Effective production prediction is vital for optimizing energy resource management, designing efficient extraction strategies, minimizing operational risks, and informing strategic investment decisions within the energy sector. This paper introduces a Dual-Stage Attention Temporal Convolutional Network (DA-TCN) model to enhance the accuracy and efficiency of gas production forecasting, particularly for wells in tight sandstone reservoirs. The DA-TCN architecture integrates feature and temporal attention mechanisms within the Temporal Convolutional Network (TCN) framework, improving the model’s ability to capture complex temporal dependencies and emphasize significant features, resulting in robust forecasting performance across multiple time horizons. Application of the DA-TCN model to gas production data from two wells in Block T of the Sulige gas field in China demonstrated a 19% improvement in RMSE and a 21% improvement in MAPE compared to traditional TCN methods for long-term forecasts. These findings confirm that dual-stage attention not only increases predictive accuracy but also enhances forecast stability over short-, medium-, and long-term horizons. By enabling more reliable production forecasting, the DA-TCN model reduces operational uncertainties, optimizes resource allocation, and supports cost-effective management of unconventional gas resources. Leveraging existing knowledge, this scalable and data-efficient approach represents a significant advancement in gas production forecasting, delivering tangible economic benefits for the energy industry.

Accurate production forecasting of tight gas reservoirs plays a critical role in effective gas field development and management. Recurrent-based deep learning models typically require extensive historical production data to achieve robust forecasting performance. This paper presents a novel approach that integrates transfer learning with the neural basis expansion analysis time series (N-BEATS) model to forecast gas well production, thereby addressing the limitations of traditional models and reducing the reliance on large historical datasets. The N-BEATS model was pre-trained on the M4 competition dataset, which consists of 100,000 time series spanning multiple domains. Subsequently, the pre-trained model was transferred to forecast the daily production rates of two gas wells over short-term, medium-term, and long-term horizons in the S block of the Sulige gas field, China’s largest tight gas field. Comparative analysis demonstrates that the N-BEATS transfer model consistently outperforms the attention-based LSTM (A-LSTM) model, exhibiting greater accuracy across all forecast periods, with root mean square error improvements of 19.5%, 19.8%, and 26.8% of Well A1 for short-, medium-, and long-term horizons, respectively. The results indicate that the pre-trained N-BEATS model effectively mitigates the data scarcity challenges that hinder the predictive performance of LSTM-based models. This study highlights the potential of the N-BEATS transfer learning framework in the petroleum industry, particularly for production forecasting in tight gas reservoirs with limited historical data.