Performance during the seed-filling process directly impacted the seed quality of the metering device. The improved BP neural network prediction model was a metering device that filled at a single-grain rate η1 and the miss rates η2 was established using the Matlab neural network toolbox. The speed n, seed equivalent diameter d, seed-filling angle β and type hole diameter D were selected as the test factors, the test was carried out on 64 groups to determine the single-particle and miss rate. 55 groups were selected from the test as training samples. The Levenberg-Marquardt training method was used to train the establishment of a network. The remaining 9 groups were selected to simulate and predict the trained and improved BP neural network. n, d, β and D were set as the network’s input layers, η1 and η2 were set as the network’s output layers, the network structure was the 4-15-2 type three-layer network containing a single hidden layer. Predicted results showed that predicted values and experimental values were almost same, the predicted performance of seed-filling with the improved BP neural network method was feasible, the method can be used to optimize metering device design and provide a basis for the selection of working parameters, in addition to reducing test time and cost. 