水电能源科学

大坝位移的精确预测是保障大坝安全运行的关键技术手段，然而传统预测方法难以精准捕捉位移数据的复杂性与非线性特征。为此，提出一种融合改进Inception模块、残差注意力机制与Informer模型、引力搜索算法(GSA)的混合位移预测模型。首先通过改进Inception模块并行采用多尺寸卷积核，实现对大坝位移影响因子的多尺度特征提取；再结合残差注意力机制与Informer模型的稀疏自注意力策略，强化模型对长时序数据依赖关系的刻画能力；然后引入引力搜索算法对模型超参数进行全局优化，提升模型泛化性能；最后以某拱坝为例将所提模型与HTT、ELM、LSTM等传统及机器学习模型进行对比验证。结果表明，所提混合模型各项指标均显著优于对比模型，在大坝位移预测中具有显著的优越性与可靠性，可为大坝安全监测提供技术支撑。

Precise prediction of dam displacement is a critical technical means for ensuring the safe operation of dams. However, the traditional prediction methods are difficult to accurately capture the complexity and nonlinear characteristics of displacement data. To address this, this paper proposes a hybrid prediction method integrating an improved Inception module, a residual attention mechanism, Informer model, and gravitational search algorithm(GSA). Firstly, the improved Inception module utilizes parallel multi-sized convolutional kernels to achieve multi-scale feature extraction of displacement influencing factors. By combining the residual attention mechanism with the ProbSparse self-attention strategy of the Informer model, the proposed method strengthens the capability to characterize dependencies in long-term time-series data. Furthermore, the GSA is introduced to globally optimize model hyperparameters, thereby enhancing generalization performance. Taking an arch dam as an engineering case study, the proposed model is validated against traditional and machine learning models, including HTT, ELM, and LSTM. The results demonstrate that the proposed model significantly outperforms the comparison models across all evaluation metrics. The proposed hybrid model demonstrates significant superiority and reliability, providing technical support for dam safety monitoring.