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

Time series forecasting is crucial for guiding capital investment, production enhancement, and optimization in the oil and gas industry. However, conventional data-driven approaches for the production prediction fail to meet the industry's criteria. This paper develops a hybrid model combining bidirectional long short-term memory (Bi-LSTM) or bidirectional gated recurrent unit (Bi-GRU) with multi-layer perceptron (MLP) and self-attention (SA), termed Bi-LSTM/GRU-MLP-SA, to predict shale oil production rates. The SHapley Additive exPlanations (SHAP) method is applied to enhance the model's interpretability. The proposed model architecture consists of five key components: input layers, Bi-LSTM/GRU layers, MLP layers, SA layers, and output layers. The Bi-LSTM/GRU captures temporal dependencies from time-series data, while the MLP captures relevant information from non-sequential data. The SA mechanism allows the model to focus on the most salient parts of the data. Compared to traditional single-technique models like standalone Bi-LSTM/GRU, Bi-LSTM/GRU with SA (Bi-LSTM/GRU-SA), and Bi-LSTM/GRU combined with MLP (Bi-LSTM/GRU-MLP), our Bi-LSTM/GRU-MLP-SA model demonstrates superior performance. Specifically, the Bi-GRU-MLP-SA variant achieved an average root mean square error (RMSE) of 0.2763, a mean absolute error (MAE) of 0.2192, and a mean absolute percentage error (MAPE) of 0.0490, indicating a higher accuracy and stability. In summary, the Bi-GRU-MLP-SA model is the most effective among the evaluated methods for identifying underlying trends in shale oil production and accurately predicting production levels.