In order to overcome the difficulties that independent selecting key temperature points and performing thermal error modeling destroy their intrinsic relation and reduce the thermal error mode prediction performance, a method of concurrently selecting key temperature points and thermal error modeling under unified framework is proposed. The least squares support vector machine (LSSVM) is used as the basic thermal error model. The selection status of temperature points and the hyperparameters of the thermal error model are regarded as the optimization variables. Furthermore, the binary whale optimization algorithm (BWOA) is used to carry out the optimization. And the cost function is designed by comprehensively considering maximizing the prediction accuracy and minimizing the number of key temperature points. Taking a horizontal machining center as the example, the thermal error experiment was conducted. Using the proposed method, the optimal key temperature points were selected in 10fold crossvalidation mode, the number of key temperature points was reduced from 20 to 3, and the model optimal hyperparameters were simultaneously obtained. Finally, the proposed method was compared and analyzed with the traditional independent method. The comparison results indicate that the thermal error prediction accuracy is improved by 628% at most using the proposed modeling method, which verifies its effectiveness and superiority, and the proposed method also provides a reference for subsequent thermal error compensation implementation.