混合應用隨機森林和簡單線性迴歸之BIM人力成本預測方法論

A BIM Labor Cost Prediction Methodology with Random Forest and Simple Linear Regression

黃建勳(Chien-Hsun Huang)

指導教授 ： 謝尚賢

國立臺灣大學/工學院/土木工程學研究所/博士(2021年)

https://doi.org/10.6342/NTU202100542

摘要

建築資訊塑模(Building Information Modeling, BIM)近年來已廣泛應用於營建產業，在建築專案中採用此技術的好處亦廣為人知。但是，導入BIM至專案時所產生的額外人力成本，以及執行時是否具有時效性，都是影響業主決定是否採用BIM技術的因素。目前，專案經理通常會在成本決策方面面臨很多風險。因為在實務上BIM人力成本被認為與總建築面積具有某種比例關係，並且專案經理亦大多採用簡單線性迴歸來進行估算。儘管此方法簡單明瞭，但往往存在相當高的估算誤差風險。
因此，本研究試圖開發一種基於跨行業資料探勘標準流程（CRISP-DM）的新方法論，並提出了一種將機器學習(ML)模型和簡單線性迴歸（Simple Linear Regression, SLR）模型相結合的混合方法，以提高模型對專案在施工階段的BIM人力成本預測的準確性。本研究首先透過台灣某大型營造工程公司過去執行的21個實際案例所記錄的工時記錄資料為基礎，提出了一種成本分解結構（Cost Breakdown Structure, CBS），以建立用於機器學習的訓練數據集。接著導入了隨機森林（Random Forest, RF）和支持向量機（Support Vector Machine, SVM）兩種機器學習技術進行結構體模型建置及結構體施工圖兩項BIM應用的執行人力成本預測，並與實務上常用的簡單線型迴歸模型進行效能比較。針對簡單線性迴歸模型，本研究亦提出了有效樓地板面積的概念，並以此參數作為簡單線性迴歸建模的依據。最後，透過時間序列分析、聚類分析、特徵值選擇等方式對於模型進行修正及優化以提出本研究之BIM人力成本預測方法論。
上述研究結果顯示：(1)隨機森林模型相較於支持向量機模型來說，對於BIM結構模型建置的人力成本估算具有更好的預測能力，但對於結構體施工圖產出的人力成本估算來說，機器學習技術則無顯著優勢，反而是簡單線性迴歸模型有較佳的表現。(2)本研究亦發現若採用有效樓地板面積參數代替總樓地板面積參數建置RF和SLR模型，其預測效能提升都會有所提升。(3)透過時間序列分析證明了在不同案例組成的數據集下，各模型可能會產生不同的表現，因此本研究混合了機器學習模型與簡單線性迴歸模型提出了一個新的模型選擇策略。(4)聚類分析及模型降維兩種方法在本研究中已被分別確認可提高模型性能及簡化建模的複雜度。(5)最後，通過比較研究結果，本研究中提出的BIM人力成本預測方法論已被證明可有效降低BIM人力成本預測時的風險。

關鍵字

建築資訊塑模 ； 成本分解結構 ； 人力成本預測 ； 隨機森林 ； 簡單線性迴歸

並列摘要

The benefits of adopting Building Information Modeling (BIM) in a construction project have been well recognized. However, it involves additional labor costs and the timeliness of task completion should be considered because a BIM application may not be required (and paid for) by the owner in the current stage of BIM development. At present, a project manager often faces much risk in making decisions about the cost because in practice the BIM labor cost is supposed to be proportional to the gross (or total) floor area or the percentage of the total construction cost, and the project managers can only adopt simple linear regression to estimate it. Although this method is straightforward, it tends to have a high risk of estimation error.
Therefore, this research tried to develop a new methodology based on Cross Industry Standard Process for Data Mining (CRISP-DM). At first, in order to evaluate whether machine learning technologies can more accurately estimate BIM labor costs, two machine learning models are built which are based on Random Forest (RF) and Support Vector Machine (SVM,) respectively. Case studies are conducted to demonstrate and validate the prediction results using twenty-one completed BIM projects from a leading construction company in Taiwan. A cost breakdown structure (CBS) was proposed to establish the training data set for machine learning.
The research results show that, based on leave-one-out cross-validation (LOOCV), by comparing the mean absolute error (MAE) and mean square error (MSE) of the two models, the RF model performs better than others for predicting the BIM labor cost on building structural BIM model. However, for the production of construction working drawing from the BIM model, the performance of both RF and SVM models have no significant advantage over the commonly used linear regression models.
Because Random forest model and Simple Linear Regression model have advantages in different situations, a hybrid approach by combining Random Forest (RF) and Simple Linear Regression (SLR) is proposed for improving the accuracy of prediction on a project's BIM labor cost in construction phase. Moreover, this study proposes the use of effective floor area, instead of gross floor area, to be one of the features for training RF and SLR models. Clustering analysis is also adopted and confirmed to improve model performance. Through comparative studies, the hybrid approach of prediction methodology proposed in this study has been proved to be effective in reducing the risk of BIM labor cost prediction.

並列關鍵字

Building Information Modeling ； Cost Breakdown Structure ； Labor Cost Prediction ； Random Forest ； Simple Linear Regression

參考文獻

Ahmad, M. W., Reynolds, J., Rezgui, Y. (2018). Predictive modelling for solar thermal energy systems: A comparison of support vector regression, random forest, extra trees and regression trees. Journal of cleaner production, 203, 810-821. https://doi.org/10.1016/j.jclepro.2018.08.207

Allweyer, T. (2016). BPMN 2.0: introduction to the standard for business process modeling. BoD–Books on Demand.

Azhar, S. (2011). Building information modeling (BIM): Trends, benefits, risks, and challenges for the AEC industry. Leadership and management in engineering, 11(3), 241-252. https://doi.org/10.1061/(asce)lm.1943-5630.0000127.

Barlish, K., Sullivan, K. (2012). How to measure the benefits of BIM—A case study approach. Automation in construction, 24, 149-159. https://doi.org/10.1016/j.autcon.2012.02.008.

Breiman, L. (2001). Random forests. Machine learning, 45(1), 5-32. doi:10.1023/ A:1010933404324