Natural gas consumption forecasting technology has been researched for 70 years. This paper reviews the history of natural gas consumption forecasting, and discusses the changes in forecasting horizons, influencing factors, and forecasting performance. According to the characteristics of forecasting models used in different periods, the history of natural gas consumption forecasting can be categorized into initial stage, conventional stage, AI stage, and all-round stage. The stage characteristics, typical models, advantages and disadvantages at different stages have been summarized. The review results show that, affected by the development of computer science and AI technology, short-term forecasting is the fastest-growing forecasting horizon, followed by long-term and medium-term. Additionally, long-term forecasting is mainly affected by production, population, and economic variables. Medium-term forecasting is mainly affected by economic and temperature variables. Influencing factors of short-term forecasting mainly depend on temperature variables, weather condition and date type. Furthermore, the statistical analysis of data characteristics, model characteristics and forecasting results presents that time series models are the best models for long-term forecasting. It has the lowest average mean absolute percentage error (1.90%) in long-term forecasting. To the medium-term and short-term forecasting, AI-based models present the best performance. Among them, artificial neural network models (2.21%) are preferred for medium-term forecasting, and support vector regression models (4.98%) are more suitable for short-term forecasting. Besides, this paper proposes a framework for model selection, and provides specific suggestions for future research directions.

天然气消耗量预测技术已经研究了70年。本文回顾了天然气消耗预测的历史，并讨论了预测范围，影响因素和预测性能的变化。根据不同时期使用的预测模型的特点，天然气消耗预测的历史可以分为初始阶段，常规阶段，人工智能阶段和全面阶段。总结了不同阶段的阶段特征，典型模型，优缺点。审查结果表明，受计算机科学和AI技术发展的影响，短期预测是增长最快的预测范围，其次是长期和中期。此外，长期预测主要受生产，人口和经济变量影响。中期预测主要受经济和温度变量的影响。短期预报的影响因素主要取决于温度变量，天气状况和日期类型。此外，对数据特征，模型特征和预测结果的统计分析表明，时间序列模型是进行长期预测的最佳模型。在长期预测中，它具有最低的平均绝对绝对百分比误差（1.90％）。对于中期和短期预测，基于AI的模型表现出最佳性能。其中，人工神经网络模型（2.21％）更适合中期预测，而支持向量回归模型（4.98％）更适合短期预测。