Shale oil production time series forecasting for multi-fractured horizontal wells with optimized artificial neural networks integrating multi-source data.

This article focuses on developing and evaluating a hybrid deep learning model, termed Bi-LSTM/GRU-MLP-SA, which integrates bidirectional long short-term memory (Bi-LSTM) or gated recurrent unit (Bi-GRU) networks with multi-layer perceptron (MLP) and a self-attention (SA) mechanism, to predict shale oil production rates. Utilizing multi-source data from geological, fracturing, and production parameters of three horizontal wells, the model employs the Hiking Optimization Algorithm (HOA) for hyperparameter tuning and SHapley Additive exPlanations (SHAP) for interpretability. Experimental results demonstrate that the Bi-GRU-MLP-SA variant outperforms other models in accuracy and stability, achieving lower root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) across wells. The study highlights the model’s ability to capture complex temporal and nonlinear relationships in shale oil production data, while providing transparent insights into feature importance, and suggests its applicability to other time series forecasting domains.