Abstract: Oil well production prediction serves as a core basis for production scheduling, cost control, and production plan optimization.This paper addresses the limitations of traditional empirical formula or physical model method in oil well production prediction, which is generally subjected to less accuracy and difficulties in establishment of effective models. On the basis of data-driven technology, traditional BP neural network (FNN-BP), improved FNN-PSO integrated with particle swarm optimization algorithm (PSO), and long-short-term memory neural network (LSTM) are applied to develop the well production prediction model to actually estimate the well production based on such parameters as production time, average downhole pressure, average downhole temperature, average choke size percentage, average wellhead pressure, average wellhead temperature, daily gas and water productivity, etc. The production data from well NO159-F-14H in Volve oilfield abroad were used as the data source to train and validate the model upon the normalization and subdivision of data set. The results indicate that all three models performed well in training, validation, and testing. LSTM exhibits the optimal performance in training, with R² exceeding 0.99 and RMSE of 12.03. During the blind validation phase, LSTM performed the best as usual, with R² greater than 0.97 and RMSE of 5.45. Additionally, when applying PSO for training, the FNN-PSO model showed improvements over the FNN-BP model, with a 1.43% increase in R² and a 4.06% improvement in RMSE, indicating that metaheuristic algorithms for hyperparameter tuning can reduce computational effort and enhance prediction accuracy. The errors of all models approximate a normal distribution, with centers close to zero error, confirming the reliability of the models. This study provides an effective method for oil well production prediction. In the future, integrating metaheuristic algorithms into LSTM models could lead to the development of even more efficient prediction models.