Abstract: In this study, the authors aim to establish a reasonable and accurate natural gas forecasting model to provide guidance information for the future work of related companies and departments in the ever-changing and complex background of the natural gas market. Firstly, 12 factors affecting natural gas consumption are selected from the aspects of economic level, industrial structure, energy structure, and natural gas price. Secondly, a Bayesian Model Average(BMA)method is used to construct a benchmark model containing six influencing factors commonly used in related literature. On this basis, a set of comparative models is established by adding various factors that affect the consumption of natural gas one by one. Then we select the model with the highest prediction accuracy to predict future natural gas consumption. Finally, in order to detect the prediction effect of the BMA model, it is compared with the ARIMA model, the ETS model, the BVAR model, the STEP model and the equal weighted average model. The results show that the optimal BMA model includes nine factors that affect the economic level, industrial structure, energy structure, population factors, natural gas prices, and natural gas supply. The prediction accuracy is better than the comparison prediction models. The model predicts that natural gas consumption in 2022 will reach 325.413 billion cubic meters with an average annual growth rate of 8%.

Key words: Bayesian model averaging, influence factors, model select, scenario analysis, gas consumption forecast

摘要： 作为一种优质、高效的绿色能源,天然气在中国能源结构中所占比重逐渐增加。但可再生能源的崛起使得天然气成为过渡能源的选择,天然气消费量的增长趋势不明晰,因此相关企业及部门需要合理、准确的天然气需求预测模型为未来的工作提供指导性信息。基于此,本文首先从经济水平、产业结构、能源结构、天然气价格等方面选取影响天然气消费的12个因素。其次,运用贝叶斯模型平均(BMA)法构建了一个包含相关文献中常用的6个影响因素的基准模型,针对该模型,围绕影响天然气消费量的各种因素,以逐个添加的方式建立对比模型,从备选模型中选出预测精度最高的对未来天然气消费量进行预测。最后,将BMA模型与ARIMA模型、ETS模型、BVAR模型、逐步回归模型以及等权重加权平均模型的预测精度进行对比。结果表明,最优的BMA模型包含了涉及经济水平、产业结构、能源结构、人口因素、天然气价格、天然气供给六个方面9个影响因素,其预测精度优于对比预测模型,且该模型预测 2022年天然气消费量将达到3254.153亿立方米,年均增长率为8%。

关键词: 贝叶斯模型平均, 影响因素, 模型选择, 情景分析, 天然气消费量预测