The production dynamics of shale oil horizontal wells are highly complex, and existing prediction techniques often fail to achieve satisfactory accuracy. Taking the Dongying Depression as the study area, this research utilizes production data from wells X1 and X2. First, Granger causality analysis is employed to identify time-dynamic factors highly correlated with shale oil production, optimizing the model input features. Subsequently, a Long Short-Term Memory (LSTM) network is utilized to construct the production prediction model, with its hyperparameters optimized using the Particle Swarm Optimization (PSO) algorithm. The predictive performance of LSTM is compared with Recurrent Neural Networks (RNN), Gated Recurrent Units (GRU), and Temporal Convolutional Networks (TCN). The results demonstrate the critical importance of feature selection for production prediction. For instance, in the case of well X1, the LSTM model based on Granger analysis reduced the root mean square error (RMSE) by 3.41 m3 compared to the model using features selected via Spearman analysis, significantly improving prediction accuracy. While various time-series models exhibit strong predictive capabilities, the LSTM model outperforms others in capturing dynamic characteristics of time series, providing a robust theoretical foundation and technical support for complex shale oil production forecasting.