【摘要】 含噪多变量预测是环境、交通、工业等领域的共性难题，其核心挑战在于平衡噪声过滤与多尺度特征捕捉。提出一种基于卡尔曼滤波-长短期记忆网络-Transformer的混合模型（Kalman-LSTMTransformer）。该模型在过滤噪声的同时捕捉局部时序与全局依赖，并结合贝叶斯优化以实现高效精准预测。以露天矿粉尘浓度预测为案例验证，基于1年监测数据的实验显示，该模型较基准模型的均方根误差降低21.70%～27.19%，平均绝对误差降低6.68%～18.30%，决定系数R~2达0.934；消融实验与超参数分析结果进一步证实了各模块有效性。该模型可迁移至同类场景，为多领域智能预警与精准治理提供支撑。更多还原

【Abstract】 Noisy multivariate prediction is a common challenge in fields such as environmental science, transportation, and industry. The core difficulty lies in balancing noise filtering with multi-scale feature capture. To address this, a hybrid model(Kalman-LSTMTransformer) based on Kalman filter, long short-term memory(LSTM), and Transformer is proposed. This model captures local temporal and global dependencies while filtering noise, and integrates Bayesian optimization to achieve efficient and accurate prediction. Using open-pit mine dust concentration prediction as a case study, experiments based on 1-year of monitoring data demonstrate that the model outperforms baseline models, reducing the root mean square error(RMSE) by 21.70%–27.19% and the mean absolute error(MAE) by6.68%–18.30%, while achieving a coefficient of determination(R~2) of 0.934. Ablation experiments and hyperparameter analysis results further confirm the effectiveness of each module. The model exhibits transferability to similar scenarios, providing support for intelligent early warning and precision management across multiple domains.更多还原