2021年11月，近200个国家在联合国气候变化大会(COP26)上达成协议，将在逐步减少化石燃料的使用。会上达成的《格拉斯哥气候公约》的目标是保持将全球气温上升限制在1.5oC以内的希望。再加上各国采取了更多行动，1.5oC目标仍然是可以实现的，但前提是每个国家都兑现承诺。IPCC第一工作组第六次评估报告指出，大气温室气体(GHG)浓度增长主要是由化石燃料和农业(以畜牧业为主)部门的排放驱动的，并发现每年全球肉类消费引起的温室气体排放与全球使用核能等清洁能源所避免的温室气体排放相当。此外，过多肉类摄入对人类健康有明显的负面影响。因此促进肉类消费减少和转换营养结构至关重要。要实现此目标，就必须要研究肉类消费的主要驱动因素和对未来肉类需求情况有准确的判断。本研究综述了现有肉类供需研究模型和肉类消费驱动因素，建立了多种机器学习模型，结合FAO 和UN数据库，全面分析了全球肉类消费的驱动因素影响力在时间序列上的变化趋势，并对未来全球肉类需求情况做出了定量预估。本文得到以下初步结论：

（1）现有的肉类供需研究模型工具，可根据其模拟依据分为两类：1)基于时间序列数据的模拟模型：不考察系统演化动力，仅通过肉类产量或消费量的时间序列数据对未来进行预测。2)基于系统演化动力的模拟模型：通过对肉类供需系统进行系统演化动力归因，通过肉类消费量和生产量的影响因素数据对消费量和生产量未来数据进行预测。其中基于时间序列数据的模型可再分为时间序列模型，灰色系统模型和机器学习模型；基于系统演化动力的模型主要有局部均衡模型。肉类消费的驱动因素主要有经济社会因素、价格因素、自然因素、文化因素、人口因素和全球化因素几大类。

（2）通过对比现有模型工具的模拟效果，发现机器学习类模型的模拟效果最好。通过建立不同类型机器学习模型，对比其模拟性能后，发现随机森林(RF)模型在各项评估指标中均显现出优于其他种类机器学习模型的性能。

（3）在肉类消费的主要驱动因素分析中，我们将机器学习模型与FAOSTAT统计数据，UN Database数据和其他不同数据源相结合，发现肉类消费主要由经济社会因素驱动，但其影响力在逐渐减弱。不同肉类消费的驱动因素差别很大，猪肉、禽肉等大消耗量肉类主要由经济社会因素驱动，其中猪肉消费受宗教影响明显，全球化因素在禽肉消费中的作用越来越显著。牛羊肉的消费量相对较小，其中牛肉消费受城市化的影响很大，羊肉消费主要受自然因素影响。没有充足证据支持之前研究声称的GDP与肉类消费量之间存在的“倒U型关系”。经过历史数据验证，模型对国家肉类消费预测的百分误差基本在15%以下。

（4）基于SSP-RCP组合未来情景路径，对全球约170个国家2020-2100年的肉类消费需求进行了定量估计，发现在绝大多数路径下，未来全球肉类需求量将出现下降，这一下降趋势主要是由东亚和太平洋地区贡献。作为历史上最大肉类消费者，西方国家的肉类需求量极大可能在未来持续保持增长。本文认为促进西方国家居民的消费行为转变，加强国际合作和改善农业生产技术是降低农业部门温室气体排放以缓解气候变化的有效方法。

In November 2021, almost 200 countries reached an agreement at COP26 to phase down the use of all fossil fuels across the energy sector, especially, coal power. The aim of the achievement was to keep alive the hope of limiting the rise in global temperature to 1.5oC, and the Glasgow Climate Pact does just that. Combined with increased ambition and action from countries, 1.5oC remains in sight, but it will only be achieved if every country delivers on what they have pledged. The IPCC Working Group I Sixth Assessment Report states that the increase in atmospheric greenhouse gas (GHG) concentration is mainly driven by emissions from fossil fuels and the agricultural (mainly livestock) sector, and finds that the annual greenhouse gas emissions averted by the world's population ate less meat are comparable to those avoided by the global use of nuclear energy. In addition, excessive meat intake has a clear negative impact on human health. Therefore, it is very important to reduce meat consumption and promote nutritional transition. To achieve this goal, it is necessary to study the main drivers of meat consumption and make accurate evaluation for future meat demand scenarios. Our study reviews the current meat supply-demand simulation methods and the factors underlying meat consumption. Besides, different machine learning methods are established and combined with FAOSTAT and UN Databases to comprehensively analyze the driving factors. Finally, we make quantitative forecasts for the future global meat demand scenarios. Our study draws the following preliminary conclusions:

(1) Current simulation tools for meat supply-demand research can be divided into two categories according to their theory: 1) Simulation models based on time series data: the future simulation is made only through the time series data of meat production or consumption without investigating the evolutionary dynamics of the system; 2) Simulation model based on system evolutionary dynamics: Through the system evolutionary dynamics attribution of meat supply-demand system. The model based on time series data can be subdivided into three types: time series model, grey system model and machine learning model. There are mainly traditional partial equilibrium models based on system evolutionary dynamics. The driving factors of meat consumption mainly include socioeconomic factors, price factors, natural factors, cultural factors, demographic factors and globalization factors.

(2) By comparing the simulation performance of existing methods, the machine learning methods has the best simulation performance. By establishing different types of machine learning methods and comparing their simulations, we identified that the random forest (RF) algorithm has better performance than other types of machine learning methods evaluated under all evaluation indicators.

(3) In the analysis of the main driving factors of meat consumption, we combined the machine learning methods with FAOSTAT, UN Database data and other data sources, and found that total meat consumption is mainly driven by socioeconomic factors, but its influence is gradually decreasing by time. The main driving factors of different meat consumption vary widely. Pork and poultry consumption is mainly driven by socioeconomic factors. Pork consumption is obviously influenced by religion, and globalization factors are playing an increasingly important role in poultry consumption. The consumption quantity of beef and mutton is comparatively small, among which beef consumption is greatly affected by urbanization, while mutton consumption is mainly affected by natural factors. In another hand, there is insufficient evidence to support the "inverted U-shaped relationship" between GDP and meat consumption claimed by previous studies. Verified by historical data, the relative error of the model for predicting national meat consumption is basically below 15%.

(4) Based on the SSP-RCP combination future pathways, we estimate the meat demand scenarios of around 170 countries in 2020-2100. We find that the global meat demand will decrease under most of development pathways, and the decline trend is mainly contributed by East Asia & Pacific region. As the largest consumers of meat in history, demand of meat in the Western Countries is likely to continue to grow. In our view, it is considered that promoting the change of consumer behavior in western countries, improving international cooperation and promoting agricultural production technology are the methods to reduce greenhouse gas emissions from agricultural sector and mitigate climate change.