关键词: 冰水堆积物；渗透系数；颗粒级配；DBO-CRNN 神经网络；渗透试验；夏曲水电站

Abstract: Ice - water accumulations are characterized by a wide range of particle sizes and heterogeneous compositions, which significantly influence their permeability properties, consequently, the safe operation of hydraulic and hydropower projects. This study focused on the ice - water accumulations in the Xiaqu Hydropower Station, Yigong - Zangbu Basin. A total of 20 sets of indoor constant - head permeability tests were conducted, and an empirical formula was developed to calculate the permeability coefficient, taking the gradation area into consideration. Based on these results, a Generalized Regression Neural Network (GRNN) model optimized by Dung Beetle Optimization (DBO) was constructed, with characteristic particle sizes  and gradation area (S) as input variables to predict the permeability coefficient of the ice - water accumulations. Four sets of field single - ring permeability tests were then carried out to verify the accuracy of the DBO - CRNN model. The results showed that the predicted permeability coefficients from this model were in excellent agreement with the experimental values, with an error margin of less than 5%. In contrast, the errors between the predictions of the empirical formula and the traditional BP neural network model and the test values reached up to 61.29% and 37.50%, respectively. These findings demonstrate that the DBO - CRNN model can accurately estimate the permeability coefficient of ice - water accumulations.

Key words: ice - water accumulation; permeability coefficient; particle size distribution; DBO - CRNN neural network; permeability test; Xiaqu Hydropower Station