面对全球气候变暖及国际形势变化，核电重新受到多国重视，我国也提出积极安全有序发展核电。铀是目前核能利用的最主要材料，也是主流核电机型所需的燃料原料，同时兼有国家重要战略资源的性质，全球对铀的交易具有非常严格的限制并要求对铀产品实施严格的保障措施。国际铀价的预测研究是国际铀市场参与方进行国际贸易采销、铀矿项目开发与生产、铀浓缩生产方案制定等贸易与管理决策的有力抓手，是核电项目落地及后续运营成本的重要参考依据，是推动完善我国核燃料价格指数，增强我国在国际铀市场话语权的重要途径。国际铀价受多种因素共同影响，且影响因素难以准确识别或及时获取数据，铀价随着时间还呈现出一定的波动性，这给铀价预测带来极大的挑战。铀价预测研究包含着回归预测和时间序列预测。目前，市场参与方对国际铀价以判断预测为主、统计预测为辅，预测结果粒度较粗、精度较低，并不能很好的满足企业需求，在实际业务中的参考价值还需提高。随着机器学习算法大幅发展，新的算法被用来分析和预测时间序列数据。在前人铀价预测研究中，主要聚焦于预测模型构建与数据结果分析，对国际铀市场与国际铀价的分析较少；所使用的模型较少考虑除时间以外的其他影响因素指标，运用到的机器学习模型也较为传统。本文从调研国际铀市场、铀交易特点与限制，以及近两年在铀现货市场上引人关注的金融投资情况等方面出发，较为完整的介绍了国际铀价的形成机制和历史情况，并梳理了1998年1月-2022年9月的月度铀价数据集。为了更好的研究与预测国际铀价，选取国际铀现货价为实验研究对象，从直接供需关系、关联价格、经济与货币影响、其他能源可替代性等4个方面选取了相关24项影响因素的历史数据，形成了一套时间跨度较长、较为完善的数据集。本文设计实现了基于LSTM的铀价预测系统，该系统由三个模块组成：数据预处理、模型推断以及最终的模型应用模块；同时构建XGBoost预测模型、ARIMA预测模型进行比较分析。通过对三个预测模型的短期预测能力和长期预测能力进行算法评估，实验证明LSTM在短期、长期预测效果均优于XGBoost和ARIMA，其短期预测结果效果最佳，MAE为1.34$/lb U3O8，MAPE为5.26%。因此基于LSTM的铀价预测系统可以较好的应用在实际铀价预测工作中。
In the face of global warming and changes in the international situation, nuclear power has once again attracted the attention of many countries. China has also proposed its active, safe and orderly development of nuclear power. Uranium is currently the most important material in nuclear energy, the fuel of mainstream nuclear reactors, and is also one of key strategic resources. Trade in uranium is very restrictive and requires safeguards for uranium products. Accurate prediction of uranium price helps uranium market participants make scientific management decisions such as procurement and sales, uranium ore project production, and enrichment production plan. Nuclear reactor operators also need good prediction of uranium price to acquire new projects and balance operating costs. It is also an important way to promote the improvement of China’s nuclear fuel price index and enhance China’s voice in the international uranium market.Uranium price is affected by a variety of factors which are difficult to accurately identify or obtain in time. The uranium price also shows certain volatility over time, which brings great challenges to price forecasting. Uranium price forecasting studies include regression forecasting and time series forecasting. At present, market participants mainly judge and predict uranium prices, supplemented by statistical forecasts. The prediction results are coarse and low accuracy, which cannot well meet the needs of enterprises. With the dramatic development of machine learning algorithms, new algorithms are being used to analyze and predict time series data. In the previous uranium price prediction research, they mainly focused on the construction of prediction models and data result analysis. Because of less analysis of uranium market, they seldom considered factors other than time in their models. Moreover, they used traditional machine learning models in general.This paper investigates the uranium market, characteristics and restrictions of uranium trading, and financial investments of uranium in the past two years. This paper also introduces the formation mechanism and history of uranium prices and sorts out the monthly uranium price data sets from January 1998 to September 2022. In order to better understand and predict uranium price, uranium spot price is selected as the experimental research object. In this paper, the historical data of 24 influencing factors are selected from four aspects, including supply and demand, related price, economic and monetary impact, and other energy substitutions. Therefore, this paper builds a set of relatively complete data with a long-time span. In this paper, an LSTM-based uranium price prediction systemis designed, which consists of three modules: data preprocessing, model inference, and model application module. The XGBoost prediction model and the classical time series forecasting model ARIMA are constructed for comparative analysis. Through the algorithm evaluation of the short-term prediction and long-term prediction, the experimental results show that LSTM is better than XGBoost and ARIMA. LSTM short-term prediction results are the best, MAE is 1.34$/lb U3O8, MAPE is 5.26%, so the LSTM-based uranium price prediction system can be better applied to the actual uranium price prediction.
