|
| 基于VSW－CDD－AuKELM的混凝土坝长期变形预测模型 |
| VSW－CDD－AuKELM-based Long-Term Deformation Prediction Model for Concrete Dams |
| 投稿时间：2025-06-06 修订日期：2026-02-19 |
| DOI： |
| 中文关键词: 混凝土坝长期变形 概念漂移检测 在线监控 可变滑动窗口 参数自适应更新 核极限学习机 |
| 英文关键词: Long－term concrete dam deformation Concept drift detection Online monitoring Variable sliding window Adaptive parameter update Kernel Extreme Learning Machine |
| 基金项目:国家重点研发计划项目资助(2022YFC3005404,2022YFC3005403)；西藏自治区科技重大专项资助(XZ202201ZD0003G)；中国电建集团核心攻关项目资助(DJ-HXGG-2022-02) |
|
| 摘要点击次数: 26 |
| 全文下载次数: 0 |
| 中文摘要: |
| 大坝变形的长效高精度预测是混凝土坝长期运行安全监控的关键。现有基于数据驱动的变形监控模型多利用历史观测序列训练模型架构，未能考虑长期服役过程中筑坝材料劣化或不可预知因素扰动导致的映射关系变化，存在预报时效短、难以满足大坝长期运行安全在线监控需求的问题。针对上述不足，在核机器学习机（Kernel Extreme Learning Machine，KELM）算法中嵌入概念漂移检测（Concept Drift Detection，CDD）策略，并将模型预测结果连续异常的频数定义为滑动窗口尺寸，提出可变滑动窗口（Variable Sliding Window，VSW）技术以增强概念飘移检测的适用性；进而，建立基于可变滑动窗口概念漂移检测自更新KELM（Adaptive Updating KELM based on CDD with VSW，VSW－CDD－AuKELM）的混凝土坝长期变形预测模型。该模型利用VSW改进的概念漂移检测技术实时监控KELM模型的预测精度，通过滑动窗口的方式自适应更新模型参数，以维持KELM模型的预测性能，从而实现混凝土坝长期变形的在线高精度预测。某混凝土拱坝案例研究及对比分析表明，所提模型具备抵抗映射规则概念漂移的能力，能够保持长时间的高精度预测性能；在相同预报周期下，平均变形预测均方根误差为0.575mm，较多元线性回归、极限学习机、支持向量机、长短期记忆网络等主流算法，预测精度提高了26%以上。 |
| 英文摘要: |
| Long term high－precision prediction of dam deformation is critical for the operational safety monitoring of concrete dams. Existing data-driven deformation monitoring models trained on historical data, neglecting the dynamic changes in the input–output mapping caused by material degradation or unforeseen disturbances during prolonged service, resulting in short prediction horizons that cannot support the long-term online safety monitoring of dams. To address these issues, the Concept Drift Detection (CDD) strategy is embedded into Kernel Extreme Learning Machine (KELM) algorithm, and the frequency of consecutive abnormal prediction results is defined as the sliding window size, thereby proposing a variable sliding window (VSW) technique to enhance the applicability of CCD strategy. Consequently, an adaptive updating KELM model based on CDD with VSW (VSW－CDD－AuKELM), is established for long－term deformation prediction of concrete dams. This model employs the VSW－enhanced CDD technique to monitor the prediction accuracy of the KELM model in real-time and adaptively updates the model parameters through the sliding window mechanism to maintain predictive performance, enabling online high－precision prediction of long-term dam deformation. A case study and comparative analysis involving a concrete arch dam demonstrate that the proposed model can resist concept drift in the mapping rules and maintain high－precision prediction over extended periods. Under identical forecast horizon, the model achieves an average Root Mean Square Error of 0.575 mm for deformation prediction, representing an improvement of over 26% in predictive accuracy compared to mainstream algorithms such as Multiple Linear Regression, Extreme Learning Machine, Support Vector Machine, and Long Short－Term Memory. |
|
查看/发表评论 下载PDF阅读器 |
| 关闭 |