The invention relates to a thermal error modeling method based on gray linear regression. The method comprises the following steps that first, on the basis of a gray thermal error model, a linear equation is introduced, a gray linear regression combination model is constructed; second, a least square method is used for solving a gray linear regression combination model parameter; third, the gray linear regression model is used for thermal error prediction; fourth, a BP nerve network is used for amending combination model residual errors, and prediction accuracy is improved. According to the method, the shortcoming that a linear regression model does not have exponential growth and cannot describe linear changing trend easily, and a gray thermal error model does not have a linear factor can be overcome, good capacity for solving linear and nonlinear problems is achieved, good effect is achieved for thermal error prediction on an accurate horizontal type machining center is achieved, linear factors and nonlinear factors of thermal error data are considered, the shortcoming of an original single gray model is overcome, and an accurate thermal error prediction value and high fitting degree are acquired.