波致海床瞬态液化对再悬浮贡献率的深度学习预测方法

PREDICTION METHOD FOR CONTRIBUTION RATE OF WAVE-INDUCED SEABED TRANSIENT LIQUEFACTION TO RESUSPENSION BASED ON DEEP LEARNING

波致瞬态液化渗流导致海床内细粒沉积物向海水中运移,这一过程对海底沉积物再悬浮的贡献率不容忽视,但是贡献率的准确估计和预测比较困难。本研究将黄河水下三角洲的观测数据(包括水深、有效波高、有效波周期、实验舱内悬沙浓度、实验舱外悬沙浓度)作为模型输入数据集,基于长短时记忆循环神经网络建立了瞬态液化对再悬浮贡献率的深度学习预测模型。为了客观评价模型的性能,以平均绝对百分比误差、均方根误差和平均平方误差-标准偏差为评判标准,将该深度学习模型与其他预测模型(支持向量回归模型、人工神经网络)的预测结果进行了比较。结果表明,基于长短时记忆循环神经网络的深度学习模型对3.5d以内的瞬态泵送再悬浮贡献率预测误差最小,其平均绝对百分比误差、均方根误差和平均平方误差-标准偏差分别为5.87%、1.6730、0.1574。因此,该模型可以有效地减少机器学习方法在连续预测中产生的误差叠加问题。

Wave-induced transient liquefaction seepage leads to the transport of fine-grained sediments from the seabed to seawater.The contribution of wave-induced transient liquefaction to the sediment resuspension cannot be ignored,and the accurate prediction of the contribution is difficult.The observation data in the Yellow River subaquatic delta include water depth,significant wave height,significant wave period,suspended sediment concentration in the benthic chamber,and suspended sediment concentration out of the benthic chamber.These data are treated as the input data sets.Then,a deep learning model of transient liquefaction contribution to resuspension is developed based on a long short-term memory recurrent neural network.To evaluate the performance of the model,the prediction results of the deep learning model based on LSTM are compared with other models.These models include Support Vector Regression and Artificial Neural Network,and have mean absolute percentage error(MAPE),root mean square error(RMSE)and mean squared error-standard deviation(RSR).The results show that the LSTM model has the smallest error for transient pumping resuspension contribution within 3.5 days,with the mean values of MAPE,RMSE,and RSR of 5.87%,1.6730,and 0.1574,respectively.Therefore,the LSTM model can effectively reduce the error superposition problem arising from machine learning methods in continuous forecasting.