水电能源科学

针对复杂分区土石坝渗透系数反演结果不唯一、算法复杂、计算量大等问题，采用正交设计方法拟定渗透系数组合，通过渗流有限元计算测点处的渗压水位组成学习样本，采用结构简单、计算量小、自学习和强泛化能力的误差反向传播(BP)神经网络构造渗压水位-渗透系数之间的非线性映射关系，为克服该网络算法易陷入误差函数的局部极小值、收敛速度慢、过拟合等缺点，利用和声搜索(HS)算法优化网络参数，从而构建了复杂分区土石坝渗透系数反演的HSBP网络算法。采用工程实例对该算法进行验证，结果表明该算法计算速度快、反演精度较高。

Aiming at the problems of non-uniqueness, complex algorithms, and large computational complexity in the inversion of permeability coefficients for complex zoned earth-rock dams, an orthogonal design method was used to determine combination of permeability coefficients. The seepage pressure water levels at the measurement point were calculated by finite element method to form learning samples. An error backpropagation(BP) neural network with simple structure, low computational complexity, self-learning, and strong generalization ability was used to establish a nonlinear mapping relationship between seepage pressure water levels and permeability coefficients. To overcome the shortcomings of this network algorithm, such as being prone to local minima of error functions, slow convergence speed, and overfitting, the harmony search(HS) algorithm was used to optimize the parameters of the network, and then an HSBP algorithm for complex zoned earth-rockfill dam permeability coefficient inversion was established.The algorithm was validated by an engineering example.The results show that the algorithm has fast calculation speed and high inversion accuracy.