初步的工程估價在工程專案中是ㄧ個重要的階段。在這個階段，承包商與業主要能評估這個專案是否可行。在印度尼西亞，一個典型的建築專案得花上數周來做初步的工程估價，且誤差值的範圍會從-12.97% 到+26.80%。以往的研究指出，有74%的成本超支是由過低的工程估價造成的。因此，本研究的目的在於：1)確立在印度尼西亞的估價因子2)發展一個支撐向量回歸模型，試圖去改善精準度以及減少初步工程估價的工作時間。文獻回顧辨別出了14個影響世界各地所有的建築專案的估價因子。考慮到這些因素做為模型之基礎，資料隨機取樣，蒐集104項包含有效資訊的印尼建築案例供模型使用。在資料修整、分析以及正規劃後，建立出伴隨徑向基函數核(radial basis function kernel)的SRV模型。以5折交叉驗證來預估以及執行模型，並產出平均95.79%正確率之工程專案初步預估。從SVR模型比原始資料提升了8.71%準確率。過往人工計算初期成本需要以周為單位計算，現行則由模型運算以秒為單位處理，由此可看出對於縮減時間亦是相當重要的。由上述可得知，SRV模型在正確率與省時都是相當優異的。;Preliminary cost estimation is an important stage for construction projects. During the stage, any contractor and owner is able to determine whether his/her project is feasible. A typical preliminary cost estimation for a building construction project in Indonesia may take weeks and have an error rate varying from -12.97% to +26.80%. Previous studies also concluded that 74% of cost overruns are caused due to underestimation. The research objectives, therefore, are (1) to determine factors that influence cost estimation in Indonesia and (2) to develop a Support Vector Regression (SVR) model in an attempt to improve accuracy and to reduce workhours for preliminary cost estimation. Literature review identified 14 factors that influence cost estimation the most for all types of construction projects around the world. Considering these factors as the model bases, data collection randomly gathered 104 building cases in Indonesia containing valid information for the proposed model. The SVR model with the radial basis function kernel was established after data trimming, analysis, and normalization. The model then was evaluated and implemented using the 5-folds cross validation and yielded the average accuracy at 95.79% for preliminary cost estimation of building construction projects. The accuracy has been improved 8.71% between the original data and the results from the SVR model. Time spent for conducting such a preliminary cost estimation has been significantly reduced from weeks by human estimators to less than one second by the model. The SVR model is efficient in both accuracy and time-saving.
