迁移学习框架下高心墙堆石坝施工仿真参数IGOA-MLP动态预测模型

IGOA-MLP dynamic prediction model for simulation parameters of high core rockfill dam construction under transfer learning framework

施工仿真参数是影响高心墙堆石坝仿真结果准确性的关键。现有方法基于历史数据来预测未来填筑层的仿真参数,忽略了不同层之间的施工差异;同时,在新一层开始时往往存在数据不足或缺失的问题;此外,施工参数受到气象条件、机械运行状态等多因素影响而动态变化。本文利用迁移学习解决了上述问题,该方法具有通过知识迁移解决少样本建模问题的优势,同时考虑气象条件、机械运行状态等多种因素的定量影响,提出迁移学习框架下的高心墙堆石坝施工仿真参数改进蝗虫算法优化的多层感知机动态预测模型。首先,建立综合考虑多因素影响的施工仿真参数IGOA-MLP预测模型;其中,采用非线性缩减因子和柯西-高斯混合变异模式改进蝗虫优化算法(IGOA),并利用IGOA高效全局最优搜索能力来优化多层感知机(MLP)的超参数。其次,引入迁移学习策略,将训练集划分为源域和目标域,并在MLP隐藏层中增加自适应层以表征源域数据与目标域数据的差异性,实现历史工况和新工况间的知识迁移,从而解决新工况下缺少数据的问题。工程实例表明,相比于传统MLP模型以及未使用迁移学习的IGOA-MLP模型,本文所提方法的平均绝对百分比误差(MAPE)分别降低了54.68%、40.57%,证明了本文所提模型能够更准确地预测仿真参数,为仿真计算提供可靠的数据基础。

For construction simulation of high core rockfill dam,the parameters are the key to ensuring its accuracy.However,existing parameter prediction methods used historical data and ignore the differences between the construction processes of different layers,and there is often insufficient or missing data at the beginning of a new layer.In addition,the parameters are affected by many factors such as meteorological conditions and operating state of the machine.To solve the above problems,this paper takes advantage of the transfer learning’s capability of modeling with small samples through knowledge transfer and considers the quantitative influence of various factors.An improved multi-layer perceptron dynamic prediction model(IGOA-MLP)is proposed for construction simulation parameters of high core rockfill dam under the framework of transfer learning.Firstly,the IGOA-MLP prediction model is established that considering the influence of multiple factors.The grasshopper optimization algorithm is improved(IGOA)by nonlinear reduction factor and Cauchy-Gaussian hybrid mutation mode,and the efficient global optimal search capability of IGOA is utilized to optimize the hyperparameters of multi-layer perceptron(MLP).Secondly,the transfer learning strategy is introduced to realize the knowledge transfer between the historical and new conditions and solve the problem of insufficient or missing data in the new conditions.The training set is divided into source domain and target domain,and an adaptive layer is added to the hidden layer of MLP to represent the difference between source domain data and target domain data.A case study shows that compared with other machine learning methods such as MLP model and IGOA-MLP model without transfer learning,the mean absolute percentage error(MAPE)of the proposed method is reduced by 54.68%and 40.57%,respectively.It is proved that the proposed model can predict the parameters of construction simulation more accurately and provide a reliable data basis for simulation.