Transfer Learning-Based Dynamic Production Prediction for Tight Oil: Considering Multimodal Features and Long-Term Temporal Dependencies

Tight oil horizontal wells in the Mahu block of the Junggar Basin commonly show rapid production decline and limited target-domain samples. These characteristics make accurate production prediction difficult. This work aims to address the small-sample overfitting problem of tight oil horizontal well production prediction in Mahu Sag with rapid production decline and limited measured well data. The concept of transfer learning is introduced to address the issue of insufficient target domain samples, and the Pearson correlation coefficient is utilized to select the main controlling factors for production from the production data. Next, based on the features extracted by the temporal convolutional network at different data scales, a multi-head attention mechanism is introduced to capture the dependencies across different time steps. Subsequently, an improved sparrow search algorithm is employed to optimize the hyperparameters of the bidirectional long short-term memory network. Finally, the bidirectional long short-term memory network is integrated to further extract the nonlinear features learned by the temporal convolutional network to conduct production prediction. Tailored to the exploitation conditions of tight oil horizontal wells in this block, a tight oil production prediction model based on transfer learning and the multi-head attention mechanism is proposed. Experimental results demonstrate that, compared with the standard bidirectional long short-term memory network, the proposed model’s evaluation metrics show a 60.93% decrease in root mean square error, a 78.53% decrease in mean absolute percentage error, and a 43.68% increase in coefficient of determination. This verifies the effectiveness of transfer learning in solving small-sample modeling challenges, providing precise technical support for the optimization of tight oil fracturing parameters and stimulation treatments in the Mahu block. The novelty of this work lies in the integration of multi-head attention temporal convolution network, quantum sparrow optimized bidirectional long short-term memory network and cross-block transfer learning for small-sample tight oil forecasting.