Subway station construction stage carbon emission prediction method and system
By constructing an integrated model of LightGBM and XGBoost, and combining the carbon emission coefficient method and data preprocessing technology, the systematic and accurate problems of carbon emission prediction during the construction phase of subway stations were solved, achieving efficient and accurate carbon emission prediction. This provides a scientific basis for subway station design and construction management, and contributes to the low-carbon construction of urban rail transit.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA RAILWAY DESIGN GRP CO LTD
- Filing Date
- 2026-01-13
- Publication Date
- 2026-05-01
AI Technical Summary
Existing research on carbon emission prediction for subway stations mainly focuses on the operation phase, lacking systematic and accurate methods for predicting carbon emission during the construction phase. Single models have limitations in handling complex nonlinear relationships and high-dimensional data, making it difficult to fully capture the complex features in carbon emission data, resulting in insufficient prediction accuracy and generalization ability.
An ensemble model combining LightGBM and XGBoost models, along with the carbon emission coefficient method and data preprocessing techniques, is used to construct an ensemble model for carbon emission prediction through feature importance identification, feature transformation, and parameter optimization. The SHAP method is then used for interpretability analysis to establish the mapping relationship between features and physical components, and to calculate the local marginal effects of key features.
It improves the accuracy and stability of carbon emission prediction during the construction phase of subway stations, provides a scientific basis for subway station design optimization and construction management, helps achieve carbon emission reduction targets in urban rail transit construction, reduces model complexity and overfitting risk, and improves prediction efficiency.
Smart Images
Figure CN121960964A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of carbon emission prediction technology, and in particular to a method and system for predicting carbon emissions during the construction phase of a subway station. Background Technology
[0002] In the field of carbon emission prediction, machine learning algorithms have been widely used due to their powerful data processing capabilities. For example, some scholars have predicted the trend of carbon emission intensity in my country based on LSTM neural network models; others have used the gradient boosting tree (LightGBM) algorithm to build a carbon emission prediction model and analyzed the influence of different factors on carbon emissions. LightGBM and XGBoost algorithms have shown significant advantages in handling large-scale data and complex model relationships.
[0003] However, most of these models are logically single models, which have problems such as failing to capture nonlinear relationships well and insufficient adaptability. There is an urgent need to introduce composite models to effectively capture complex carbon emission data in order to scientifically predict carbon emissions from subway stations.
[0004] Specifically, while the LightGBM model alone incorporates a gradient one-sided sampling (GOSS) strategy, which can significantly reduce computational resource consumption while maintaining prediction accuracy, it still has certain limitations when dealing with complex nonlinear relationships. The XGBoost model, on the other hand, optimizes the objective function by introducing regularization terms and second-order derivative expansion, balancing model fitting accuracy and complexity control, but it may still experience overfitting in some cases, leading to a decrease in prediction accuracy.
[0005] In summary, existing research on carbon emission prediction for subway stations focuses primarily on the operational phase, with limited research on carbon emission prediction during the construction phase, and a lack of systematic and accurate prediction methods. Furthermore, existing single models have limitations in handling complex nonlinear relationships and high-dimensional data, failing to fully capture the complex characteristics of carbon emission data, resulting in room for improvement in prediction accuracy and generalization ability. Current technologies have not fully utilized the complementary advantages of multiple machine learning algorithms, lacking effective ensemble models to enhance prediction performance.
[0006] Therefore, there is an urgent need for a method and system for predicting carbon emissions during the construction phase of subway stations. Summary of the Invention
[0007] To address the aforementioned technical issues, this invention provides a method and system for predicting carbon emissions during the construction phase of subway stations. This method overcomes the limitations of a single model, improves prediction accuracy and stability, and provides a scientific basis for subway station design optimization, construction management, and building material selection, thereby contributing to the achievement of carbon emission reduction targets in urban rail transit construction.
[0008] This invention proposes a method for predicting carbon emissions during the construction phase of subway stations, including: Carbon emissions at different stages of a metro station are calculated using the carbon emission coefficient method, and data preprocessing is performed to obtain carbon emission monitoring data. Feature and target variables are separated from the carbon emission monitoring data to obtain feature matrix and target variable vector. A carbon emission prediction model is constructed to perform feature processing on carbon emission monitoring data to obtain carbon emission feature data; this includes processing carbon emission monitoring data based on LightGBM split gain to generate feature importance; and performing logarithmic transformation on all features. A carbon emission prediction ensemble model was constructed based on the LightGBM and XGBoost models. The carbon emission prediction ensemble model was used to identify carbon emission characteristic data. Grid search was used to optimize the parameters in the LightGBM and XGBoost models to obtain the predicted carbon emissions during the construction phase of subway stations. The process of building the model and obtaining the predictions also includes: constructing a monotonicity constraint vector, applying monotonicity constraints to preset physical features during model training to restrict the splitting of decision tree nodes and weight updates; calibrating the model's point prediction results using a conformity prediction method, constructing a conformity prediction interval with statistical coverage guarantees by calculating the quantile threshold of inconsistency scores on an independent calibration set; performing interpretability analysis on the prediction results based on the SHAP method, establishing the mapping relationship between features and physical components, calculating the sum of SHAP contributions of each physical component to achieve carbon emission source tracing, and calculating the local marginal effects of key features.
[0009] The carbon emissions of a metro station at different stages are calculated using the carbon emission coefficient method. These stages include: building material production, transportation, and construction. The feature matrix includes the number of ventilation shafts, station length, width, foundation pit depth, top slab soil cover thickness, number of station entrances / exits, station area, carbon emissions of building materials during the production stage, carbon emissions of building materials during the transportation stage, and carbon emissions of machinery during the construction stage. The target variable vector includes total carbon emissions.
[0010] The data preprocessing includes: using column mean imputation to complete the data in the original carbon emission monitoring data of the station; and using statistical principles to fill in missing samples by calculating the mean of non-missing values in each feature column.
[0011] The carbon emission prediction model includes: Based on the split gain generation feature importance of the LightGBM model, key features affecting station carbon emissions are identified and screened; and the dimensionality of the feature matrix is reduced based on these key features to reduce model complexity and the risk of overfitting. Logarithmic transformation is applied to all features to address the skewness of the feature distribution, making the feature distribution closer to a normal distribution and improving the model's stability, convergence speed, and prediction accuracy.
[0012] The carbon emission prediction ensemble model includes a LightGBM sub-model and an XGBoost sub-model, and the construction process includes: Based on the carbon emission calculations for different stations, a weighted average method was used to establish a LightGBM-XGBoost integrated prediction model. At the same time, a grid search method was used to optimize the parameters in the LightGBM model and the XGBoost model respectively.
[0013] The LightGBM sub-model includes: Histogram binning technology is used to discretize carbon emission characteristic data and clarify the reasonable range of characteristic variables; The GOSS sampling method is used to select representative samples from the data; By leveraging leaf node histogram optimization technology, the model structure is further refined to capture subtle carbon emission variation patterns at different stages of subway construction.
[0014] The XGBoost sub-model includes: The feature variables in the carbon emission feature data are sorted by a pre-sorting algorithm to screen out the key factors that have the most significant impact on carbon emissions. A decision tree model is constructed using a precise greedy splitting method to ensure that the splitting operation of each node can improve the prediction accuracy of the model; Meanwhile, sparse perception optimization technology is introduced to address the sparsity problem commonly found in subway construction data, thereby improving the model's adaptability and generalization ability to complex data environments.
[0015] A carbon emission prediction system for the construction phase of a subway station includes: The data acquisition module is used to calculate the carbon emissions of the metro station at different stages according to the carbon emission factor method, and to perform data preprocessing to obtain carbon emission monitoring data. The feature separation module is used to separate features and target variables from the carbon emission monitoring data to obtain a feature matrix and a target variable vector. The feature matrix includes the number of ventilation shafts, station length, width, foundation pit depth, top slab soil thickness, number of station entrances and exits, station area, carbon emissions from building materials during the production stage, carbon emissions from building materials during the transportation stage, and carbon emissions from machinery during the construction stage. The target variable vector includes the total carbon emissions. The feature processing module is used to construct a carbon emission prediction model and perform feature processing on carbon emission monitoring data to obtain carbon emission feature data; including processing carbon emission monitoring data based on LightGBM split gain to generate feature importance; and performing logarithmic transformation on all features. The model building module is used to build an integrated carbon emission prediction model based on LightGBM and XGBoost. The integrated carbon emission prediction model identifies carbon emission characteristic data and uses grid search to optimize the parameters in the LightGBM and XGBoost models to obtain the carbon emission prediction during the construction phase of subway stations.
[0016] The embodiments of the present invention have the following technical effects: 1. The carbon emission data preprocessing method provided by this invention has significant advantages in processing carbon emission data from subway integrated stations. Based on the internationally accepted IPCC carbon emission factor method, this method comprehensively calculates carbon emissions from multiple stages, including building material production, transportation, construction, and engineering, ensuring data integrity and accuracy. Simultaneously, addressing the common problem of missing values in raw carbon emission monitoring data, it employs column mean interpolation for efficient data completion, effectively avoiding the impact of missing data on prediction results. This provides a high-quality data foundation for subsequent carbon emission prediction model training, thereby improving the accuracy and reliability of model predictions.
[0017] 2. The feature processing method of the carbon emission prediction model in this invention can significantly improve the model's performance and prediction accuracy. Based on the split gain generation feature importance of the LightGBM model, key features that have a significant impact on the carbon emissions of subway stations are accurately identified and screened, such as the depth of the foundation pit and the carbon emissions of building materials during the station construction stage. This effectively reduces the feature dimensionality, model complexity, and the risk of overfitting. Through polynomial feature generation and logarithmic transformation, not only are the nonlinear relationships between features captured, but the skewness of the feature distribution is also addressed, making the feature distribution closer to a normal distribution. This further improves the model's ability to fit complex patterns and its prediction accuracy, enhances the model's stability and convergence speed, and ensures that the carbon emission prediction model can more accurately reflect the carbon emissions during the construction phase of subway stations in practical applications.
[0018] 3. The carbon emission prediction ensemble model based on LightGBM and XGBoost of this invention has significant advantages in the field of carbon emission prediction during the construction phase of subway stations. This model integrates LightGBM's GOSS data sampling strategy and XGBoost's regularization term introduction, enabling efficient processing of large-scale carbon emission data and effectively avoiding overfitting, thus improving the model's robustness and generalization ability. By employing a weighted average method to establish the ensemble prediction model and utilizing grid search to optimize the model parameters, the model's performance is further optimized, allowing for more accurate prediction of carbon emissions for different stations. From a practical application perspective, this model only requires inputting basic data such as different engineering quantities according to engineering budget quota standards, without the need for complex on-site monitoring methods, to quickly and accurately obtain carbon emission prediction results during the subway station construction phase. This provides scientific and efficient technical support for the formulation of energy-saving and emission-reduction measures and carbon emission management during subway construction. Attached Figure Description
[0019] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0020] Figure 1 This is a flowchart of a carbon emission prediction system for the construction phase of a subway station, provided in an embodiment of the present invention. Figure 2 This is a flowchart of the data training process for the carbon emission prediction integrated model provided in this embodiment of the invention; Figure 3 This is a schematic diagram illustrating the importance of subway carbon emission characteristics provided in an embodiment of the present invention; Figure 4 This refers to the training set prediction results during the training process of the carbon emission prediction integrated model provided in this embodiment of the invention. Figure 5 These are the test set prediction results during the training process of the carbon emission prediction integrated model provided in this embodiment of the invention. Figure 6 This is a graph of the carbon emission prediction integrated model evaluation data provided in this embodiment of the invention; in the graph, a represents the mean absolute error (MAE), b represents the mean squared error (MSE), and c represents the coefficient of determination (R²). 2 ), d represents the root mean square error (RMSE). Detailed Implementation
[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are part of this invention.
[0022] Current prediction methods mostly employ single models, such as the LightGBM or XGBoost models. While these models can predict carbon emissions to some extent, they still have limitations when handling complex nonlinear relationships and high-dimensional data, and their prediction accuracy and generalization ability need improvement. Furthermore, existing research largely focuses on energy consumption and carbon emission analysis during the operation phase, with relatively few studies on carbon emission prediction during the subway station construction phase, lacking systematicity and precision. Therefore, the purpose of this invention is to construct a carbon emission prediction method for the subway station construction phase, overcoming the limitations of single models, improving prediction accuracy and stability, providing a scientific basis for subway station design optimization, construction management, and building material selection, and contributing to the achievement of carbon reduction targets in urban rail transit construction.
[0023] Example 1: This invention proposes a method for predicting carbon emissions during the construction phase of subway stations, including: Carbon emissions at different stages of a metro station are calculated using the carbon emission coefficient method, and data preprocessing is performed to obtain carbon emission monitoring data. Feature and target variables are separated from the carbon emission monitoring data to obtain feature matrix and target variable vector. A carbon emission prediction model is constructed to perform feature processing on carbon emission monitoring data to obtain carbon emission feature data; this includes processing carbon emission monitoring data based on LightGBM split gain to generate feature importance; and performing logarithmic transformation on all features. A carbon emission prediction ensemble model was constructed based on the LightGBM and XGBoost models. The carbon emission prediction ensemble model was used to identify carbon emission characteristic data. Grid search was used to optimize the parameters in the LightGBM and XGBoost models to obtain the predicted carbon emissions during the construction phase of subway stations. The process of building the model and obtaining the predictions also includes: constructing a monotonicity constraint vector, applying monotonicity constraints to preset physical features during model training to restrict the splitting of decision tree nodes and weight updates; calibrating the model's point prediction results using a conformity prediction method, constructing a conformity prediction interval with statistical coverage guarantees by calculating the quantile threshold of inconsistency scores on an independent calibration set; performing interpretability analysis on the prediction results based on the SHAP method, establishing the mapping relationship between features and physical components, calculating the sum of SHAP contributions of each physical component to achieve carbon emission source tracing, and calculating the local marginal effects of key features.
[0024] The carbon emissions of a metro integrated station at different stages are calculated using the carbon emission coefficient method. These stages include: building material production, transportation, construction, and engineering construction. The feature matrix includes the number of ventilation shafts, station length, width, foundation pit depth, top slab soil cover thickness, number of station entrances / exits, station area, carbon emissions of building materials during the production stage, carbon emissions of building materials during the transportation stage, and carbon emissions of machinery during the construction stage (X0 to X9). The target variable vector includes total carbon emissions (Y).
[0025] The carbon emission data calculation mainly adopted the national greenhouse gas inventory guidelines provided by the Intergovernmental Panel on Climate Change (IPCC), and used the carbon emission factor method to calculate the carbon emissions of the metro complex at different stages.
[0026] The calculation formula is: (1) (2) (3) (4) (5) (6) In equations (1)-(6): Total carbon emissions; This refers to the total carbon emissions during the building materials production stage. This represents the total carbon emissions during the transportation of building materials. This refers to the total carbon emissions during the construction phase of building materials; Let i be the consumption amount of the i-th main building material; is the carbon emission factor of the i-th main building material; Let be the average transportation distance for the i-th type of building material; The carbon emission factor for the transportation of the i-th type of building material; For carbon emissions during the design phase, simplified calculations may exclude carbon emissions during the design phase. This refers to direct carbon emissions during the engineering construction phase; Let be the number of shifts consumed by the i-th type of construction machinery powered by gasoline, diesel, or other fuels; The energy consumption per shift for the i-th type of construction machinery powered by gasoline, diesel, or other fuels is expressed in kg fuel / shift. Let be the carbon emission factor for the combustion of the i-th fuel, kgCO2e / kg; This represents indirect carbon emissions during the construction phase, expressed as kgCO2e. The total power consumption during the construction phase of the project is expressed in kWh. This is a carbon emission factor for the power grid.
[0027] The data preprocessing includes: using column mean imputation to complete the data in the original carbon emission monitoring data of the station; and using statistical principles to fill in missing samples by calculating the mean of non-missing values in each feature column.
[0028] The carbon emission prediction model includes: Based on the split gain generation feature importance of the LightGBM model, key features affecting station carbon emissions are identified and screened; and the dimensionality of the feature matrix is reduced based on these key features to reduce model complexity and the risk of overfitting. Logarithmic transformation is applied to all features to address the skewness of the feature distribution, making the feature distribution closer to a normal distribution and improving the model's stability, convergence speed, and prediction accuracy.
[0029] The LightGBM model-based split gain generation of feature importance measures the information gain or Gini gain of a feature when building a decision tree through split nodes.
[0030] Specifically, each time a feature is used to split a node, its splitting gain is accumulated, resulting in the total gain value for each feature. Features with larger gain values contribute more to the model's predictive ability. These features represent important characteristics affecting the carbon emission modeling of railway stations, such as the depth of the station's foundation pit and the carbon emissions from building materials during the station's construction phase. By reducing the feature dimensionality, not only can the model's complexity be reduced, but the risk of overfitting can also be decreased, while simultaneously improving the computational efficiency and predictive performance of the carbon emission prediction model.
[0031] Furthermore, the feature matrix can be processed using polynomial feature generation, which includes interaction terms and higher-order terms between features. Specifically, a quadratic polynomial feature generation method is used to generate interaction terms and higher-order terms between features, enabling the prediction model to capture nonlinear relationships between features and thus improving the model's ability to fit complex patterns. After generating polynomial features, a logarithmic transformation is performed on all features to address the skewness of the feature distribution, making the feature distribution closer to a normal distribution, thereby improving the model's stability and convergence speed.
[0032] The carbon emission prediction ensemble model includes a LightGBM sub-model and an XGBoost sub-model. The construction process includes: establishing a LightGBM-XGBoost ensemble prediction model using a weighted average method based on the carbon emission calculation amount of different stations; and simultaneously optimizing the parameters in the LightGBM model and the XGBoost model using a grid search method to obtain the LightGBM sub-model and the XGBoost sub-model.
[0033] The LightGBM sub-model includes: Histogram binning technology is used to discretize carbon emission characteristic data and clarify the reasonable range of characteristic variables; The GOSS sampling method is used to select representative samples from the data; By leveraging leaf node histogram optimization technology, the model structure is further refined to capture subtle carbon emission variation patterns at different stages of subway construction.
[0034] The XGBoost sub-model includes: The feature variables in the carbon emission feature data are sorted by a pre-sorting algorithm to screen out the key factors that have the most significant impact on carbon emissions. A decision tree model is constructed using a precise greedy splitting method to ensure that the splitting operation of each node can improve the prediction accuracy of the model; Meanwhile, sparse perception optimization technology is introduced to address the sparsity problem commonly found in subway construction data, thereby improving the model's adaptability and generalization ability to complex data environments.
[0035] GBDT can predict target variables by training multiple decision trees. LightGBM uses the GOSS data sampling strategy, which sorts the absolute values of gradients and then only retains the gradients with larger absolute values, while sampling the gradients with smaller absolute values.
[0036] The objective function of the LightGBM algorithm is: (7) In the formula: This represents the objective function value at the t-th iteration. This represents the true value of the i-th sample. Let be the model's prediction value for the i-th sample after the t-th iteration; The number of samples; For the i-th weak learner The regularization term, where t is the current iteration number. This is the loss function.
[0037] XGBoost is also a widely used algorithm within the gradient boosting framework. This algorithm can efficiently process large-scale data and has strong processing and generalization capabilities in various classification problems. It can also handle carbon emission data with missing values. XGBoost innovatively introduces a regularization term and a second derivative into the objective function, considering both the model's fit to the true values and its computational complexity, thereby further improving the model's robustness and effectively avoiding overfitting. The output of the XGBoost algorithm is the sum of all loss functions and regularization terms, and the model is continuously optimized through gradient boosting, iteratively obtaining the optimal prediction model. The objective function of the XGBoost method is as follows: (8) In the formula: The objective function predicted by XGBoost is used as a benchmark to evaluate model performance; The model training error is used to measure the prediction results of the prediction model. and the actual results The differences between them; The regularization term is used to measure the... The complexity of each decision classification tree is used to control the overall complexity of the model and avoid overfitting. For the first The number of leaf nodes in the tree; For the first The weight of the s-th leaf node in the tree; The penalty term is the number of leaf nodes; for Regularization penalty term; K in the objective function of the XGBoost algorithm represents the total number of decision trees in the model ensemble.
[0038] Based on the calculated carbon emissions of different stations, a weighted average method was used to establish a LightGBM-XGBoost ensemble prediction model. Simultaneously, a grid search method was used to optimize the parameters in both the LightGBM and XGBoost models. In this invention, the weights of XGBoost and LightGBM are dynamically calculated based on the root mean square error (RMSE) of the models. The parameters in the LightGBM and XGBoost models affect the model performance and the final prediction results; therefore, selecting an appropriate parameter combination is crucial for obtaining good prediction results.
[0039] The specific implementation steps are as follows: Data Reading and Preprocessing: In predicting carbon emissions at stations, based on the principles of carbon emission calculation, the total carbon emissions for each station at each stage are calculated and the results are saved. This data is then read and processed according to preprocessing methods to ensure data quality and integrity.
[0040] Feature selection: The feature selection method is used to evaluate the contribution of each feature to the total carbon emissions and remove unimportant features, retaining the features that are most helpful for carbon emission prediction.
[0041] Feature generation: To capture the nonlinear relationships between features and improve the model's expressive power, polynomial features are generated. Analysis revealed that features such as the depth of the station's foundation pit, the thickness of the top slab cover, and the number of entrances / exits do not always have a linear impact on the station's total carbon emissions. Therefore, feature generation is used to improve prediction accuracy and more precisely predict carbon emissions.
[0042] Feature transformation and standardization: Given the massive scale of total carbon emissions from each station, data processing may become skewed, impacting model performance. Therefore, feature transformation is performed to stabilize the variance of carbon emission predictions. Data analysis reveals that features influencing carbon emissions have different scales, such as the depth of the station's foundation pit, the thickness of the roof overburden, the number of entrances / exits, and the station area. Failure to address these factors could lead to significant prediction errors in the fitted model. Therefore, feature standardization ensures that different features have the same scale, thereby improving the model's convergence speed and prediction accuracy.
[0043] Model Training and Prediction: 80% of the total sample size was used for model training, and 20% was used for testing. Samples were randomly selected during both training and testing, and the model underwent 1000 random training predictions. Carbon emission data from railway stations was read and trained using the LightGBM-XGBoost ensemble model to generate prediction results.
[0044] Model evaluation: Calculate the R-values of the predicted values in the test set.2 RMSE, MAPE, MAE, and MSE are used to evaluate prediction performance. Among them, the mean absolute error (MAE), mean squared error (MSE), and coefficient of determination (R²) are used to evaluate prediction performance. 2 The root mean square error (RMSE) was used; in addition, the mean absolute percentage error (MAPE) was also used to further evaluate the model's predictive performance; R 2 This indicates the proportion of variation explained by the model, reflecting the goodness of the model's fit.
[0045] Furthermore, the process of building the model and obtaining the predictions also includes: constructing a monotonicity constraint vector, applying monotonicity constraints to preset physical features during model training to restrict the splitting of decision tree nodes and weight updates; calibrating the model's point prediction results using a conformity prediction method, constructing a conformity prediction interval with statistical coverage guarantees by calculating the quantile threshold of inconsistency scores on an independent calibration set; performing interpretability analysis on the prediction results based on the SHAP method, establishing the mapping relationship between features and physical components, calculating the sum of SHAP contribution values of each physical component to achieve carbon emission source tracing, and calculating the local marginal effects of key features.
[0046] Specifically, this includes uncertainty quantification and confidence interval output; to ensure that engineering decisions have a statistically guaranteed description of uncertainty, conformal prediction intervals are used to calibrate point predictions and construct confidence intervals.
[0047] The key points and logic are as follows: First, train the prediction model using a training subset, and then use an independent calibration subset to calculate the model's "inconsistency score" (e.g., absolute residual) under known true values. By taking an appropriate quantile of the calibration set residuals as a threshold, a prediction interval for new samples can be constructed, so that the interval contains the true value with a certain probability.
[0048] The premise of coverage guarantee: Under the assumption of sample commutability, conformity prediction can provide a marginal coverage guarantee for a finite number of samples. In probabilistic terms: For a sample set sampled using the same mechanism, the coverage guarantee for a new sample... Falling within the calibrated interval The probability in the equation satisfies: ; This guarantee is marginal coverage, meaning that the average coverage rate under long-term repeated sampling is no less than [a certain percentage]. ;in, This is a probability operator, used here to describe the randomness under the sampling mechanism; , This represents a new sample randomly selected from its features (which follows the same exchangeable sampling mechanism as calibration / training). The specific implementation process includes: (1) Divide the available data into a training set and a calibration set (the calibration set should be independent of the training process and used to evaluate the residual distribution). (2) Train the model on the training set ; (3) Calculate the inconsistency score (e.g., absolute residual) on the calibration set and obtain the corresponding quantile threshold. The non-conformal scores are calculated as follows: ; in, Let be the inconsistency score of the i-th calibration sample, representing the magnitude of the model prediction error; To determine the true target value (total carbon emissions) of the i-th sample in the calibration set, the unit is consistent with the training target (e.g., kgCO2e). Given a fixed point prediction function (ensemble model), output point estimates for input features x; The feature vector of the i-th sample in the calibration set (containing all modeling features); The number of samples in the calibration set (positive integer).
[0049] Sort these scores in ascending order and take the first... Each value is used as a threshold. ,in significance level (e.g.) =0.05 corresponds to the 95% interval) (4) For new samples Point prediction Construct the compliance interval: ; in, For new samples of Conformity prediction interval (closed interval); For the point prediction values of the new sample (by the ensemble model) (Given) The feature vector of the sample to be predicted.
[0050] (5) As new data arrives, the intervals are re-divided and recalibrated according to the established update strategy (e.g., periodic or quantity-triggered) to maintain the validity of the intervals.
[0051] Furthermore, this also includes physical constraints and monotonicity embedding; The essence of monotonicity constraints: for each feature that needs to be guaranteed to be monotonic. Specify a monotonic direction (e.g., "monotonically non-decreasing") to constrain the model's predictions to prevent a decrease or anomalous fluctuations that contradict engineering physics when the value of that feature changes. That is, when a feature (such as the consumption of a certain building material) increases, the constrained model's prediction of the target (carbon emissions) will not decrease. Expressed as a vector: arrange all candidate features in a fixed order into a vector x=(x1,…,xp). Assign a monotonicity identifier mj to each feature: mj=+1 indicates that the feature is required to be "monotonically non-decreasing," and mj=0 indicates that no monotonic constraint is imposed on the feature. The resulting array m=(m1,…,mp) is the "monotonical constraint vector." The training algorithm must satisfy the monotonic relationship specified by m when searching for the optimal tree structure and leaf node values.
[0052] Monotonicity is generally applicable to numerical or naturally ordered features (such as "material usage" or "transportation distance"). For categorical variables, monotonicity constraints should only be considered if they can be safely mapped to ordered values (and the mapping has engineering physical meaning); otherwise, monotonicity constraints should be avoided for categories.
[0053] Before imposing constraints, verify whether the monotonic relationship holds in the data through exploratory analysis (such as grouped average trends, local regression, or partial dependency plots); if the data is noisy or the sample is sparse, monotonic constraints may lead to a decrease in model performance or underfitting.
[0054] Interrelated characteristics and conflict resolution: When multiple highly correlated features are constrained to the same monotonic direction, implicit conflicts may arise or the model may become too rigid. Strategies to address this include: imposing constraints only on a small number of core features with a clear causal direction in engineering; or reducing conflicts through feature transformation (such as merging correlated variables into a single composite index).
[0055] Further, this includes interpretability analysis and carbon emission tracing; To translate model predictions into engineering-ready emission reduction recommendations, the SHAP (SHapley Additive ex Planations) method is used to interpret and trace the model predictions: SHAP, based on Shapley values in cooperative game theory, provides a consistent and additive decomposition of feature contributions: model predictions can be represented as the baseline value plus the contribution of each feature. This method can provide both global (full sample) importance and detailed local (individual sample / scheme) decompositions, facilitating attribution and decision-making.
[0056] The approach to item-by-item tracing is to establish a mapping relationship between "physical items → feature sets", for example: Materials production: All characteristics that directly represent the quantity / type of materials used; Transportation: transportation distance, mode of transportation, ton-kilometers, etc.; Construction machinery: number of machine shifts, consumption per unit shift, etc.; Indirect electricity consumption: construction power consumption and grid factors, etc. After calculating the SHAP value of each feature for each prediction sample, the SHAP values of features belonging to the same physical item are summed to obtain the net contribution of that item to the prediction (which can be positive or negative). In this way, the total emission prediction can be "traced" to the operational items in the project.
[0057] Marginal effects and decision quantification: Combining the local interpretation of SHAP with local finite difference (small-amplitude perturbation) calculations, the "impact of unit variation on forecasts" can be estimated, which can be used to measure the "emission reduction potential of unit modifications." This is very valuable when selecting material substitutes or deciding on construction configurations.
[0058] Verification and robustness: During implementation, the stability of the SHAP decomposition should be verified (e.g., by multiple resampling or calculating the consistency of SHAP across different model versions), and the uncertainty results (intervals generated by compliance intervals or calibration residuals) should be delivered to decision-makers along with the SHAP contribution to avoid over-reliance on a single point estimate.
[0059] Calculate the SHAP value for the final prediction model using its additive decomposition property: ; in, This serves as a baseline (e.g., the average prediction on the training set). Let be the Shapley contribution value of the j-th feature to the prediction, representing the incremental contribution of that feature to the current prediction. The definition of SHAP is based on the Shapley value in cooperative game theory, satisfying properties such as additiveness and local accuracy. This is the model's predicted value (point estimate) for a specific sample x.
[0060] Map the SHAP values at the feature level to physical components (material production, transportation, construction, machinery fuel, and indirect electricity consumption): establish a corresponding feature set for each physical component (e.g., "reinforcing steel usage" and "concrete usage" are attributed to material production, while "transportation distance" and "transportation mode" are attributed to transportation), and summarize the SHAP contribution by component: ; in, The set of feature indices corresponding to a certain physical component (e.g., all relevant features of material production); The net contribution of this sub-item to the prediction (can be positive or negative); this yields the net contribution of each sub-item to the model prediction (can be positive or negative), thus achieving "source tracing".
[0061] Calculating marginal effects for decision-making reference: Predicting changes in key features using small-amplitude perturbations (finite differences) to estimate local marginal effects. ; Combining unit cost or construction feasibility, it can be used to assess "emission reduction efficiency per unit input"; among which, The finite perturbation step size (small positive number) should be selected in combination with the characteristic dimensions and the engineering feasible step size. This is an approximation of the local marginal effect of the j-th feature.
[0062] Calculate at multiple points for different x values to estimate local nonlinearity (rather than using only a single point). And combine confidence intervals to evaluate stability.
[0063] The commutability assumption states that the coverage of the conformity prediction interval is guaranteed based on the commutability of samples. In the presence of time series drift or strong distribution changes, segmented calibration or sliding / conditional calibration should be used.
[0064] Use of monotonicity constraints: Apply monotonicity constraints only to numerical features that have a unidirectional physical causal relationship in engineering; be cautious or merge features that are categorical or highly correlated first.
[0065] This invention also proposes a carbon emission prediction system for the construction phase of subway stations, comprising: The data acquisition module is used to calculate the carbon emissions of the metro station at different stages according to the carbon emission factor method, and to perform data preprocessing to obtain carbon emission monitoring data. The feature separation module is used to separate features and target variables from the carbon emission monitoring data to obtain a feature matrix and a target variable vector. The feature matrix includes the number of ventilation shafts, station length, width, foundation pit depth, top slab soil thickness, number of station entrances and exits, station area, carbon emissions from building materials during the production stage, carbon emissions from building materials during the transportation stage, and carbon emissions from machinery during the construction stage. The target variable vector includes the total carbon emissions. The feature processing module is used to construct a carbon emission prediction model and perform feature processing on carbon emission monitoring data to obtain carbon emission feature data; including processing carbon emission monitoring data based on LightGBM split gain to generate feature importance; and performing logarithmic transformation on all features. The model building module is used to build an integrated carbon emission prediction model based on LightGBM and XGBoost. The integrated carbon emission prediction model identifies carbon emission characteristic data and uses grid search to optimize the parameters in the LightGBM and XGBoost models to obtain the carbon emission prediction during the construction phase of subway stations.
[0066] Example 2: The invention will be illustrated using 21 stations on Tianjin Metro Line 3 as an example. However, it should be noted that the content and methods covered by this invention are not limited to Tianjin Metro, nor are they limited to any specific region.
[0067] Subway station construction encompasses several key engineering types, such as the main station structure, entrances and ventilation shafts, and foundation pit support. While these engineering types differ in function, they share commonalities in the engineering classification and methods used for carbon emission calculation. Therefore, this invention focuses on the main station structure, which is the most complex and central aspect of subway station construction. This engineering component spans the entire station construction process, involving extensive civil engineering, material usage, and mechanical operations, making its carbon emission calculation highly representative.
[0068] By thoroughly analyzing the carbon emissions of the main structure of the station, the accuracy and effectiveness of the method of this invention are fully demonstrated, thus providing a strong basis for predicting carbon emissions of the entire subway station construction and even similar underground construction projects, and contributing to the low-carbon construction and sustainable development of the urban rail transit sector.
[0069] like Figure 1 The diagram illustrates a flow chart of the carbon emission system for the subway station construction phase of this invention, including the main carbon emission sources during the subway station construction process. Specifically, taking the Tianjin subway construction project as an example, within the system boundaries of this invention, the consumption of raw materials is one of the key factors, starting from the production stage of building materials.
[0070] Different building materials, such as steel bars, cement, and sand, consume significant amounts of energy and generate corresponding carbon emissions during their production. For example, the calcination of limestone and the combustion of fuels during cement production both release carbon dioxide. In the transportation of building materials, those needed for subway station construction often need to be transported from distant production sites to the construction site. The distance of material transportation directly affects energy consumption during transport, thus impacting carbon emissions. Differences in transportation methods, such as road transport and rail transport, also result in varying energy efficiency and carbon emission intensity. During subway construction, various construction machinery burns fuel oil or consumes electricity, generating substantial carbon dioxide emissions. Excavators, cranes, concrete mixers, and other equipment operate continuously during earthwork excavation and structural construction in subway stations, resulting in enormous energy consumption.
[0071] Specifically, this involves examining the various components within the system boundary, including carbon emission activities at different stages of subway station construction. Taking ventilation shaft construction as an example, it involves multiple processes such as the construction of the ventilation shaft structure and equipment installation, each of which generates carbon emissions. During the excavation and support of the station foundation pit, the use of large excavating equipment and the production and installation of support structure materials all generate carbon emissions. Detailed analysis and data collection of these specific activities can cover most of the relevant emission sources during subway station construction.
[0072] Figure 2 The diagram illustrates the data training process for the LightGBM-XGBoost carbon emission prediction ensemble model proposed in this invention. In the initial stage, a comprehensive collection and systematic preprocessing of massive amounts of raw data related to subway construction is performed. This data comes from a wide range of sources and is diverse in dimensions, covering detailed layout parameters of subway stations, station pit depths, the number of ventilation shafts, specific construction processes, detailed usage of various materials, precise selection of construction equipment, and energy consumption.
[0073] Furthermore, after data preprocessing, the process proceeds to the feature engineering stage. Considering the unique and complex characteristics of Tianjin Metro construction data, feature variables closely related to carbon emissions are extracted. In this invention, for the carbon emission dataset, the feature variables are: number of ventilation shafts, station length, width, pit depth, top slab soil thickness, number of station entrances / exits, station area, carbon emissions from building materials during the production phase, carbon emissions from building materials during the transportation phase, and carbon emissions from machinery during the construction phase (X0 to X9), and a target variable (Y), where the target variable represents the total carbon emissions.
[0074] The importance of the features generated by the above feature engineering The formula is as follows: (9) in, and These are the sums of the gradient and second derivative of the left child node, respectively; and These are the sums of the gradient and second derivative of the right child node, respectively; This is a regularization parameter used to prevent overfitting.
[0075] Furthermore, based on the aforementioned feature importance formula, a graph showing the feature importance of subway carbon emissions is generated, as shown below. Figure 3 As shown in the figure. In this invention, only the first eight important features are counted. Features after these have low relevance and are therefore excluded as unimportant features. These features represent important characteristics affecting the carbon emission modeling of stations, such as the depth of the station's foundation pit, the carbon emissions of building materials during the station's construction phase, the station area, and the depth of the station's foundation pit.
[0076] Furthermore, in terms of model construction, the system innovatively consists of two complementary main sub-models. The first is the LightGBM sub-model, which first uses histogram binning technology to scientifically and reasonably discretize the data, thus clarifying the reasonable ranges of the feature variables. Then, the GOSS sampling method is used to select the most representative samples from the subway construction data. Finally, leaf node histogram optimization technology is used to further refine the model structure, enabling the model to accurately capture the subtle carbon emission changes in different details of subway construction.
[0077] The second is the XGBoost sub-model, which uses an advanced pre-sorting algorithm to accurately sort feature variables and quickly identify the key factors with the most significant impact on carbon emissions. Based on this, a high-performance decision tree model is constructed using a precise greedy splitting method, ensuring that the splitting operation at each node maximizes the model's predictive accuracy. Simultaneously, sparsity-aware optimization technology is introduced to fully consider the sparsity problem commonly found in subway construction data, comprehensively improving the model's strong adaptability and generalization ability to complex data environments.
[0078] In this method, the two sub-models are not processed independently but integrated through a dynamic weight allocation mechanism. Based on the model's performance in different subway construction scenarios and feedback data, the weights of the LightGBM and XGBoost sub-models are dynamically adjusted to allow them to complement each other's strengths. This fully leverages the advantages of LightGBM in handling large-scale data and XGBoost in capturing complex nonlinear relationships, thereby constructing an integrated prediction model with superior overall performance. The weight calculation method for LightGBM and XGBoost in the proposed integrated model prediction method is as follows: (10) (11) in, Indicates LightGBM weights; Indicates XGBoost weights; This indicates the root mean square error of LightGBM; This represents the root mean square error of XGBoost.
[0079] Furthermore, in the model evaluation phase, multiple indicators are used to comprehensively and objectively evaluate the performance of the integrated prediction model. In this invention, four evaluation indicators are selected to assess the performance of the carbon emission prediction model: mean absolute error (MAE), mean squared error (MSE), coefficient of determination (R²), and mean squared error (MSE). 2The root mean square error (RMSE) was used to evaluate the model's prediction performance. Additionally, the mean absolute percentage error (MAPE) was employed to further assess the model's predictive ability. MAE represents the average absolute error between the predicted and actual values; MSE represents the average squared error between the predicted and actual values, which helps identify prediction errors in high-carbon emission scenarios; R... 2 The variance explained by the model is expressed as a percentage, reflecting the goodness of the model fit; RMSE is the square root of MSE, suitable for evaluating the model's predictive accuracy; MAPE represents the average absolute percentage error between predicted and actual values, expressed as a percentage. It is suitable for explaining the model's relative error.
[0080] (12) (13) (14) (15) (16) in, This represents the sample mean of the observed values of the dependent variable; After determining the values of the above indicators, the reliability and effectiveness of the model can be accurately judged based on their numerical meanings, ensuring that it can be accurately and stably applied to the carbon emission calculation of subway construction projects.
[0081] Figure 4 The image shows the training set prediction results of the LightGBM-XGBoost carbon emission prediction ensemble model proposed in this invention. Figure 5 The results are shown for the corresponding test set. In this study, the LightGBM and XGBoost ensemble model was used to predict carbon emissions from subway stations, and the model performance was systematically and thoroughly analyzed using the results from the training and test sets. The LightGBM and XGBoost models were trained separately using the training set, and their hyperparameters were continuously adjusted through grid search.
[0082] For the LightGBM model, we tried different combinations of learning rate between 0.01 and 0.2 with intervals of 0.05, number of leaf nodes between 5 and 50 with intervals of 5, minimum data size of leaf nodes between 1 and 10 with intervals of 1, feature sampling ratio between 0.1 and 1 with intervals of 0.1, L1 regularization between 0 and 2 with intervals of 0.2, and L2 regularization between 0 and 2 with intervals of 0.2.
[0083] For the XGBoost model, different combinations of learning rates (0.05 intervals between 0.01 and 0.2), random seeds (1 intervals between 0 and 100), data sampling ratios (0.1 intervals between 0.5 and 1), feature sampling ratios (0.1 intervals between 0.5 and 1), L1 regularization (0.2 intervals between 0 and 1), and L2 regularization (0.2 intervals between 0 and 1) are used.
[0084] Furthermore, through grid search and cross-validation, the prediction errors of the model on the training and test sets for each combination were calculated, such as mean squared error (MSE), root mean square error (RMSE), and mean absolute error (MAE). After extensive trials and calculations, the optimal hyperparameter combination for the model's performance on the validation set was finally found. The optimal learning rate for the LightGBM model was determined to be 0.05, the minimum number of data points per leaf node was 8, the feature sampling ratio was 0.5, L1 regularization was 1.0, and L2 regularization was 1.0. The optimal learning rate for the XGBoost model was 0.05, the random seed was 42, the data sampling ratio was 0.8, the feature sampling ratio was 0.8, L1 regularization was 1.0, and L2 regularization was 1.0. Finally, the trained LightGBM and XGBoost models were combined using a weighted average. In-depth analysis of the training set data revealed a high degree of consistency between predicted and actual carbon emissions across station sample numbers 0 to 17. Particularly in the sub-intervals of station sample numbers 4 to 7 and 12 to 15, the fluctuation trends of predicted and actual values almost completely overlapped, indicating that the model possesses a strong ability to capture and fit data features during training.
[0085] Figure 6 The R-value of the LightGBM-XGBoost carbon emission prediction ensemble model proposed in this invention is... 2 RMSE, MAE, and MSE values.
[0086] Specifically, the model accurately reflects subtle changes in carbon emissions, precisely predicting both upward and downward trends. This high degree of consistency not only demonstrates the model's stability and reliability when handling complex carbon emission data but also validates its generalization ability and adaptability in subway engineering scenarios. Through this accurate prediction, the model can provide reliable data support for carbon emission management in subway station construction, helping engineers better understand and control the dynamic changes in carbon emissions.
[0087] Furthermore, the analysis of the test set data further validated the robustness and practicality of the model. Within the range of station sample numbers 0 to 4, the predicted carbon emissions also showed a remarkable degree of agreement with the actual values. Particularly at station sample numbers 1, 2, and 3, the predicted carbon emissions were almost perfectly consistent with the actual values, indicating that the model can maintain its high-precision predictive performance even on unseen data.
[0088] Through this reliable and accurate prediction, the model can effectively capture the changing trends of carbon emissions, providing valuable decision-making support for carbon emission management in subway projects. Engineering technicians can use the model's predictions to develop carbon reduction plans in advance and optimize resource allocation.
[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.
Claims
1. A method for predicting carbon emissions during the construction phase of a subway station, characterized in that, include: Carbon emissions at different stages of a metro station are calculated using the carbon emission factor method, and data preprocessing is performed to obtain carbon emission monitoring data. The carbon emission monitoring data is separated into feature matrix and target variable vector; A carbon emission prediction model is constructed to perform feature processing on carbon emission monitoring data to obtain carbon emission characteristic data. This includes processing carbon emission monitoring data using LightGBM-based split gain to generate feature importance; and performing logarithmic transformation on all features. A carbon emission prediction ensemble model was constructed based on the LightGBM and XGBoost models. The carbon emission prediction ensemble model was used to identify carbon emission characteristic data. Grid search was used to optimize the parameters in the LightGBM and XGBoost models to obtain the predicted carbon emissions during the construction phase of subway stations. The process of building the model and obtaining the predictions also includes: constructing a monotonicity constraint vector, applying monotonicity constraints to preset physical features during model training to restrict the splitting of decision tree nodes and weight updates; calibrating the model's point prediction results using a conformity prediction method, constructing a conformity prediction interval with statistical coverage guarantees by calculating the quantile threshold of inconsistency scores on an independent calibration set; performing interpretability analysis on the prediction results based on the SHAP method, establishing the mapping relationship between features and physical components, calculating the sum of SHAP contributions of each physical component to achieve carbon emission source tracing, and calculating the local marginal effects of key features.
2. The method for predicting carbon emissions during the construction phase of a subway station according to claim 1, characterized in that, The carbon emissions of a metro station at different stages are calculated using the carbon emission coefficient method. These stages include: building material production, transportation, and construction. The feature matrix includes the number of ventilation shafts, station length, width, foundation pit depth, top slab soil cover thickness, number of station entrances / exits, station area, carbon emissions of building materials during the production stage, carbon emissions of building materials during the transportation stage, and carbon emissions of machinery during the construction stage. The target variable vector includes total carbon emissions.
3. The method for predicting carbon emissions during the construction phase of a subway station according to claim 1, characterized in that, The data preprocessing includes: using column mean imputation to complete the data in the original carbon emission monitoring data of the station; and using statistical principles to fill in missing samples by calculating the mean of non-missing values in each feature column.
4. The method for predicting carbon emissions during the construction phase of a subway station according to claim 1, characterized in that, The carbon emission prediction model includes: Based on the split gain generation feature importance of the LightGBM model, key features affecting station carbon emissions are identified and screened; and the dimensionality of the feature matrix is reduced based on these key features to reduce model complexity and the risk of overfitting. Logarithmic transformation is applied to all features to address the skewness of the feature distribution, making the feature distribution closer to a normal distribution and improving the model's stability, convergence speed, and prediction accuracy.
5. The method for predicting carbon emissions during the construction phase of a subway station according to claim 1, characterized in that, The carbon emission prediction ensemble model includes a LightGBM sub-model and an XGBoost sub-model, and the construction process includes: Based on the carbon emission calculations for different stations, a weighted average method was used to establish a LightGBM-XGBoost integrated prediction model. At the same time, a grid search method was used to optimize the parameters in the LightGBM model and the XGBoost model respectively.
6. The method for predicting carbon emissions during the construction phase of a subway station according to claim 5, characterized in that, The LightGBM sub-model includes: Histogram binning technology is used to discretize carbon emission characteristic data and clarify the reasonable range of characteristic variables; The GOSS sampling method is used to select representative samples from the data; By leveraging leaf node histogram optimization technology, the model structure is further refined to capture subtle carbon emission variation patterns at different stages of subway construction.
7. The method for predicting carbon emissions during the construction phase of a subway station according to claim 5, characterized in that, The XGBoost sub-model includes: The feature variables in the carbon emission feature data are sorted by a pre-sorting algorithm to screen out the key factors that have the most significant impact on carbon emissions. A decision tree model is constructed using a precise greedy splitting method to ensure that the splitting operation of each node can improve the prediction accuracy of the model; Meanwhile, sparse perception optimization technology is introduced to address the sparsity problem commonly found in subway construction data, thereby improving the model's adaptability and generalization ability to complex data environments.
8. A carbon emission prediction system for the construction phase of a subway station, characterized in that, include: The data acquisition module is used to calculate the carbon emissions of the metro station at different stages according to the carbon emission factor method, and to perform data preprocessing to obtain carbon emission monitoring data. The feature separation module is used to separate features and target variables from the carbon emission monitoring data to obtain a feature matrix and a target variable vector. The feature matrix includes the number of ventilation shafts, station length, width, foundation pit depth, top slab soil thickness, number of station entrances and exits, station area, carbon emissions from building materials during the production stage, carbon emissions from building materials during the transportation stage, and carbon emissions from machinery during the construction stage. The target variable vector includes the total carbon emissions. The feature processing module is used to construct a carbon emission prediction model and perform feature processing on carbon emission monitoring data to obtain carbon emission feature data; including processing carbon emission monitoring data with split gain based on the LightGBM model to generate feature importance; and performing logarithmic transformation on all features. The model building module is used to construct an integrated carbon emission prediction model based on the LightGBM and XGBoost models. The integrated carbon emission prediction model identifies carbon emission characteristic data and uses grid search to optimize the parameters in the LightGBM and XGBoost models to obtain the carbon emission prediction amount during the construction phase of subway stations.