为提高基坑降水水位预测精度，加强对基坑降水水位的控制，在标准鹈鹕优化算法（POA）的基础上引入Logistic混沌映射、自适应动态权重因子w及萤火虫扰动最优位置3种优化策略，提出改进鹈鹕优化算法（IPOA）。选取4种基准测试函数对IPOA算法进行仿真测试，并选用其他5种优化算法进行对比分析。以沈阳北站基坑降水工程为实例进行研究，采用IPOA算法优化CNN-BiGRU模型中的隐含层节点个数、初始学习率以及L2正则化系数，形成IPOA-CNN-BiGRU水位预测模型，选用GWO、WOA、POA算法优化CNN-BiGRU模型并搭建Midas GTS NX基坑降水有限元数值模拟模型，进行对比分析，验证本文提出的预测模型在水位预测任务中的优越性。结果表明： 1）IPOA优化算法在单峰及多峰函数的寻优任务中均体现出较好的性能，其在4种基准测试函数中的寻优速度更快，寻优结果要明显优于其余5种算法。2）在水位预测任务中，IPOA-CNN-BiGRU水位预测模型的EMAE、ERMSE、EMAPE分别为0.213 8、0.253 9、0.771
3，预测精度比GWO、WOA、POA优化下的CNN-BiGRU模型分别提高了32.36%、20.96%、15.79%。3）相比Midas GTS
NX有限元模型，IPOA-CNN-BiGRU模型的参数依赖性更低，其预测精度提高了80.65%，验证了采用IPOA-CNN-BiGRU模型对基坑降水水位进行预测的可行性。
To enhance the prediction
accuracy of foundation pit dewatering level and strengthen the control of
dewatering level, three optimization strategies, Logistic chaos mapping,
adaptive dynamic weighting factor w and firefly disturbance optimal
position, are introduced to improve standard pelican optimization algorithm
(POA). Four benchmark test functions are selected to simulate and test the
improved POA (IPOA), and five optimization algorithms are selected for
comparison and analysis. Taking a foundation pit in Shenyang North Railway
Station as an example, the IPOA algorithm is used to optimize the number of
nodes in the hidden layer, the initial learning rate, and the L2 regularization
coefficient in the convolutional neural networks (CNN)-bidirectional
gated recurrent unit (BiGRU) model to form the IPOA-CNN-BiGRU water level prediction model. Additionally, the grey wolf
optimizer (GWO), whale optimization algorithm (WOA), and POA are selected to
optimize the CNN-BiGRU model and build the Midas GTS NX
finite element numerical simulation model. The comparative results among these
models in predicting dewatering level of the foundation pit reveal the
following: (1) The IPOA optimization algorithm shows better performance in both
single-peak and multi-peak
function optimization tasks, and its optimization speed is the fastest in the
four benchmark functions, and the optimization results are significantly better
than the remaining five algorithms. (2) In the water level prediction task, the EMAE, ERMSE, and EMAPE of the IPOA-CNN-BiGRU
model are 0.213 8, 0.253 9, and 0.771 3, respectively, and the prediction
performances are 32.36%, 20.96%, and 15.79% higher than those of the CNN-BiGRU model under GWO, WOA, and POA optimization, respectively. (3)
The IPOA-CNN-BiGRU model has a
low dependence on the parameters compared with the Midas GTS NX finite element
model, and its prediction accuracy is improved by 80.65%, which verifies the
feasibility of the IPOA-CNN-BiGRU
model in predicting dewatering level of the foundation pits.
