耦合ALO-LSTM和特征注意力机制的土石坝渗压预测模型

Coupled ALO-LSTM and feature attention mechanism prediction model for seepage pressure of earth-rock dam

现有土石坝渗压预测模型缺乏对渗压效应量中各影响因子影响程度的定量分析,难以探究渗压效应量变化的内在影响机制。针对此问题,本文从时间维度和特征维度两个层面出发,提出一种耦合ALO-LSTM和特征注意力机制的土石坝渗压预测模型。该模型首先采用主成分分析方法对各个影响因子进行降维处理;然后,提出基于蚁狮优化算法(Ant Lion Optimizer,ALO)的长短时记忆网络(Long Short-Term Memory,LSTM),利用ALO中的随机游走以及轮盘赌等策略优化LSTM中的神经元数量,以捕捉渗压数据在时间维度上的深层信息;进一步地,在特征维度上引入特征注意力机制计算各影响因子对渗压效应量的贡献率,以自适应挖掘渗压效应量变化的内在原因。案例分析表明,本文所提模型具有较高的精度,其在测试样本上的MAPE、RMSE和MAE分别为0.28%、0.022 m和0.17 m;此外,水位分量对渗压效应量的贡献率最大,为47.9%;其次是降雨、温度和时效分量,其贡献率分别为33.5%、9.8%和8.8%。

The existing prediction model of seepage pressure of earth-rock dam lacks quantitative analysis of the influence degree of each influencing factor in seepage effect quantity,which makes it difficult to ex-plore the internal influencing mechanism of seepage effect quantity change.In view of the above prob-lems,this paper proposes a prediction model of seepage pressure of earth-rock dam by coupling ALO-LSTM and feature attention mechanism from the perspectives of time dimension and characteristic di-mension.The model firstly adopts principal component analysis to reduce dimension of each influencing fac-tor.Then,a long short-term memory network based on ant lion optimization algorithm is proposed to opti-mize the number of neurons in long short-term memory network by using random walk in ant lion optimiza-tion and roulette to capture the deep information of osmotic pressure data in time dimension.Furthermore,the feature attention mechanism is introduced in the feature dimension to calculate the contribution rate of each influencing factor to the osmotic effect volume,so as to find out the internal reasons for the change of the seepage effect volume adaptively.The case analysis shows that the proposed model has high accura-cy,and its MAPE,RMSE and MAE on test samples are 0.28%,0.022m and 0.17m,respectively.In addi-tion,the contribution rate of water level component to osmotic effect is 47.9%,followed by precipitation temperature and aging component,which are 33.5%,9.8%and 8.8%,respectively.