In large low-lift pump stations, the pump assembly comprises an inlet conduit, a pump, and an outlet conduit. A short conical frustum section that connects the elbow section with the impeller inlet directly affects the impeller inflow state, thereby influencing the overall performance. Therefore, investigating the conical frustum section contributes to studying the effect of inflow states on the performances of pump assemblies and similar pumping systems. To improve the pump assembly efficiency, three parameters of the conical frustum section, i.e., the contraction angle, height, and centerline inclination angle, are investigated and optimized via univariate and multivariate analyses. The flow field and external characteristics of the pump assembly are investigated via computational fluid dynamics simulation with a constant head. Furthermore, a comprehensive analysis and discussion of the performance improvement mechanisms are presented. The results indicate that the axial velocity distribution at the impeller inlet conforming to the cascade high-efficiency characteristics will achieve a better pump performance compared with a uniform distribution. The pump efficiency distribution can be predicted and visualized based on the cascade efficiency characteristics and the flow state at the impeller inlet using a machine learning method. In addition, the directions and distribution of the lateral and axial components of the inflow velocities have great impacts on the circulation distribution. A sensible circulation distribution at the guide vane outlet can suppress the entropy production and reduce hydraulic loss of the outlet conduit. In this case, a significant increase in the pump assembly efficiency is obtained.