MDS-LOF and GBRT fused project cost prediction method

Abstract

The invention discloses a project cost prediction method fusing MDS-LOF and GBRT. The method comprises the steps of S1, selection and primary processing of project cost data; s2, performing project feature analysis and reserving core data; s3, performing dimension reduction processing by using a multi-dimensional scaling analysis (MDS) method; s4, using a local outlier factor (LOF) algorithm to identify abnormal values and removing the abnormal values; s5, training is carried out by using the GBRT prediction model and the processed data, and a trained model is obtained; and S6, outputting and verifying the trained prediction model. According to the invention, by constructing the adaptive prediction model, the influence caused by sudden policy regulation and control is weakened, so that the project cost prediction result is more in line with the market law of the building construction cost, and scientific basis and reference value are provided for project early-stage decision making and cost prediction of investment subjects such as enterprises and governments in a complex policy environment period.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of engineering economy, and particularly relates to an engineering cost prediction method fusing MDS-LOF and GBRT. BACKGROUND

[0002] Engineering cost prediction is a core link in construction engineering management and decision-making, and its result is directly related to the rationality of project investment control, cost management and resource allocation. With the continuous expansion of the scale of construction and the increasing diversification of construction types, engineering cost prediction has gradually transitioned from experience estimation to data-driven quantitative analysis methods.

[0003] At present, common cost prediction methods are mainly divided into two categories: Firstly, traditional statistical methods such as regression analysis, time series method, etc. This kind of method can obtain a certain precision of prediction under the condition of stable data and small external disturbance, but due to its dependence on the stationary distribution assumption of data, it is difficult to cope with the cost fluctuation caused by policy factors and unexpected events in the construction market.

[0004] Secondly, machine learning methods such as BP neural network, support vector machine (SVM), random forest, etc. This kind of method can model complex nonlinear relationships and improve prediction accuracy to a certain extent. However, due to the fact that engineering cost data often contains outliers and noise data caused by policy regulation, machine learning models are easily affected by outliers and produce overfitting or prediction bias in the absence of effective data cleaning mechanism.

[0005] In addition, in recent years, policy regulation (such as real estate regulation, environmental standard upgrading, material capacity adjustment) has an increasingly significant impact on the price of building materials and engineering cost, making engineering cost data present sudden fluctuations and non-stationary characteristics. Under this background, traditional prediction methods lack the ability to handle outliers, and common machine learning models are too sensitive to noise data, making it difficult to maintain the stability and reliability of prediction results in a complex policy environment.

[0006] Therefore, how to effectively identify and eliminate abnormal data in engineering cost prediction, and combine prediction models with strong generalization ability and anti-noise performance, has become a technical problem that needs to be solved at present. This not only helps to improve the accuracy and stability of engineering cost prediction, but also provides a more scientific basis for project decision-making of enterprises and government investors in a complex policy environment. SUMMARY

[0007] The problem to be solved by the present application is that the existing neural network-based engineering cost prediction is affected by sudden policy regulation, resulting in abnormal fluctuations in engineering cost data, which affects the prediction accuracy of machine learning models.

[0008] In order to solve the above technical problems, the application provides an engineering cost prediction method combining MDS-LOF and GBRT, comprising the following steps: S1: selection and preprocessing of project cost data; S2: project feature analysis and core data retention; S3: dimensionality reduction processing using the multidimensional scaling (MDS) method; S4: identification and elimination of outliers using the local outlier factor (LOF) algorithm; S5: training using the GBRT prediction model and the processed data to obtain a trained model; S6: outputting the trained prediction model, verifying it, and outputting the prediction result.

[0009] Preferably, the specific steps of data preprocessing in S1 are as follows: Step 1: Quantitative processing of engineering cost data according to different engineering project feature types; Step 2: Using the median completion method to process missing data in the data collection process; Step 3: To eliminate dimensional differences, Max-Min normalization is used to process each item of engineering data to obtain preprocessed data.

[0010] Preferably, the specific steps of project feature analysis and core data retention in S2 are as follows: Step 1: Correlation analysis using the Pearson correlation coefficient, with the analysis results output in chart form; Step 2: Eliminating or merging highly correlated feature values.

[0011] Preferably, in S3, the multidimensional scaling (MDS) method is used for dimensionality reduction processing, and the Manhattan distance method is used. wherein and represent the coordinate values of the i-th data point and the j-th data point in the k-dimensional space, respectively, and the sample data points in the high-dimensional space are mapped to the low-dimensional space to obtain a visual chart; Preferably, in S4, the sample data obtained after dimensionality reduction processing in S3 is analyzed according to its coordinate point data using the local outlier factor (LOF) algorithm. wherein, is the local reachable density of point p, is the k-field, which identifies the deviation of the project point and eliminates outliers; Preferably, in S5, the data processed in S4 is grouped into a training set and a prediction set in a ratio of 8:2, and the training set is used to train the GBRT model to obtain a trained prediction model. Preferably, the S6 uses the prediction set to verify the trained model in S5, outputs the prediction result; The application has the advantages that: The application introduces a data processing mechanism combining multidimensional scaling (MDS) and local outlier factor (LOF) in the engineering cost prediction process, which can not only effectively reduce the dimension of high-dimensional cost parameters and retain the core features, but also identify and eliminate outliers caused by sudden policy regulation or market abnormal fluctuations, thereby improving the reliability of data from the source. In the prediction stage, the gradient boosting regression tree (GBRT) model is used to overcome the problem of overfitting of traditional neural network models when dealing with abnormal data, and the prediction result is significantly improved in terms of accuracy and stability. In addition, the application also establishes a policy-sensitive abnormal value determination threshold, so that the prediction model can still maintain strong adaptability and robustness in complex policy environment, and finally makes the engineering cost prediction result closer to the actual construction cost market rule, and provides a more scientific theoretical basis and practical reference for project investment subjects such as enterprises and governments in the early stage of project decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 The figure is a flowchart of the engineering cost prediction method combining data processing and prediction model of the application; Figure 2 The figure is a distribution diagram of coordinate points obtained after dimensionality reduction of cost data by multidimensional scaling (MDS) in the application; Figure 3 The figure is a structure diagram of the prediction model constructed by integrating multiple regression trees by gradient boosting regression tree (GBRT) in the application; Figure 4 The figure is a prediction result diagram obtained in the verification stage after model training in the application. DETAILED DESCRIPTION

[0013] In order to clearly and completely describe the technical solutions and technical effects of the application, the following examples are used for detailed description.

[0014] Example 1 As shown in the accompanying Figure 1 The method for project cost prediction based on MDS-LOF and GBRT model includes the following steps: S1, obtaining project cost data, and selecting and preprocessing the obtained experimental data; In this embodiment, the specific method for selecting and preprocessing the project cost data includes the following steps: The characteristic indicators are divided into four categories: the first category is core cost factors (unit volume of steel reinforcement, unit volume of concrete, etc.), which are quantified quantitatively (kg / m², m³ / ㎡) and are the core factors affecting unit cost; the second category is price fluctuation factors (unit price of steel reinforcement, unit price of ready-mixed concrete, etc.), which are quantified quantitatively (yuan / ton, yuan / m³) and are the factors most sensitive to policy disturbances; the third category is engineering design factors (project type, structural type, foundation type, etc.), which are converted from qualitative indicators to ordinal numbers for quantification, reflecting the complexity of the construction project and truly reflecting the project needs of the construction unit, and are important indicators affecting unit cost; the fourth category is environmental constraint factors (project location, seismic intensity, etc.), which are converted from qualitative indicators to ordinal numbers for quantification. Considering the possibility of a very small number of random missing data points during data collection, and to maximize sample integrity while avoiding variance compression caused by mean imputation, this paper employs median imputation. To avoid extreme value disturbances, for each variable with missing values, the median value of the position is taken after sorting the samples to fill in the missing data.

[0015] To adapt to the numerical data requirements of machine learning models, this study designs transformation strategies for qualitative and quantitative data respectively to address issues of attribute heterogeneity, multicollinearity, and dimensional differences. Qualitative data (such as project type and delivery form) needs to be converted into numerical indicators for inclusion in the model. Based on the differences in qualitative data attributes, ordinal variables are used for ordinal encoding. To address the common issues (correlation coefficients) among the original quantitative features (such as total building area and land area), and to enhance feature interpretability, composite indicators are created: floor area ratio (above-ground building area / land area) reflects land use efficiency, single-story building area [total building area / (number of above-ground floors + number of underground floors)] reflects the complexity of construction organization, and single elevator square meters (above-ground building area / number of elevators) reflects the requirements of the building project type.

[0016] To eliminate dimensional differences, Max-Min normalization is used: S2. Conduct correlation analysis on project characteristics and retain core data; In this embodiment, the specific method for project feature analysis includes the following steps: Correlation analysis using the Pearson correlation coefficient can remove or merge highly correlated features to improve the model's prediction accuracy and generalization ability.

[0017] S3. Use the Multidimensional Scale Analysis (MDS) method to reduce the dimensionality of the project cost data and obtain the point coordinate map after dimensionality reduction. The specific method of data dimension reduction processing in this embodiment includes the following steps: In this study, Manhattan Distance is used for multidimensional scaling analysis: Wherein and represent the coordinate values of the i-th data point and the j-th data point in the k dimension, and the final calculation result represents the distance or difference between the i-th data point and the j-th data point. By MDS dimension reduction to a two-dimensional chart, a visualization chart as shown in FIG. 2 is obtained. Figure 2

[0018] S4, using local outlier factor algorithm LOF to identify and eliminate outliers; In this embodiment, the specific method of identifying outliers of the chart coordinate points after MDS dimension reduction includes the following steps: The density-based unsupervised anomaly detection algorithm identifies anomalies by quantifying the local density deviation of data points. Its core formula is: Wherein, is the local reachable density of point p, k domain. The sample number of this method is 150, and the optimized neighbor number K=12.24.

[0019] Combined with policy correlation analysis, the effect of the prediction model is improved, and the samples of are defined as outliers and eliminated.

[0020] S5, using gradient boosting regression tree GBRT prediction model to train the processed data and obtain the trained model; In this embodiment, the gradient boosting regression tree GBRT model used is a regression algorithm based on the idea of ensemble learning. By iteratively improving a plurality of weak learners (regression decision trees) in sequence, a strong learner is constructed. In each round of training, the new model will fit the residual of the last model, so as to continuously optimize the overall prediction ability, as shown in FIG. 3. Figure 3

[0021] In the context of regression, the first weak learner is a constant, and the subsequent weak learners are regression trees, and the formula is: ​​In the method, the structure design of constructing the gradient boosting model is as follows: base learner: Regression Tree; maximum number of splits of weak learner: 5; number of iterations (learning rounds): 100; learning rate (shrinkage): 0.1; loss function: least square error (LSBoost); The data processed in S4 is grouped into a training set and a prediction set according to a ratio of 8:2, and the training set is used to train the GBRT model, and a trained prediction model is obtained.

[0022] S6, the trained prediction model is verified using the prediction set, and a prediction result is output. In the test set, the root mean square error (RMSE) of the GBRT is significantly lower than that of the BP neural network and the linear regression model, the determination coefficient R² is improved by 9% compared with the BP neural network and by 15% compared with the linear regression model, and for the same set of project data, it is explained that the GBRT model is more suitable for processing the prediction task of the “single-party cost” in the engineering project. Since the GBRT model essentially belongs to an integrated model, it has an anti-overfitting capability, and the prediction result is shown in the accompanying drawings. Figure 4

[0023] The above only describes some embodiments of the present application. For those skilled in the art, without departing from the inventive concept, several modifications and improvements can be made, and any modification, equivalent replacement and improvement within the inventive concept are all within the protection scope of the present application.​

Claims

1. A method for predicting engineering costs by integrating MDS-LOF and GBRT, characterized in that, Includes the following steps: S1: Selection and preprocessing of project cost data; S2: Project feature analysis retains core data; S3: Dimensionality reduction is performed using the Multidimensional Scaling (MDS) method; S4: Use the local outlier factor algorithm to identify and remove outliers; S5: Use the GBRT prediction model and the processed data to train the model and obtain the trained model; S6: Output the trained prediction model and validate it.

2. The engineering cost prediction method integrating MDS-LOF and GBRT according to claim 1, characterized in that, The specific steps of data preprocessing in S1 are as follows: Step 1: The project cost data is quantified according to different project characteristics; Step 2: For missing data present during the data collection process, use the median completion method to handle them; Step 3: To eliminate dimensional differences, Max-Min normalization is used to process the various engineering data to obtain preprocessed data.

3. The engineering cost prediction method integrating MDS-LOF and GBRT according to claim 1, characterized in that, The specific steps for retaining core data in the project feature analysis of S2 are as follows: Step 1: Perform correlation analysis using the Pearson correlation coefficient and output the results in chart form; Step 2: Remove or merge highly correlated features.

4. The engineering cost prediction method integrating MDS-LOF and GBRT according to claim 1, characterized in that, In S3, dimensionality reduction is performed using the Multidimensional Scaling (MDS) method, employing the Manhattan distance method. ,in and These represent the coordinates of data point i and data point j in the k-dimensional space, respectively. This maps the sample data points from the high-dimensional space to the low-dimensional space and obtains a visualization chart.

5. The engineering cost prediction method integrating MDS-LOF and GBRT according to claim 1, characterized in that, The data chart obtained in S4, after dimensionality reduction of the sample data obtained in S3, uses the local outlier factor algorithm based on its coordinate point data. ,in, It is the locally reachable density of point p. For the k-domain, identify deviation anomalies at project locations and remove outliers.

6. The engineering cost prediction method integrating MDS-LOF and GBRT according to claim 1, characterized in that, In step S5, the data processed in step S4 is grouped into training and prediction sets in an 8:2 ratio. The training set is then used to train the GBRT model, resulting in a trained prediction model.

7. The engineering cost prediction method integrating MDS-LOF and GBRT according to claim 1, characterized in that, In step S6, the training model completed in step S5 is validated using the prediction set.

Images