基于模型融合的国际短期天然铀价格预测研究  

International short-term natural uranium price forecast based on model fusion

作　　者：孙若凡 SUN Ruofan(Institute of Nuclear Energy and New Energy Technology Tsinghua University,Beijing 100084,China;China National Nuclear Corporation Overseas Limited,Beijing 100013,China)

机构地区：[1]清华大学核能与新能源技术研究院,北京100084 [2]中核海外有限公司,北京100013

出　　处：《世界核地质科学》2024年第4期712-719,共8页World Nuclear Geoscience

摘　　要：天然铀是核电发展的重要矿产物质基础,对核电发展速度和运行成本具有重要影响,因此国际天然铀价格(简称天然铀价格)预测工作对核电发展有重要的应用价值。数据驱动模型是目前效果较好的价格预测方法之一,通过分析历史天然铀价格及其协变量数据特征,可以预测得到未来天然铀价格的趋势和具体数值。一般而言,为了提升数据驱动模型的预测效果,需要从数据内容和模型结构两方面着手进行改进。首先,深入分析影响天然铀价格的协变量因素,将天然铀价格协变量因素分为基本供需关系、市场交易情况和经济金融环境三类,选择较为完善的协变量数据构建天然铀价格数据库,通过相关性检验验证有效性。然后,根据协变量时序数据的时间维度自相关和变量间互相关特点,基于机器学习的模型融合方法,提出在模型中引入卷积神经网络(Convolutional neural network,CNN)深入挖掘互相关和自相关特征,以改善预测模型对数据特征分析能力。最后,以长短期记忆(Long and short-term memory,LSTM)网络为时间序列预测的基础,建立5种不同的模型融合结构并预测未来天然铀价格,结合敏感性测试方法进行分析对比,发现采用CNN插入的LSTM-CNN-LSTM模型的天然铀价格预测效果最好且受超前预测步和时间窗的影响较小。研究表明,通过模型融合和敏感性测试方法构建的优化数据驱动预测模型,可以稳定、准确地预测未来6个月的短期天然铀价格,为核电发展和运行提供数据参考。Natural uranium is the indispensable mineral material foundation for the development of nuclear power,which significantly impacts the development growing speed and operating costs of nuclear power.Hence,the international natural uranium price(abbreviated as natural uranium price)forecast work has important application value for developing nuclear power.The data-driven model is one of the currently most effective price prediction methods.It can forecast the trend and specific value of future natural uranium prices by analyzing the characteristics of past natural uranium prices and covariate data.Generally,the improvement should commence from two aspects of data content and model structure to promote the data-driven model’s forecast performance.Firstly,an in-depth analysis was conducted on the covariate factors that affect natural uranium prices.The covariate factors of natural uranium prices were divided into three categories:basic supply-demand relationship,market transaction situation,and economic and financial environment.A natural uranium database was constructed by selecting relatively comprehensive covariate data,and the validation was verified by correlation tests.Then,based on the covariate datas autocorrelation on time dimension and crosscorrelation between covariables,the convolutional neural network(CNN)was introduced,based on the model fusion method of machine learning,to perfect the forecast models ability to analyze data features.Finally,five different model fusion structures were formed on the basis of the long and short-term memory(LSTM)network,and the future prices were forecasted respectively.The comparison of sensitivity testing results revealed the CNN-inserted LSTM-CNN-LSTM model has better natural uranium price forecast performance and is less susceptible to the ahead forecast steps and time window.This research demonstrates the short-term natural uranium prices of the next six months can be stably and precisely forecasted through the data-driven model constructed by model fusion and sensit

关 键 词：国际天然铀 价格预测 模型融合 供需关系 

分 类 号：P619.14[天文地球—矿床学] F830.9[天文地球—地质学]