In order to solve the problems of premature convergence, uneven search ability, and tendency to fall into local optimality in differential grey wolf prediction algorithm, an improved hybrid grey wolf optimization (HGWO) prediction algorithm is proposed, which can adaptively improve and adjust the mutation operator, crossover operator, and mutation strategy. Support vector machine (SVM) with classification prediction function is embedded, while Levy flight global search is used to update the position of the wolves, and the SVM kernel function parameter γ and penalty factor C are optimized. Thus, an HGWO-SVM prediction algorithm is built to predict the large lane of the coke pusher. The results show that, compared with the existing algorithms, the relative errors of position prediction of pedestrian, bicycle, battery car, electric tricycle, and large, medium and small four-wheel vehicle are reduced by 4.21, 4.14, 7.91, 2.03, and 25.53 percentage points, respectively, and the prediction time is reduced by 8.8-10 s. It can overcome the harsh environmental impact of coke oven, accurately predict the trajectory of the moving targets in the lane of the coke pushing vehicle, and provide an active and safe predictive control method for the unmanned operation of coke pushing truck.