Aiming at the optimization of injection-production parameters in the process of waterflooding development，it is proposed to use neural network instead of numerical simulation to predict the distribution of remaining oil，and combine the gradient-free differential evolution algorithm to optimize injection and production parameters. This model not only establishes the nonlinear relationship between injection-production parameters and the objective function，but also accurately predicts the distribution characteristics of remaining oil in different production stages. The prediction principle of the model is that the injection-production parameters and production time are regarded as advanced features of the remaining oil distribution image. Then，the convolution layer is used to extract image features and the transposed convolution layer is used to carry out up-sampling on the image. The original image is recovered step by step through a combination of multi-layer “convolution+transpose convolution”，so as to achieve the effect of accurate prediction. During the construction of the neural network，multiple 3×3 small convolution kernels are selected to replace the large convolution kernels. Without affecting the receptive field，the amount of parameters is reduced，and the calculation cost is saved. The iteration efficiency during model training is effectively improved. Taking the five-point well pattern of 4 injection wells and 5 production wells in a block as an example，a neural network prediction model was established by taking the bottom hole pressure of the production well，the injection volume of the injection well and production time as input parameters. Taking the net present value （NPV）as the objective function，the injection-production parameters of the four stages were optimized through the differential optimization algorithm. Compared with the basic plan，the NPV of the optimized plan has increased by about 21%.