Production forecasting is one of the important contents of the research and development of oilfield production dynamics. The long-term production of oil fields has accumulated a large amount of data, but the fluctuation range is very large. The long-short term memory neural network is directly applied to predict the production indicators of oil fields, and the generalization of neural networks is very poor. Therefore, this paper firstly uses the two-layer LSTM and random inactivation layer to adjust the neural network architecture, establishes a deep learning neural network model; then proposes a new fruit fly aggregation method, through improved fruit fly optimization. The algorithm optimizes the established neural network model to avoid falling into the local optimal solution and searching for the optimal solution of the solution space. Finally, the field example verification shows that the network generalization ability and prediction accuracy of the optimized deep learning network have been greatly improved, and the data with large fluctuations in the oilfield can be fitted well. The oilfield production prediction model established in this paper can be applied to the actual development of the mine.