大型洞室群稳定性与优化的并行进化神经网络有限元方法研究——第一部分:理论模型

RESEARCH ON PARALLEL EVOLUTIONARY NEURAL NETWORK FEM FOR STABILITY ANALYSIS AND OPTIMIZATION OF LARGE CAVERN GROUP——PART I : THEORY MODEL

提出并行进化神经网络有限元方法,通过有限元计算构造样本,用并行进化搜索到的神经网络学习并建立计算方案与地下洞室关键点最大位移和破损区体积之间的映射关系;随机产生一组初始方案,以关键点最大位移和破损区体积大小与参考值的差值比加权和作为评价指标,对该组方案进行遗传操作,产生下一代可行方案,由此进行下一步操作直至找到最优方案。利用自主开发的基于WINDOWS平台的并行计算环境(RsmVPC)来实现并行计算,该方法使得大规模优化问题在PC机群上实现了并行求解,提高了计算速度、规模与精度。

With ever enlarging scale of the underground cavern group, recognition of relevant parameters and scheme optimization would be highly nonlinear and of multi-extreme values in the solution space. It is much in need of finding an available method of global optimization and parallel computation. So a new parallel evolutionary neural network FEM is put forward. Through the sample construction by FEM calculating, the mapping relationship among the calculating schemes, the maximum displacement and the volume of damage zone is set up by parallel evolutionary neural network, and a group of initial feasible schemes are given by genetic algorithms(GAs). Evaluated by the maximum displacement of key spots and volume of damage zone, a group of new schemes are generated by operation of GAs. The operation is done until the reasonable scheme is found. The parallel computation is carried out on independently developed parallel environment(RsmVPC) based on WINDOWS platform. The methodology makes it possible to solve the large scale optimization problems parallelly on PC machine groups, and to improve the computing speed, scale and precision to a large extent with the proposed method.