Abstract: Variable fertilization is an important technical approach in implementing precision agriculture. The method of external groove wheel-type variable fertilization with dual regulation of speed and aperture is a typical operation method for crop production (planting) in rice-wheat rotation areas. In response to current issues with variable fertilizer applicators such as slow control system response, inaccurate prediction models, large fertilizer amount errors, and insignificant effectiveness, this study, based on a self-developed dual-variable precision fertilizer applicator for rice and wheat, proposed a method for constructing a fertilizer amount prediction model based on a multilayer perceptron artificial neural network using mathematical statistics and machine learning methods, and verified its effectiveness and applicability. By analyzing the algorithm mechanisms of the levy flight algorithm (LFA), particle swarm optimization (PSO), and multilayer perceptron (MLP) neural network models, and combining the dual-variable fertilization method of aperture-speed, a fertilizer amount prediction model based on LFA-PSO-MLP (LMP) was constructed. The model incorporated the aperture-speed-fertilizer amount relationship, improved algorithm structure through normalization, regularization, etc., conducted parameter optimization and model training, and compared the MLP and PSO-MLP models to obtain the optimal LFA-PSO-MLP fertilizer amount prediction model. Furthermore, an inverse LFA-PSO-MLP (ILMP) prediction model was constructed to quickly calculate the required aperture and speed based on the target fertilizer amount. Experimental results showed that the LFA-PSO-MLP model converged in about 50 iterations, with an R² value of 0.999 after 500 iterations and a mean absolute percentage error (MAPE) of 1.83%, which was better than the other two models. Validation tests of the LMP model yielded an average relative error of 2.47% between predicted and validation values, while field experiments showed an average relative error of 3.49% between predicted and measured values. For the ILMP model, the average relative error for rotation speed prediction was 1.82%, and in field experiments, the maximum relative error between target and actual fertilization rates was 7.22%, with an average fertilization accuracy of 93.92%. This indicated that the fertilizer applicator equipped with the ILMP model performed well in fertilizer application. The study demonstrated that the proposed model construction method can ensure the accuracy of fertilizer amount prediction while improving computational efficiency, achieving fast, precise, and efficient variable fertilization, and improving ecological and economic benefits.