Abstract: Accurate prediction of gas load holds significant practical value for maintaining the dynamic balance between gas supply and demand. With advancements in artificial intelligence technology, gas load forecasting algorithms have undergone substantial development. This study first categorizes forecasting periods into three distinct phases: short-term (ST), medium-toshort-term (MST), and medium-to-long-term (MLT). From an algorithmic perspective, we systematically analyze six representative methods including eXtreme Gradient Boosting (XGBoost) for ST forecasting, two approaches such as Long Short-Term Memory (LSTM) networks for MST forecasting, and two techniques including Prophet for MLT forecasting, and evaluate their advantages, limitations, and application scenarios. Through empirical validation using operational data, we conduct multidimensional comparative analysis of 12 selected models. Our experimental framework encompasses critical aspects including dataset construction, data preprocessing, extrapolative prediction, parameter optimization, and model evaluation across different temporal scales. Finally, we propose forward-looking perspectives on future research directions in gas load forecasting, particularly focusing on practical applications in natural gas dispatch management. This comprehensive investigation provides valuable references for advancing algorithmic research in gas load prediction and its implementation in smart energy management systems.

Key words: gas load forecasting, short-term forecasting, medium-to-short-term forecasting, medium-to-long-term forecasting, machine learning, deep learning, review