only handle a limited number of tuning parameters. This is especially true for multiple-input-multiple-output (MIMO) models due to their intrinsic complexity. This dissertation proposes a new automatic parameter tuning and validation approach to address this problem. In this approach, a fast and accurate model is derived using linearization. Discrete-time convolution is then applied on this linearized model to generate the model outputs. These outputs and data are then processed through wavelet decomposition, and the corresponding wavelet coefficients obtained from it are used to establish the objective function. Wavelets are advantageous in capturing the dynamic information hidden in the time series. The objective function is then optimized iteratively using a hybrid method consisting of a global search genetic algorithm (GA) and a local gradient search method. In order to prove the feasibility and robustness of the proposed approach, it is applied on different dynamic models. These models include an HVAC system model with moving boundary (MB) heat exchanger models, a heat pump model with finite control volume (FCV) heat exchanger models, and a lumped parameter residential conditioned space model. These models generally have a large number of parameters which need tuning. The proposed method is proved to be efficient in tuning single data set, and can also tune the models using multiple experimental or field data sets with different operating conditions. The tuned parameters are further cross-validated using other data sets with different operating conditions. The results also indicate the proposed method can effectively tune the model using both static and transient data simultaneously.
