A hybrid MLP-CNN-transformer architecture for high-accuracy dynamic production forecasting in shale oil

This paper addresses the challenge of insufficient accuracy in shale oil well production forecasting by proposing a deep learning model that integrates static engineering constraints with dynamic temporal features—MLP-CNN-Transformer. The model encodes static parameters such as geological attributes and fracturing scale using a multilayer perceptron (MLP), incorporating them as physical constraints. It extracts local dependencies and short-term trends from historical production sequences via a convolutional neural network (CNN), and employs Transformer's self-attention mechanism to model long-term dynamic interdependencies and global temporal patterns among multiple variables. The synergistic operation of these three components achieves a multi-scale prediction mechanism characterized by "local perception-global correlation-knowledge fusion." Validation based on actual production data from 80 wells in the Changqing Oilfield, China, demonstrates that the model achieves a coefficient of determination (R²) of 0.8991 on the test set, significantly outperforming comparative models such as GRU, LSTM, and ANN. Ablation experiments confirm the necessity of each module: the absence of CNN leads to a sharp increase in prediction error, highlighting its core role in capturing local temporal patterns; the removal of MLP reduces the model's goodness-of-fit, reflecting the critical contribution of static feature integration to prediction rationality. Furthermore, static feature importance analysis reveals that engineering parameters such as "number of fracturing stages" and "total sand volume per well" have the most significant impact on production forecasting. This research provides a data-driven approach for shale oil production prediction that is characterized by high accuracy, strong generalization capability, and engineering interpretability, offering significant application value for development plan optimization and real-time productivity assessment.