Objective Natural gas demand is influenced by intertwined factors such as economic fluctuations, policy regulations, and seasonal variations. Inadequate integration of dynamic policy effects with the long- and short-term dependencies in time-series data significantly hinders forecasting accuracy. Therefore, it is essential to develop a coupled forecasting model that incorporates policy time-series characteristics to enhance model adaptability to complex scenarios and improve prediction accuracy.

Methods First, a combined approach of linear interpolation and historical monthly mean interpolation was used to address missing data in the feature sequence of influencing factors. Second, the BorutaShap algorithm was applied for feature importance screening and dimensionality reduction, eliminating redundant features and retaining core information to reduce model input dimensions. Third, a policy-related feature sequence was constructed, incorporating policy hierarchy, seasonal adjustments, time-decay effects, and synergistic or conflicting influences to quantitatively represent policy factors. Meanwhile, a policy gating mechanism was introduced to dynamically adjust feature weights, leveraging the Temporal Convolutional Network (TCN) for capturing long-range trends, Temporal 2D-Variation Modeling Network (TimesNet) for multi-scale periodic analysis, and Bidirectional Long Short-Term Memory (BiLSTM) for local time-series dependencies, thus achieving deep coupling of policy and time-series data. Finally, the Improved Black-winged Kite Algorithm (IBKA) was employed to optimize model hyperparameters, resulting in the integrated IBKA-TCN-TimesNet-BiLSTM natural gas demand forecasting model.

Results To accurately screen input features and optimize dimensionality, the eXtreme Gradient Boosting (XGBoost) model with default parameters evaluated prediction errors across the original dataset, interpolated data, feature sets filtered by Deep Lasso and BorutaShap, their combinations, and the combined feature set. BorutaShap demonstrated the best screening performance. The proposed IBKA-TCN-TimesNet-BiLSTM model, driven by policy quantification, outperformed comparative models, achieving an average absolute percentage error of 2.64％, an average absolute error of 9.42, and a root-mean-square error of 11.44.

Conclusion This method effectively adapts to natural gas demand forecasting under the influence of policies and multiple factors, providing valuable guidance for planning production, supply, storage, sales, and industry decision-making.