水电能源科学

混凝土坝变形的长期预测是实际运行中保持其结构完整性的重要要求。为提高混凝土长期变形预测的准确性，构建了基于多层感知机(MLP)和编码器—解码器(Ecoder-Decoder)架构的大坝变形长期预测模型MLP-Ecoder-Decoder(MED),该模型通过深度自动相关(Deep-Auto-Correlation)机制捕获大坝变形与环境荷载的长期依赖性，利用时间序列分解和深度自动相机制进行多步变形预测，并采用该模型对青海省的一座250 m高拱坝在复杂环境条件下的变形进行预测。结果表明，MED模型有效提升了预测精度，在提取长期时间特征方面具有强大的优势。

The long-term prediction of concrete dam deformation is an important requirement for maintaining its structural integrity during actual operation. To improve the accuracy of long-term deformation prediction of concrete, a long-term dam deformation prediction model based on multi-layer perceptron(MLP) and ecoder-decoder(Ecoder-Decoder) architecture, MLP-Ecoder-Decoder(MED), was constructed. This model captured the long-term dependence of dam deformation and environmental loads through a deep auto-correlation(Deep-Auto-Correlation) mechanism, and used time series decomposition and deep auto-correlation mechanism for multi-step deformation prediction. The model was used to predict the deformation of a 250 m height arch dam in Qinghai Province under complex environmental conditions. The results show that the MED model effectively improves the prediction accuracy and has a strong advantage in extracting long-term time features.