TAMNet: Temporal and adaptive‐frequency network with MixStyle for cross‐region oil and fluid production forecasting

Disparities in oil production among wells within the same oilfield often arise from differences in geological structures, production modes, and environmental conditions. These inconsistencies present major challenges for accurate and robust modeling and forecasting of multi‐well productivity. To address this issue, this paper proposed a deep time series modeling framework called temporal and adaptive‐frequency network with MixStyle (TAMNet), which combines transformer and long short‐term memory (LSTM) architectures for comprehensive temporal feature learning. First, a temporal gate unit (TGU) is introduced to enhance the model's ability to capture dynamic production trends. By adopting time gating, TGU selectively extracts informative time‐step features, improving sensitivity to temporal fluctuations and enabling more precise short‐term modeling. Second, an adaptive frequency transform (AFT) module is designed to incorporate frequency‐domain information into temporal representations. This allows the model to effectively capture periodic patterns commonly observed in oil production signals, thereby enhancing its perception of long‐term production behaviors. Finally, to improve generalization across heterogeneous well domains, we incorporate a MixStyle‐based domain generalization mechanism. By perturbing intermediate feature statistics during training, this technique encourages the model to learn domain‐invariant representations, reducing overfitting and improving adaptability to unseen conditions. Through extensive experiments on real‐field datasets, TAMNet shows strong predictive performance in multi‐well productivity prediction compared to the state‐of‐the‐art baseline models. Whether in the same oilfield area or different oilfield areas, the prediction results of TAMNet can effectively fit the changing trend of production, showing good stability and cross‐domain generalization ability.