P-KTFNet: A prior knowledge enhanced time-frequency forecasting model for natural gas consumption

Natural gas is a crucial transitional fuel in the shift toward cleaner energy systems, offering substantial environmental advantages over traditional fossil fuels. Accurate forecasting of natural gas consumption is vital for effective energy planning, system operating, and management, which contributes to carbon emission reduction targets. However, existing forecasting models often struggle to capture complex time-frequency features and incorporate domain-specific prior knowledge, which hinders accuracy improvement. To overcome the shortcomings of existing studies, this work proposed a Prior Knowledge Enhanced Time-Frequency Network (P-KTFNet), to achieve accurate natural gas consumption forecasting. Time-frequency features are extracted through a dedicated module combining discrete wavelet transformation and convolutional neural networks, enabling a robust fusion of temporal and frequency-domain patterns, which are often underrepresented in traditional methods. Nonlinear features are effectively captured by a parallelized multi-layer temporal memory architecture, enhancing the model's generalization capability and stability across diverse forecasting scenarios. Domain-specific constraints are seamlessly incorporated into the loss function, embedding prior knowledge to improve both prediction accuracy and model robustness significantly. The proposed P-KTFNet model was evaluated using three natural gas consumption datasets from different regions and time granularities. Experimental results demonstrate that P-KTFNet consistently outperforms other state-of-the-art models, such as XGBoost, LSTM, and CNN-LSTM, across all datasets and seasons. Compared with these advanced models, P-KTFNet achieved the lowest Mean Absolute Percentage Error (MAPE), with improvements ranging from 2.76 % to 74.53 %. These results highlight the superior robustness and predictive accuracy of P-KTFNet in diverse scenarios. An ablation study further proves the contributions of each model component, confirming the effectiveness of integrating prior knowledge and time-frequency feature extraction in enhancing model robustness. This research presents a valuable tool for natural gas consumption forecasting, providing insights that support strategic decision-making in energy management.