| 王艳丽*,陈远**.基于联合特征筛选与混合深度学习框架的项目成本估算方法[J].高技术通讯(中文),2025,35(10):1108~1119 |
| 基于联合特征筛选与混合深度学习框架的项目成本估算方法 |
| Project cost estimation method based on joint feature selection and hybrid deep learning framework |
| |
| DOI：10. 3772 / j. issn. 1002-0470. 2025. 10. 008 |
| 中文关键词: 成本估算； 混合深度学习框架； 时序卷积网络-双向门控循环单元； 联合特征分析； 参数优化 |
| 英文关键词: cost estimation, hybrid deep learning framework, temporal convolutional network-bidirectional gated recurrent unit, joint feature analysis, parameter optimization |
| 基金项目: |
| 作者 | 单位 | | 王艳丽* | （*河南建筑职业技术学院工程管理系郑州 450064）
（**郑州大学土木工程学院郑州 450001） | | 陈远** | |
|
| 摘要点击次数: 21 |
| 全文下载次数: 39 |
| 中文摘要: |
| 可靠的施工成本是项目规划和资源分配的关键。针对建筑工程领域高维数据冗余和时变数据的异构造成施工成本难以准确估算的问题，本研究提出了一种基于联合特征筛选的混合深度学习框架，融合了多层感知机（multi-layer perceptron， MLP）、时序卷积网络（temporal convolutional network ，TCN）和双向门控循环单元（bidirectional gated recurrent unit，BiGRU），使用人工智能的方式用于项目成本的估算。首先，设计了基于Pearson-Spearman联合检验的双维度特征筛选机制，去除冗余特征，降低数据噪声。其次，构建具有并行架构的MLP-TCN-BiGRU混合模型，通过双通路设计分离处理静态特征和时序变量, TCN层通过膨胀卷积捕捉时序变量的长程依赖，结合BiGRU的双向门控机制对提取的特征进行非线性建模，MLP层则融合静态特征，所有的信息通过拼接层生成融合表征。最后，采用牛顿-拉夫森方法的优化器实现超参数的自适应调优。实例分析表明，本文所提方法通过隐式关联规律建模，在建筑项目的成本估算中具有更高的预测准确性，为建筑项目管理者提供了实时风险预警和决策支持工具。 |
| 英文摘要: |
| Reliable estimation of construction costs is crucial for project planning and resource allocation. To address the challenges of high-dimensional data redundancy and heterogeneity in time-varying data in the construction engineering domain, this study proposes a hybrid deep learning framework based on joint feature selection, which integrates multi-layer perceptron (MLP), temporal convolutional network (TCN), and bidirectional gated recurrent unit (BiGRU) for cost prediction. First, a dual-dimensional feature selection mechanism based on the Pearson-Spearman joint test is designed to eliminate redundant features and reduce data noise. Second, a hybrid framework with a parallel processing architecture, MLP-TCN-BiGRU, is constructed, where static features and temporal variables are processed separately through dual pathways: the TCN layer captures the long-range dependencies of temporal variables using dilated convolutions, while the BiGRU’s bidirectional gating mechanism nonlinearly models the extracted features. Meanwhile, the MLP layer integrates static features and generates collaborative representations through a concatenation layer. Finally, an optimizer based on the Newton-Raphson method is introduced to achieve adaptive tuning of hyperparameters. Case studies demonstrate that the proposed method, through modeling implicit association patterns, achieves higher prediction accuracy in construction project cost estimation, providing project managers with a real-time risk-early warning and decision-support tool. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
