铀资源是保障核电稳定运行的物质基础, 近年来中国核电快速发展, 对铀资源需求日益增加, 需建立可靠的铀价预测系统, 为企业海外投资及采购提供有效参考。该文以中长期铀价预测为目标, 从基本供需面出发, 设计了基于政策或事件的供需决策量化方法, 邀请5位铀资源行业专家对105个铀市场中具有代表性的政策或事件进行评价, 对评价结果的量化和分析表明选用均值化处理方式可以消除个人偏好干扰; 提出了描述供需决策与铀价关系的响应模型, 利用多分量时移指数衰减模型进行验证, 证明了供需决策响应比供需数据更适用于铀价预测; 利用全连接网络的逼近能力, 学习和表达“决策-价格”响应特征, 基于响应叠加建立了中长期铀价预测模型。将均值化处理的供需决策量化结果作为输入信息, 该模型在测试集上的中长期铀价预测精度优于模型融合方法, 可以为铀价预测和市场交易提供有效参考。

Objective: Uranium resources are the substantial foundation for ensuring the stable operation of nuclear power. In recent years, China's nuclear power has developed rapidly and the demand for uranium resources is increasingly growing. Hence, a reliable uranium price prediction system needs to be established to provide an effective reference for enterprise's overseas investment and procurement. This paper proposes a response superposition model (RSM) to depict the relationship between requirement and demand decisions and uranium price. Then the fully connected network (FCN) is combined with RSM to achieve a better forecast performance than model fusion methods. Forecasting uranium prices by data-driven models is in vogue because of less dependent on mathematical models and more precise than traditional methods. However, data-driven methods are highly rely on datasets and difficult to illustrate due to their black-box mechanism. Methods: This paper solves this problem by two major approaches. First, the covariance variables are split into supply, demand, and financial. A window-based correlation estimation applied to those covariance variables shows the forecast character of different covariance variables. First, the supply and demand covariance variables stably influence the uranium price over a long period. Second, the financial covariance variables are highly correlated with uranium price but significantly decrease or are reversely correlated in several months. From those results, supply and demand data are more suitable for medium and long-term uranium price forecasting. To obtain a quantified supply and demand dataset from the major events, policies, and accidents, five uranium resource industry experts are invited to evaluate the questionnaires. Then the questionnaires are quantified and analyzed by mathematical methods and the results are averaged to eliminate personal preference interference. Second, the RSM is proposed to describe the relationship between supply and demand decisions and uranium prices. The multi-component time-varying exponential decay model is used for verification, which proves that supply and demand decision response is more suitable for uranium price prediction than supply and demand data. To improve the forecasting capability, the decision-price-response characteristics are learned and expressed by utilizing the approximation capability of fully connected networks (FCN). Then a medium and long-term uranium price forecasting model is established by the RSM. Finally, the average quantified decision results are used as input, and the uranium price is forecasted by FCN-RSM. Results: The FCN-RSM's mean absolute percentage error (MAPE) for forecasting the future 6 months, 12 months, 18 months, and 24 months is 12.6, 6.98, 12.9, and 26.1, respectively. In comparison, the model fusion mothod's better MAPE is 9.27, 13.8, 35.6, and 32.0, which proves that the FCN-RSM is more suitable for medium and long-term uranium price forecasting. From the perspective of the forecasted trend, the FCN-RSM can follow a rapid increment or decrement better than the pure data-driven model by analyzing the decision result. Conclusions: By using the RSM to describe uranium price, the medium and long-term uranium price forecasting precision is remarkably increased compard with the model fusion method on the test dataset. The FCN-RSM can achieve medium and long-term uranium price forecasting goals and provide an effective reference for uranium price forecasting and market transactions.