The vertical inline pump, a single-stage centrifugal pump with a bent elbow inlet, is widely used in marine engineering. The unique water inlet passage combined with uneven inflow at the impeller inlet tends to form an inlet vortex and secondary flow area, which reduces performance and causes vibration. To predict the performance of the elbow inline pump, this study uses spline curve fitting for the centerline and cross-sectional shape of the elbow passage. With four elbow inlet variables from experimental design as the input layer and targeting efficiency under pump operating conditions, a pump performance prediction model based on an improved sand cat swarm optimization algorithm combined with a BP neural network (MSCSO-BP) is proposed. Six test functions are used to effectively test the improved sand cat swarm optimization algorithm. The results show that compared to the unimproved algorithm, the improved algorithm has significantly faster convergence speed, shorter parameter optimization time, and higher accuracy. For more demanding multidimensional test functions, the improved optimization algorithm can more accurately find the optimal solution, enhancing the prediction accuracy and generalization ability of inline pump performance. This provides a more effective engineering solution for the design and optimization of inline pumps.