基于深度可分离卷积混合网络模型的地浸采铀注液量预测研究  

Injection Volume Prediction for In-situ Leaching of Uranium Based on Depthwise Separable Convolution Mixer Network Model

作　　者：刘志锋[1,2,3] 唐俊贤 林芝宁 周义朋 LIU Zhifeng;TANG Junxian;LIN Zhining;ZHOU Yipeng(State Key Laboratory of Nuclear Resources and Environment,East China University of Technology,Nanchang 330013,China;National Key Laboratory of Uranium Resources Exploration-Minning and Nuclear Remote Sensing,Nanchang 330013,China;East China University of Technology,Nanchang 330013,China)

机构地区：[1]东华理工大学核资源与环境国家重点实验室,江西南昌330013 [2]铀资源探采与核遥感全国重点实验室,江西南昌330013 [3]东华理工大学,江西南昌330013

出　　处：《铀矿冶》2025年第1期9-17,共9页Uranium Mining and Metallurgy

基　　金：中国铀业有限公司-东华理工大学核资源与环境国家重点实验室联合创新基金(2022NRE-LH-14);国家国防科技工业局核能开发项目“铀裂变瞬发n-γ融合测井及航空监测关键技术研究”;江西省自然科学基金(20242BAB25084)。

摘　　要：地浸采铀作为铀矿的绿色开采技术,在生产运行中产生海量数据,利用这些海量数据进行大数据分析和趋势预测,能够提升技术人员制定生产计划的可靠性。目前采用的基于编码器-解码器结构的时序预测模型,由于存在注意力机制,导致计算复杂、内存消耗大。本研究提出深度可分离卷积混合模型,通过动态序列分割模块降低固定分割带来的语义破坏,通过深度可分离卷积混合模块降低模型运行时间并捕获局部和全局特征。结果表明,深度可分离卷积混合网络模型的均方误差(Mean Square Error,MSE)与平均绝对误差(Mean Absolute Error,MAE)相较于时间序列分块自注意力模型(Patch Time Series Transformer,PatchTST)分别降低了1.04%和4.13%,提出的动态序列分割模块的MSE与MAE相较于原有模型分别降低了7.32%和5.03%;在性能对比分析上,深度可分离卷积混合模型的训练速度相较于趋势季节分解线性模型(Decomposition Linear,DLinear)提高了59.91%。建立的模型能够准确预测采区生产运行中硫酸注液量的变化趋势,改善了现有预测模型针对地浸铀矿数据集存在的运行时间长、运行内存大、数据拟合差的问题,可为地浸铀矿生产决策提供理论和实践参考。In-situ leaching of uranium,as a green uranium mining technology,generates massive data in production operation,which are available for the big data analysis and trend prediction to improve the reliability of technicians in making production plans.In the current prediction algorithms,the attention mechanism in the temporal prediction model based on the encoder-decoder structure has the problems of computational complexity and high memory consumption.In this paper,we proposed a depthwise separable convolutional model,in which the semantic damage caused by fixed segmentation was reduced by the dynamic sequence segmentation module,and the depthwise separable convolutional mixer module was used to reduce the model running time and capture local features as well as global features.The results show that the Mean Square Error(MSE)and Mean Absolute Error(MAE)of the depthwise separable convolutional hybrid network model are reduced by 1.04%and 4.13%respectively,compared with Patch Time Series Transformer(PatchTST),and the proposed dynamic sequence segmentation module MSE and MAE are reduced by 7.32%and 5.03%respectively,compared to the original model;in the comparative performance analysis,the training speed of this model is 59.91%faster than the Trend Seasonal Decomposition Linear(Decomposition Linear,DLinear)model.The depthwise separable convolutional model can accurately predict the future trend of sulfuric acid injection volume in the production operation of the mining area,improve the existing prediction model for in-situ leaching data by solving the problem of long running time,large running memory,poor data fitting problems,which provide a theoretical and practical reference for the decision-making of in-situ leaching production.

关 键 词：地浸采铀 注液量预测 深度可分离卷积 预测模型 

分 类 号：P631[天文地球—地质矿产勘探] TP39[天文地球—地质学] TL212[自动化与计算机技术—计算机应用技术]