摘要
合理的仓面排序方案对于加快工程进度和优化资源配置有着重要影响。然而,现有仓面排序方法将这一序贯决策问题简化,多数采用多属性决策方法,存在仅对大坝实时施工状态进行分析以及未考虑未来仓面浇筑方案对当前排序策略影响的问题;部分采用多目标优化方法进行仓面排序多目标优化问题分析,但主要是采用静态权重,存在忽略了仓面排序策略随环境动态变化的不足。针对以上问题,本文提出基于深度蒙特卡洛树搜索的拱坝仓面排序方法。首先,分析仓面排序问题的约束条件和目标函数,建立仓面排序强化学习模型;其次,针对仓面排序强化学习模型具有复杂且庞大的离散状态空间,为提高搜索效率,提出融合深度学习的蒙特卡洛树搜索方法,分别利用深度神经网络进行先验动作概率分布预测和策略函数评估;最后,以乌东德拱坝工程为例进行研究,结果表明本文方法可以有效地分析拱坝仓面排序问题,且相比于粒子群方法、证据理论方法,本文方法分析的施工工期可分别提前6天、14天,平均机械利用率分别提高1.19%、1.35%。本研究为拱坝仓面排序分析与优化提供了新思路。
Reasonable schemes of concrete placement sequencing have an important impact on accelerating construction progress and optimizing resource allocation.However,previous sequencing methods have simplified this sequential decision-making issue.Most of them adopt multi-attribute decision-making methods,which have the problem of analyzing only the real-time construction state of a dam and neglecting the influence of future concrete placing schemes on the current sequencing strategy;some adopt multi-objective optimization methods for analysis of the multi-objective optimization of the sequencing,but mainly using static weights and neglecting the dynamic changes in the sequencing strategy with the environment.To address these issues,a new concrete placement sequencing method for arch dams based on deep Monte Carlo tree search is presented.First,the constraints and objective function are examined,and a reinforcement learning model of the concrete placement sequencing for arch dams is developed.Then,for this learning model that demands a complex and large discrete state space,to optimize the sequencing strategy with better efficiency,we develop a new Monte Carlo tree search method combined with a deep neural network that is used for the priori action probability distribution prediction and strategy function evaluation.The case study of the Wudongde arch dam in China shows our method is effective in analysis of the sequencing.And compared with the particle swarm method and the evidence theory method,it shortens the construction period by 6 days and 14 days respectively,and raises the average mechanical utilization rate by 1.19%and 1.35%respectively.
作者
宋文帅
任炳昱
关涛
SONG Wenshuai;REN Bingyu;GUAN Tao(State Key Laboratory of Hydraulic Engineering Intelligent Construction and Operation,Tianjin University,Tianjin 300350,China)
出处
《水力发电学报》
CSCD
北大核心
2024年第3期120-130,共11页
Journal of Hydroelectric Engineering
基金
国家自然科学基金(52222907,52379131)。
关键词
拱坝
仓面排序
深度强化学习
蒙特卡洛树搜索
门控循环单元
arch dams
concrete placement sequencing
deep reinforcement learning
Monte Carlo tree search
gated recurrent unit
