军事知识服务平台
数字孪生驱动的铜锍品位和温度时序预测方法
Digital twin-driven time series prediction method for matte grade and temperature
摘要 针对现有铜锍品位和温度预测方法存在的动态特性与非线性影响考虑不足、理论研究与工程需求脱节、数字孪生技术系统性研究不足等问题,本研究提出Transformer-KAN预测方法.该预测方法以数字孪生框架下的虚实同步映射为基础,利用Transformer编码器的多头自注意力机制提取多变量工艺参数间的长期时序依赖与全局交互特征,同时引入KAN网络,以动态可学习的B样条基函数作为激活单元,替代传统Transformer中的固定激活前馈网络,从而增强模型对铜熔炼过程复杂非线性耦合关系的精细化校正与拟合能力,有效提升了多目标预测的精度与鲁棒性.与其他预测模型的对比实验结果表明,该方法在铜锍品位预测中R²达到0.933,优于其他对比模型;在铜锍温度预测中RMSE为1.450、R²为0.952,综合性能最优.研究还基于此方法构架了物理实体层-数据感知层-孪生模型层-功能服务层模型框架,在云南某铜冶炼企业完成部署进行了工业现场应用,从工艺参数采集、数据传输、预测模型到Unity界面更新,整个流程的总耗时满足工业现场实时监测的需求.所提方法可实现虚实熔炼过程的协同交互与优化迭代运行,为铜锍品位和温度预测提供了解决思路和技术支撑.
Abstract:
To address the issues of existing copper matte grade and temperature prediction methods,such as insufficient consideration of dynamic characteristics and nonlinear effects,a disconnect between theoretical research and engineering needs,and a lack of systematic research on digital twin technology,this study proposed a Transformer-KAN prediction method.Based on the virtual-real synchronous mapping within a digital twin framework,this prediction method leverages the multi-head self-attention mechanism of the Transformer encoder to extract long-term temporal dependencies and global interaction features among multi-variable process parameters.At the same time,it introduces a KAN network,using dynamically learnable B-spline basis functions as activation units to replace the fixed activation feedforward network in a traditional Transformer,thereby enhancing the model's ability to finely calibrate and fit the complex nonlinear coupling relationships of the copper smelting process.This effectively improves the accuracy and robustness of multi-objective prediction.Comparative experimental results with other prediction models show that this method achieves an R² of 0.933 for copper matte grade prediction,outperforming other comparison models.In copper matte temperature prediction,it attains an RMSE of 1.450 and an R² of 0.952,demonstrating the best comprehensive performance.Furthermore,based on this method,the study constructed a model framework comprising a physical entity layer,a data perception layer,a twin model layer,and a functional service layer.This framework was deployed and applied in an industrial field at a copper smelting enterprise in Yunnan Province.The total processing time,from process parameter collection and data transmission to prediction model execution and Unity interface updates,meets the real-time monitoring requirements of the industrial site.The proposed method enables collaborative interaction and optimized iterative operation of the virtual and real smelting processes,providing a solution and technical support for predicting copper matte grade and temperature.