Optimization selection of the temperature measuring point is crucial during thermal error modeling. A method using correlation analysis was present to analyze the relationship between the spindle thermal drift and point temperature of measurement. The temperature distribution points of the optimal choice were achieved by finding a higher correlation measuring point of location. On basis of this, by using simulated annealing and genetic algorithm (GSA) optimized BP neural network method, the thermal error model was established, and was verified by experiments. The results show optimized thermal error model can escape from local optimal and achieve global optimal solution. The resulting error model can fit values more closer to the actual measured error values. Based on simulated annealing genetic algorithm (GSA) optimization, BP neural network model has higher accuracy and greater robustness than that of the traditional neural network model.