A production prediction method for well groups in fracture-vuggy carbonate reservoirs based on CNN-iTransformer by integrating geological and operational constraints

Fractured-vuggy carbonate reservoirs pose major challenges for production forecasting because of strong geological heterogeneity, complex flow behavior, and the difficulty of jointly incorporating static geological information and dynamic operational data. To address this issue, this study develops an interpretable cross-modal forecasting framework for the YM2-8 well group by combining a one-dimensional convolutional neural network (CNN) with an inverted Transformer (iTransformer). The CNN extracts static geological features from well-log data, whereas the iTransformer models inter-variable dependencies in dynamic production and operational sequences through variable-wise attention. The results show that the proposed method improves rolling forecasting accuracy for both oil-rate and water-rate predictions, reducing the Mean Absolute Error by ∼21.4% and 16.3%, respectively, compared with the iTransformer, while also improving the tracking of production variations under water-drive conditions. Overall, the results demonstrate that integrating geological and operational constraints can improve both the accuracy and interpretability of well-group production forecasting in highly heterogeneous reservoirs.