An adaptive vector quantization (AVQ) network clustering algorithm based on feed-forward neural network for the selection of temperature measuring points was proposed in thermal error compensation on machine tools. The method reduces the measuring points from 18 to 3, when adopted on a machining center. Then the relationship between thermal error and key temperature measuring points was established based on the output-input feedback Elman neural network model. The experimental results show that the method proposed can effectively eliminate the coupling among temperature measuring points and improve the accuracy and robustness of the thermal error model.