The liquid viscosity of hydrocarbon compounds is essential in the chemical engineering process design and optimization. In this paper, we developed a quantitative structure property relationship (QSPR) model to predict the hydrocarbon viscosity at different temperatures from the chemical structure. We collected viscosity data at different temperatures of 261 hydrocarbon compounds (C-3-C-64), covering n-paraffins, isoparaffins, olefins, allcynes, monocyclic and polycyclic cycloalkanes, and aromatics. We regressed the experimental data using an improved Andrade equation at first. Hydrocarbon viscosity versus temperature curves were characterized by only two parameters (named B and T-0). The QSPR model was then built to capture the complex dependence of the Andrade equation parameters upon the chemical structures. A total of 36 key chemical features (including 15 basic groups, 20 united groups, and molecular weights) were manually selected through the trial and-error process. An artificial neural network was trained to correlate the Andrade model parameters to the selected chemical features. The average relative errors for B and T-0 predictions are 2.87 and 1.05%, respectively. The viscosity versus temperature profile was calculated from the predicted Andrade model parameters, reaching the mean absolute error at a value of 0.10 mPa s. We also proved that the established QSPR model can describe the viscosity versus temperature profile of different isomers, such as isoparaffins, with different branch degrees and aromatic hydrocarbons with different substituent positions. At last, we applied the QSPR model to predict gasoline and diesel viscosities based on the measured molecular composition. A good agreement was observed between predicted and experimental data (absolute mean deviation equals 0.21 mPa s), demonstrating that it has capacity to calculate viscosity of hydrocarbon mixtures.