A fracture prediction model for tight oil reservoirs based on improved particle swarm algorithm is constructed, introducing dynamic inertia weight adjustment, neighborhood perturbation and entropy weight adaptation strategy on the traditional PSO-SVR framework, and integrating the multi-source features of logging, seismic and core for feature screening and parameter optimization. The experimental results show that the model has RMSE of 0.128, MAE of 0.121, MAPE of 7.46%, and R2 of 0.957 on the test set, which is 23.9% and 14.7% lower compared with SVR and PSO-SVR, and shows higher fitting accuracy and stability in both high and low logging density zones, which verifies the superiority of the method.