A machine learning modeling method and system for dynamic prediction of building construction project progress

Abstract

This invention discloses a machine learning modeling method and system for dynamic prediction of construction project progress. The method includes: collecting multi-type related data throughout the entire lifecycle of the construction project; removing invalid data; and optimizing data quality through time series unification, standardization, and dimensional reduction. A feature system incorporating process correlation, resource constraints, and risk quantification is constructed, with weights dynamically adjusted according to differences in construction stages to achieve feature-scenario adaptation. A hybrid architecture combining time series prediction and ensemble learning is built, introducing an attention mechanism to enhance key factor extraction; and the optimal model configuration is determined through hyperparameter optimization combined with cross-validation. Model parameters are updated incrementally based on a preset time window, automatically optimizing when event conditions are met or prediction errors exceed a threshold. A multi-level risk warning system is constructed, setting warning thresholds, locating the causes of deviations, and providing correction suggestions. This improves the precision of construction project progress management.

Description

Technical Field

[0001] This invention belongs to the field of dynamic prediction technology for engineering progress, and more specifically, relates to a machine learning modeling method and system for dynamic prediction of the progress of building construction projects. Background Technology

[0002] In the field of building construction project management, schedule control is directly related to project cost control, quality assurance, and delivery efficiency. Accurate schedule forecasting is a core prerequisite for achieving refined management. As building construction projects expand in scale and construction processes become more complex, the factors influencing schedule exhibit characteristics of multi-source, dynamic, and coupled nature, making traditional schedule forecasting methods insufficient to meet actual management needs.

[0003] Currently, project schedule forecasting for building construction projects largely relies on the experience and judgment of management personnel or simple mathematical statistical models. Experience-based judgment is limited by individual cognitive limitations and lacks a systematic analysis of the synergistic effects of multiple factors, making it prone to deviations, especially in schedule projections across construction phases. While simple mathematical statistical models can quantify the impact of some factors, they cannot effectively capture the time-series dependencies of schedule data, are difficult to adapt to scenarios with clear time-series characteristics such as work process connections and resource scheduling, and are insufficient in characterizing the non-linear relationships between multiple factors such as resource constraints, extreme weather, and design changes, resulting in significant deviations between predicted results and actual schedules.

[0004] Meanwhile, existing methods generally lack dynamic adaptability. During project implementation, dynamic factors such as fluctuations in material supply, adjustments in construction personnel, and changes in policy requirements can easily cause the prediction model to fail, requiring manual recalibration of model parameters. This not only increases management costs but also affects the timeliness of schedule control due to the lag in adjustments. In addition, most prediction methods only output schedule prediction results and do not establish supporting risk warning and cause identification mechanisms. When schedule deviations occur, managers find it difficult to quickly identify the core influencing factors and can only passively take remedial measures, failing to achieve early risk prevention and control, further exacerbating the risk of schedule delays. Therefore, developing a modeling method that can adapt to the complex scenarios of building construction projects and has dynamic prediction and risk control capabilities is of great practical significance for improving the level of schedule management. Summary of the Invention

[0005] This invention aims to address the problems of traditional building construction project schedule prediction methods, such as reliance on experience, poor model adaptability, lack of dynamic adjustment and risk warning capabilities. It provides a machine learning-based dynamic prediction modeling method to achieve accurate schedule prediction, dynamic model optimization and proactive risk control, thereby improving the level of precision in building construction project schedule management.

[0006] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a machine learning modeling method for dynamic prediction of construction project progress, comprising: S1. Collect various types of related data that affect the progress of the building construction project throughout its entire lifecycle and remove invalid data. Regularize the time dimension of the multi-source data through a time series unification method, and optimize data quality and reduce processing complexity by using standardization and dimension reduction methods. S2. Construct a feature system that includes process associations, resource constraints, and risk quantification; based on the differences in scenarios at different construction stages of the project, adopt a dynamic weight adjustment method to adaptively optimize the influence weight of each feature, and complete the real-time adaptation of features to construction scenarios; S3. Construct a hybrid architecture combining a time-series prediction model and an ensemble learning model. The time-series prediction model captures the time-series dependencies of progress, while the ensemble learning model fits the nonlinear relationships of multiple factors. An attention mechanism is introduced into the time-series prediction model to enhance the extraction of key factor features in the building construction scenario. Hyperparameter optimization combined with cross-validation is used to determine the optimal configuration of the model. S4. Based on a preset time window, incremental learning is used to update the local parameters of the model; when the preset event conditions are met or the prediction error exceeds the threshold, the model optimization is automatically started; a multi-level risk early warning system is constructed, and the early warning threshold is set by combining the deviation index and the risk propagation probability. The cause of the deviation is located by the contribution of the influencing factors and correction suggestions are given.

[0007] Furthermore, the multi-type associated data in S1 specifically includes: Process execution data: at least covering the actual time spent on each sub-project, the interval between process connections, the quality acceptance results of the completed process, and the progress of the critical path process; Resource allocation related data: at least covering material supply cycle, material inventory level, equipment failure rate, equipment maintenance cycle, number of personnel, distribution of personnel skill levels, and resource scheduling response time; External environmental data should include at least daily precipitation, average daily temperature, wind speed, geological disaster early warning information, lighting conditions at the construction site, and frequency of extreme weather events. Project management related data: at least including design change records, supervisor's rectification opinions, safety accident records, progress inspection reports, contract period adjustment documents, and minutes of on-site coordination meetings.

[0008] Furthermore, the dimensionality reduction method in S1 specifically involves using Principal Component Analysis (PCA) or Linear Discriminant Analysis (LDA) to extract key influencing factors from multi-source data and eliminate redundant features, thereby reducing data dimensionality and subsequent model computational complexity while retaining core progress-related information.

[0009] Furthermore, the feature system in S2 is constructed based on the process characteristics of building construction projects, resource constraint rules, and risk transmission logic. The specific process is as follows: Based on the process characteristics of building construction projects, the connection coefficients, cross-influence factors and parallel operation efficiency of different sub-projects are extracted to form a subset of process-related features. Based on the resource constraint patterns of building construction projects, we quantify the impact weight of the construction period corresponding to the material inventory early warning threshold, the transmission coefficient of equipment failure repair time on process delays, and the correlation between the proportion of manpower shortage and construction efficiency, thus forming a subset of resource constraint features. Based on the risk transmission logic of building construction projects, a mapping relationship and quantitative indicators between risk factors, including weather, geology, and design changes, and construction progress are established to form a subset of risk quantitative characteristics. By integrating the aforementioned process-related feature subset, resource constraint feature subset, and risk quantification feature subset, a complete feature system for predicting the progress of building construction projects is formed.

[0010] Furthermore, the real-time adaptation process of the features and the construction scenario in S2 is as follows: The expression for its construction scenario adaptation function is: , in, The feature output values ​​are adapted to the current construction scenario; The actual execution intensity of the core process in the current construction phase is defined as the ratio of the actual progress rate of the core process in the current phase to the standard progress rate of that type of process. The resource consumption rate at the current construction stage is defined as the ratio of the actual consumption to the planned consumption of core resources per unit time at the current stage. The probability of risk occurrence in the current construction phase is defined as the ratio of the actual frequency of risk events occurring within the preset period of the current phase to the average frequency of risk events occurring in similar historical phases. This refers to the total number of construction stages included in the entire lifecycle of a building construction project. For the first The standard execution intensity of core procedures in each construction phase For the first Standard resource consumption rate for each construction phase For the first The standard risk probability for each construction phase is obtained based on statistical data from similar historical building construction projects. These are the basic feature values ​​in the feature system, namely the initial quantification values ​​of the process-related feature subset, the resource constraint feature subset, and the risk quantification feature subset.

[0011] Furthermore, the specific method for capturing the time series dependencies of progress through the time series prediction model in S3 is as follows: Long Short-Term Memory (LSTM) network is selected as the core model for time series prediction. Based on multi-source associated data with regularized time dimensions, a progress time series sample set is constructed. With a preset time granularity as the step size, feature data within consecutive time periods are associated with the progress completion status of the corresponding time periods, forming a mapping relationship between input sequences and output labels. The intrinsic correlation between time series is quantified through a time dependency capture function, the expression of which is: , in, For the first The progress time series dependency of a time period is used to characterize the strength of the correlation between historical time period data and the progress of the current time period; For the first The actual progress completed during the time period is derived from the pre-processed data related to the execution of the work process. For the first Time period and the The temporal correlation degree of the time period is defined as the ratio of the overlap of the core processes of two time periods to the time interval. The overlap of the core processes is determined based on the process correlation characteristics, and the time interval is the time span of the two time periods. The data comes from the process correlation characteristics in the feature system. The total number of historical time periods participating in the time-series dependency calculation is determined based on the cyclical characteristics of the construction project's work processes; For the first The actual progress completed during the time period The first output of the time series prediction model Forecasted progress for the specified time period; The hidden layers of the LSTM model calculate temporal dependencies on a time-by-time basis. The study analyzes the temporal correlation of process connections between adjacent time periods, the correlation of the progress patterns of similar processes in different cycles, and the temporal transmission correlation of critical path processes to the overall progress. During model training, time-series dependency To constrain this, the parameters of the model's gating units are iteratively optimized to minimize the deviation between the predicted and actual progress in the time dimension. This enhances the model's ability to capture time series dependencies in long-cycle, multi-process scenarios, ensuring that the model accurately extracts the inherent correlation patterns of the progress of building construction projects over time.

[0012] Furthermore, the specific process of fitting multi-factor nonlinear associations using an ensemble learning model in S3 is as follows: Random forest is selected as the core model for ensemble learning. The feature output values ​​adapted to the construction scenario are used as input data to construct a multi-factor association training sample set. Each sample in the set contains dynamic quantified values ​​of process association, resource constraints, and risk quantification-related features, along with the actual progress deviation value for the corresponding time period. The impact of the synergistic effect of multiple factors on the progress is quantified through a nonlinear association fitting function. The expression of the nonlinear association fitting function is as follows: , in, For the first The nonlinear correlation impact value of multiple factors on the progress during a time period is used to characterize the comprehensive effect of multiple factors on the progress of the current time period under synergistic effect; For the first The output values ​​of the class features after being adapted to the current construction scenario cover the class features of process association, resource constraints, and risk quantification. The data comes from the adaptation results of features and construction scenarios. For the first Class features in the first The scene correlation degree of a time period is defined as the ratio of the actual occurrence frequency of the influencing factors corresponding to this type of feature in the current time period to the average occurrence frequency of this type of factor throughout the entire construction cycle. The data comes from preprocessed multi-source correlation data. This refers to the total number of feature categories included in the feature system, namely, three types of features: process association, resource constraints, and risk quantification. For the first The frequency of multi-factor interactions during a time period is defined as the number of times that influencing factors corresponding to different categories of features occur simultaneously within the current time period. The data is obtained by statistically analyzing the occurrence records of various features in the feature system. Random forest models learn from the training sample set in parallel by constructing multiple decision trees, with each decision tree based on the nonlinear association influence value. The local nonlinear relationship between a single factor and progress is analyzed, and then the outputs of all decision trees are merged through an integrated voting mechanism to fit the global nonlinear correlation under the synergistic effect of multiple factors. During model training, the influence value of the nonlinear correlation is used. The difference between the actual value of the schedule deviation and the actual value is used as the optimization target. The branch nodes and depth of the decision tree are iteratively adjusted to improve the fitting accuracy of complex nonlinear relationships of multiple factors.

[0013] Furthermore, the specific process of determining the optimal model configuration using hyperparameter optimization combined with cross-validation in S3 is as follows: First, a hyperparameter candidate space is constructed, which covers the network structure parameters and training process parameters of the temporal prediction model, as well as the basic structure parameters and ensemble rule parameters of the ensemble learning model. The hyperparameters of the temporal prediction model include the number of hidden layer nodes and the training step size; the hyperparameters of the ensemble learning model include the number of base learners and the base learner decision boundary splitting threshold. The model training dataset is constructed based on the feature data adapted to the construction scenario. The training dataset is divided into several mutually exclusive subsets using a stratified sampling method. The number of subsets is determined based on the total number of construction stages of the building project. Each set of hyperparameter configurations in the hyperparameter candidate space is quantitatively evaluated using a hyperparameter configuration evaluation function, which is: , in, The comprehensive evaluation value for the hyperparameter configuration is used to characterize the predictive performance and adaptability of the model under the evaluation function expression configuration; This is the number of folds in the cross-validation, i.e., the number of subsets. For the first The actual progress values ​​corresponding to each subset of data are derived from preprocessed process execution data. For the first The predicted progress values ​​output by the model for each subset of the dataset under the current hyperparameter configuration; For the first The planned schedule values ​​corresponding to each subset of data are derived from project management data. For the first The scenario complexity coefficient of each subset of data is defined as the ratio of the number of construction stage types contained in the subset of data to the total number of construction stages in the entire project. It is used to correct the evaluation bias under different scenario complexities. Each subset of the dataset is used as the validation set, and the remaining subset is used as the training set. A hybrid machine learning model is trained based on the current hyperparameter configuration, and the comprehensive evaluation value of the configuration on the validation set is calculated through the hyperparameter configuration evaluation function. Traverse all configurations in the hyperparameter candidate space, repeat the cross-validation process, and record the comprehensive evaluation value corresponding to each configuration. Select the hyperparameter configuration with the largest comprehensive evaluation value as the optimal configuration of the model. Under this configuration, the model can achieve the optimal balance between prediction performance and adaptability in each construction stage scenario.

[0014] As a second aspect of the present invention, a machine learning modeling system for dynamic prediction of the progress of building construction projects is also provided, comprising: The multi-source data acquisition and preprocessing unit is used to collect various types of related data that affect the progress of the building construction project throughout its entire lifecycle and remove invalid data. It uses a time-series unification method to organize the time dimension of the multi-source data and adopts standardization and dimension reduction methods to optimize data quality and reduce processing complexity. The building construction feature system adaptation unit is used to construct a feature system that includes process associations, resource constraints, and risk quantification. Based on the differences in scenarios at different construction stages of the project, the influence weight of each feature is adaptively optimized using a dynamic weight adjustment method to achieve real-time adaptation between features and construction scenarios. A hybrid model construction optimization unit is used to build a cascaded hybrid architecture of a time series prediction model and an ensemble learning model. The time series prediction model captures the time series dependencies of progress, and the ensemble learning model fits the nonlinear correlation of multiple factors. An attention mechanism is introduced into the time series prediction model to enhance the extraction of key factor features in the building construction scenario. Hyperparameter optimization combined with cross-validation is used to determine the optimal configuration of the model. The dynamic risk warning unit is used to update local parameters of the model using incremental learning based on a preset time window; when preset event conditions are met or the prediction error exceeds the threshold, the model optimization is automatically initiated; a multi-level risk warning system is constructed, and the warning threshold is set by combining deviation indicators and risk propagation probability, and the cause of deviation is located by the contribution of influencing factors and correction suggestions are given.

[0015] As a third aspect of the invention, a computer-readable storage medium is also provided, on which a computer program is stored, which is executed by a processor, according to any one of the machine learning modeling methods for dynamic prediction of the progress of a building construction project.

[0016] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: 1. The machine learning modeling method for dynamic prediction of construction project progress in this invention collects multi-type related data throughout the entire project lifecycle and removes invalid data. It then uses a time-series unification method to organize the data's time dimension and combines standardization and dimensional reduction techniques to optimize data quality. This effectively solves the problems of scattered and heterogeneous data in construction projects, ensuring comprehensive coverage of core influencing factors of progress while reducing the complexity of subsequent modeling, thus providing high-quality basic data support for accurate prediction. Simultaneously, it constructs a feature system encompassing process correlation, resource constraints, and risk quantification, and dynamically adjusts feature weights according to differences in construction stages. This allows features to adapt to different construction scenarios in real time, avoiding the problem of insufficient adaptability between general features and construction scenarios. This makes the feature's portrayal of the progress's impact more closely aligned with actual construction logic and improves the accuracy of feature mapping to progress patterns.

[0017] 2. The machine learning modeling method for dynamic prediction of building construction project progress in this invention constructs a hybrid architecture combining a time-series prediction model and an ensemble learning model. The time-series prediction model captures the temporal dependencies of the project progress, while the ensemble learning model fits the nonlinear relationships of multiple factors. Furthermore, an attention mechanism is introduced into the time-series prediction model to enhance feature extraction of key factors. This approach accurately covers the temporal continuity of building construction progress and effectively characterizes the synergistic effects of multiple factors, avoiding the limitations of single models in adapting to complex scenarios. Based on this, hyperparameter optimization combined with cross-validation is used to determine the optimal model configuration, further improving the prediction accuracy and stability of the hybrid model and ensuring more balanced and reliable performance across different construction stages.

[0018] 3. The machine learning modeling method for dynamic prediction of building construction project progress in this invention updates local model parameters through incremental learning based on a preset time window. It automatically initiates model optimization when preset event conditions are met or the prediction error exceeds a threshold, achieving real-time adaptation of the model to dynamic project changes. This solves the problem that traditional static models cannot keep up with project changes and ensures the timeliness of prediction results. Simultaneously, a multi-level risk early warning system is constructed, setting early warning thresholds by combining deviation indicators and risk propagation probabilities. Furthermore, it identifies the causes of deviations and provides correction suggestions based on the contribution of influencing factors, transforming passive response to schedule delays into proactive risk management. This helps project owners identify and resolve schedule risks in advance, improving the initiative and efficiency of building construction project progress management. Attached Figure Description

[0019] Figure 1 This is a flowchart of a machine learning modeling method for dynamic prediction of the progress of a building construction project, according to an embodiment of the present invention. Figure 2 This is an overall schematic diagram of the Yantan renovation project in the Shibei District of a certain city, according to an embodiment of the present invention. Figure 3 This is a system unit diagram of an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0021] Example 1 Please refer to Figure 1 This embodiment 1 provides a machine learning modeling method for dynamic prediction of the progress of building construction projects, including: S1. Collect various types of related data that affect the progress of the building construction project throughout its entire lifecycle and remove invalid data. Regularize the time dimension of the multi-source data through a time series unification method, and optimize data quality and reduce processing complexity by using standardization and dimension reduction methods. S2. Construct a feature system that includes process associations, resource constraints, and risk quantification; based on the differences in scenarios at different construction stages of the project, adopt a dynamic weight adjustment method to adaptively optimize the influence weight of each feature, and complete the real-time adaptation of features to construction scenarios; S3. Construct a hybrid architecture combining a time-series prediction model and an ensemble learning model. The time-series prediction model captures the time-series dependencies of progress, while the ensemble learning model fits the nonlinear relationships of multiple factors. An attention mechanism is introduced into the time-series prediction model to enhance the extraction of key factor features in the building construction scenario. Hyperparameter optimization combined with cross-validation is used to determine the optimal configuration of the model. S4. Based on a preset time window, incremental learning is used to update the local parameters of the model; when the preset event conditions are met or the prediction error exceeds the threshold, the model optimization is automatically started; a multi-level risk early warning system is constructed, and the early warning threshold is set by combining the deviation index and the risk propagation probability. The cause of the deviation is located by the contribution of the influencing factors and correction suggestions are given.

[0022] This embodiment 1 further elaborates on the above steps.

[0023] (1) Multi-source data acquisition and preprocessing In the context of building construction project progress management, the comprehensiveness and effectiveness of multi-source data directly determine the accuracy of subsequent progress prediction models. However, in actual projects, process, resource, environmental, and management-related data are often scattered and heterogeneous, with inconsistent time dimensions and a large amount of invalid and redundant information, which hinders the modeling work.

[0024] Based on this, it is necessary to collect various types of related data affecting the progress throughout the entire project lifecycle, specifically covering four core categories of data: Data related to process execution, including the actual time spent on each sub-project, the interval between process connections, the quality acceptance results of completed processes, and the progress of critical path processes; this type of data directly reflects the actual pace of construction progress. Data related to resource allocation, including material supply cycles, material inventory levels, equipment failure rates, equipment maintenance cycles, manpower allocation, personnel skill level distribution, and resource scheduling response timeliness; this type of data reflects the resource guarantee capability during construction. Data related to the external environment, including daily precipitation, average daily temperature, wind speed, geological disaster early warning information, on-site lighting conditions, and the frequency of extreme weather events; this type of data is a key factor affecting the efficiency of outdoor construction. Data related to project management, including design change records, supervisor rectification opinions, safety accident records, progress inspection reports, contract period adjustment documents, and on-site coordination meeting minutes; this type of data covers various adjustments and interventions at the project management level.

[0025] After data collection, invalid data and noise interference must be removed from the dataset. Then, a time-series unification method is used to standardize the time dimension of the multi-source data, ensuring consistency in time scale across data from different sources and collection frequencies, laying the foundation for subsequent data analysis. Building on this, standardization techniques are used to eliminate dimensional differences between different data points, and then dimensionality reduction techniques are employed to optimize data quality. Specifically, Principal Component Analysis (PCA) or Linear Discriminant Analysis (LDA) are used to extract key influencing factors from the multi-source data, remove redundant features, and reduce data dimensionality and subsequent model computational complexity while retaining core progress-related information, thereby improving modeling efficiency and accuracy.

[0026] In the multi-source data processing of building construction projects, the time-series unification method is a standardized processing method that organizes progress-related data from different sources, with different collection frequencies and different timestamps into a unified time scale. The core objective is to eliminate the heterogeneity of data in the time dimension, ensure the consistency of data on the time base, and provide a unified time-series foundation for subsequent feature construction and model input.

[0027] In practice, the first step is to define a unified time granularity. In combination with the construction rhythm of building projects, "day" or "week" is usually selected as the basic time unit. The "day" granularity can be used in short-term, intensive work phases, while the "week" granularity can be used in long-term phases such as the construction of the main structure. This ensures that the time unit can accurately capture the details of progress changes while avoiding data redundancy.

[0028] Subsequently, an adaptive processing method was adopted for the time series characteristics of different types of data: For data with a collection frequency higher than the uniform granularity, such as daily precipitation and real-time equipment operating status, the average or summation method was used to aggregate them into a uniform time unit; For data with a collection frequency lower than the uniform granularity, such as design change records and monthly progress inspection reports, the timestamp mapping method was used to associate it with the corresponding uniform time unit, clarifying the specific period of effect of the data on the time series axis; For data without a clear timestamp, such as relatively stable information like the distribution of personnel skill levels, its first effective time was used as the starting point, extending to all uniform time units before the information was changed.

[0029] Finally, a unified time series data matrix is ​​constructed, with a unified time unit as the row index and various data types as column vectors, forming a structured time series dataset to ensure that different data can be associated and matched in the same time dimension in subsequent analysis.

[0030] (2) Adaptation of building feature system In the process of predicting the progress of building construction projects, the rationality of the feature system directly determines the model's adaptability to the construction scenario. However, there are significant differences in the key procedures, resource requirements, and risk types of different construction stages. The modeling method with fixed feature weights is difficult to accurately match the dynamically changing construction scenario. Therefore, it is necessary to build a feature system that fits the actual project and achieve real-time adaptation between features and construction scenarios.

[0031] First, a complete schedule prediction feature system is constructed based on the process characteristics, resource constraints, and risk transmission logic of building construction projects. Based on the process characteristics, the connection coefficients, cross-influence factors, and parallel operation efficiency between different sub-projects are extracted to form a process-related feature subset, which characterizes the inherent connections and progress patterns between processes. Combining the resource constraints of building construction projects, the impact of material inventory warning thresholds on the construction period, the transmission relationship of equipment failure repair time on process delays, and the relationship between manpower shortage ratios and construction efficiency are quantified to form a resource constraint feature subset, which reflects the constraint effect of resource availability on construction progress. Based on the risk transmission logic of building construction projects, the mapping relationship and quantitative indicators between risk factors such as weather, geology, and design changes and construction progress are established to form a risk quantification feature subset, which reflects the degree of interference of various risk factors on the schedule. Integrating the process-related feature subset, resource constraint feature subset, and risk quantification feature subset constitutes a complete building construction project schedule prediction feature system.

[0032] Subsequently, based on the different scenarios at different construction stages of the project, a dynamic weight adjustment method was used to adaptively optimize the influence weights of each feature, completing the real-time adaptation of features to the construction scenario. The process of real-time adaptation of features to the construction scenario is as follows: The construction scenario adaptation function expression is constructed as follows: , in, The feature output values ​​are adapted to the current construction scenario; The actual execution intensity of the core process in the current construction phase is defined as the ratio of the actual progress rate of the core process in the current phase to the standard progress rate of the process. The data comes from the pre-processed process execution data. The resource consumption rate at the current construction stage is defined as the ratio of the actual consumption to the planned consumption of core resources per unit time at the current stage. The data comes from preprocessed resource allocation data. The probability of risk occurrence in the current construction phase is defined as the ratio of the actual frequency of risk events occurring within the preset period of the current phase to the average frequency of risk events occurring in similar historical phases. The data comes from the pre-processed external environment and project management-related data. This refers to the total number of construction stages (such as foundation construction, main structure construction, etc.) included in the entire lifecycle of a building construction project. For the first The standard execution intensity of core procedures in each construction phase For the first Standard resource consumption rate for each construction phase For the first The standard risk probability for each construction phase is obtained based on statistical data from similar historical building construction projects. These are the basic feature values ​​in the feature system, namely the initial quantification values ​​of the process-related feature subset, the resource constraint feature subset, and the risk quantification feature subset; Through the above functions, the feature output value can dynamically change with the current construction stage's process execution status, resource consumption, and risk situation, realizing real-time adaptation of features to the construction scenario. The adaptation process is directly related to the preprocessing results of multi-source associated data, ensuring the coherence and consistency of the adaptation logic and the entire modeling method, so that the features always fit the actual situation of the current construction scenario and improve the accuracy of subsequent model predictions.

[0033] (3) Hybrid model construction optimization In the scenario of predicting the progress of building construction projects, a single model is difficult to accurately capture the time series dependence of progress and the nonlinear correlation of multiple factors at the same time. Therefore, it is necessary to build a hybrid architecture that combines time series prediction model and ensemble learning model to achieve high-precision prediction by combining the advantages of the two models.

[0034] First, a Long Short-Term Memory (LSTM) network is selected as the core model for time series prediction. Based on multi-source associated data after normalization of the time dimension, a progress time series sample set is constructed. With a preset time granularity as the step size, feature data within consecutive time periods are associated with the progress completion status of the corresponding time periods, forming a mapping relationship between input sequences and output labels. Simultaneously, an attention mechanism is introduced into the time series prediction model. By calculating the progress time series dependency, the temporal correlation of process connections between adjacent time periods, the correlation of the advancement patterns of similar processes in different cycles, and the temporal transmission correlation of critical path processes to the overall progress are analyzed. During model training, the parameters of the model gating units are iteratively optimized using the temporal dependency as a constraint to minimize the deviation between predicted and actual progress in the time dimension. This strengthens the model's ability to capture time series dependencies in long-cycle, multi-process overlapping scenarios, accurately extracting the inherent correlation patterns of construction project progress over time.

[0035] Specifically, in a preferred embodiment, the method for capturing progress time series dependencies using a time series prediction model is as follows: Long Short-Term Memory (LSTM) network is selected as the core model for time series prediction. Based on multi-source associated data with regularized time dimensions, a progress time series sample set is constructed. With a preset time granularity as the step size, feature data within consecutive time periods are associated with the progress completion status of the corresponding time periods, forming a mapping relationship between input sequences and output labels. The intrinsic correlation between time series is quantified through a time dependency capture function, the expression of which is: , in, For the first The progress time series dependency of a time period is used to characterize the strength of the correlation between historical time period data and the progress of the current time period; For the first The actual progress completed during the time period is derived from the pre-processed data related to the execution of the work process. For the first Time period and the The temporal correlation degree of the time period is defined as the ratio of the overlap of the core processes of two time periods to the time interval. The overlap of the core processes is determined based on the process correlation characteristics, and the time interval is the time span of the two time periods. The data comes from the process correlation characteristics in the feature system. The total number of historical time periods participating in the time-series dependency calculation is determined based on the cyclical characteristics of the construction project's work processes; For the first The actual progress completed during the time period The first output of the time series prediction model Forecasted progress for the specified time period; The hidden layers of the LSTM model calculate temporal dependencies on a time-by-time basis. The study analyzes the temporal correlation of process connections between adjacent time periods, the correlation of the progress patterns of similar processes in different cycles, and the temporal transmission correlation of critical path processes to the overall progress. During model training, time-series dependency To constrain this, the model's gating unit parameters are iteratively optimized to minimize the deviation between predicted and actual progress in the time dimension. This enhances the model's ability to capture time-series dependencies in long-cycle, multi-process scenarios, ensuring that the model accurately extracts the inherent correlation patterns of construction project progress over time. Furthermore, this method utilizes temporal correlation... It is associated with process-related features in the feature system, ensuring logical consistency with the entire modeling method.

[0036] Subsequently, a random forest was selected as the core ensemble learning model. Using the feature output values ​​adapted to the construction scenario as input data, a multi-factor correlation training sample set was constructed. Each sample in the set contains dynamic quantified values ​​of process correlation, resource constraints, and risk quantification-related features, along with the actual progress deviation value for the corresponding time period. The random forest model learns in parallel from the training sample set by constructing multiple decision trees. Each decision tree analyzes the local nonlinear relationship between a single factor and progress. Then, an ensemble voting mechanism is used to merge the outputs of all decision trees to fit the global nonlinear correlation under the synergistic effect of multiple factors. During model training, the difference between the nonlinear correlation impact value of multiple factors on progress and the actual progress deviation value is used as the optimization objective. The branch nodes and depth of the decision trees are iteratively adjusted to improve the fitting accuracy of complex nonlinear relationships among multiple factors.

[0037] In a preferred embodiment, the specific process of fitting multi-factor nonlinear correlations using an ensemble learning model is as follows: Random forest is selected as the core model for ensemble learning. The feature output values ​​adapted to the construction scenario are used as input data to construct a multi-factor association training sample set. Each sample in the set contains dynamic quantified values ​​of process association, resource constraints, and risk quantification-related features, along with the actual progress deviation value for the corresponding time period. The impact of the synergistic effect of multiple factors on the progress is quantified through a nonlinear association fitting function. The expression of the nonlinear association fitting function is as follows: , in, For the first The nonlinear correlation impact value of multiple factors on the progress during a time period is used to characterize the comprehensive effect of multiple factors on the progress of the current time period under synergistic effect; For the first The output values ​​of the class features after being adapted to the current construction scenario cover the class features of process association, resource constraints, and risk quantification. The data comes from the adaptation results of features and construction scenarios. For the first Class features in the first The scene correlation degree of a time period is defined as the ratio of the actual occurrence frequency of the influencing factors corresponding to this type of feature in the current time period to the average occurrence frequency of this type of factor throughout the entire construction cycle. The data comes from preprocessed multi-source correlation data. This refers to the total number of feature categories included in the feature system, namely, three types of features: process association, resource constraints, and risk quantification. For the first The frequency of multi-factor interactions during a time period is defined as the number of times that influencing factors corresponding to different categories of features occur simultaneously within the current time period. The data is obtained by statistically analyzing the occurrence records of various features in the feature system. Random forest models learn from the training sample set in parallel by constructing multiple decision trees, with each decision tree based on the nonlinear association influence value. The local nonlinear relationship between a single factor and progress is analyzed, and then the outputs of all decision trees are merged through an integrated voting mechanism to fit the global nonlinear correlation under the synergistic effect of multiple factors. During model training, the influence value of the nonlinear correlation is used. The difference between the actual and non-linear progress deviations is used as the optimization objective. The branch nodes and depth of the decision tree are iteratively adjusted to improve the fitting accuracy for complex non-linear relationships involving multiple factors. This process... Associated with feature adaptation results, through It is correlated with the preprocessed multi-source data to ensure logical consistency with the entire modeling method and accurately depict the nonlinear impact of multiple factors such as resource constraints, environmental interference, and process coupling on the progress of building construction projects.

[0038] Finally, hyperparameter optimization combined with cross-validation is used to determine the optimal model configuration. Specifically, a hyperparameter candidate space is first constructed, which covers the network structure parameters and training process parameters of the temporal prediction model, as well as the basic structure parameters and ensemble rule parameters of the ensemble learning model. The hyperparameters of the temporal prediction model include the number of hidden layer nodes and the training stride; the hyperparameters of the ensemble learning model include the number of base learners and the base learner decision boundary splitting threshold. The model training dataset is constructed based on the feature data adapted to the construction scenario. The training dataset is divided into several mutually exclusive subsets using a stratified sampling method. The number of subsets is determined based on the total number of construction stages of the building project, ensuring that each subset contains feature data and corresponding progress data for each construction stage. Each set of hyperparameter configurations in the hyperparameter candidate space is quantitatively evaluated using a hyperparameter configuration evaluation function, which is: , in, The comprehensive evaluation value for the hyperparameter configuration is used to characterize the predictive performance and adaptability of the model under the evaluation function expression configuration; This is the number of folds in the cross-validation, i.e., the number of subsets. For the first The actual progress values ​​corresponding to each subset of data are derived from preprocessed process execution data. For the first The predicted progress values ​​output by the model for each subset of the dataset under the current hyperparameter configuration; For the first The planned schedule values ​​corresponding to each subset of data are derived from project management data. For the first The scenario complexity coefficient of each subset of data is defined as the ratio of the number of construction stage types contained in the subset of data to the total number of construction stages in the entire project. It is used to correct the evaluation bias under different scenario complexities. Each subset of the dataset is used as the validation set, and the remaining subset is used as the training set. A hybrid machine learning model is trained based on the current hyperparameter configuration, and the comprehensive evaluation value of the configuration on the validation set is calculated through the hyperparameter configuration evaluation function. Traverse all configurations in the hyperparameter candidate space, repeat the cross-validation process, and record the comprehensive evaluation value corresponding to each configuration. Select the hyperparameter configuration with the largest comprehensive evaluation value as the optimal configuration of the model. Under this configuration, the model can achieve the optimal balance between prediction performance and adaptability in each construction stage scenario.

[0039] This process is achieved through , Associated with preprocessed multi-source correlation data, through By associating with the characteristics of the project construction phase, we ensure that hyperparameter optimization and cross-validation are closely aligned with the actual scenario of building construction projects, avoiding configuration mismatch issues caused by general optimization methods. At the same time, we achieve accurate measurement of hyperparameter configuration performance through quantitative evaluation functions, ensuring the prediction accuracy and stability of the hybrid model.

[0040] (4) Dynamically update risk warnings During the construction of building projects, factors such as process progress, resource supply, and external environment are constantly changing. Static training models are prone to a decrease in prediction accuracy due to changes in the scenario. At the same time, relying solely on prediction results cannot meet the need for advance management of schedule risks. Therefore, it is necessary to establish a dynamic model update mechanism and a corresponding risk early warning system.

[0041] Dynamic model updates are achieved through a combination of incremental learning and triggered optimization. Incremental learning is conducted based on preset time windows, the length of which is determined by the construction process cycle and data update frequency to ensure timely incorporation of the latest construction data. After each time window ends, only local parameters of the model are updated using the newly added construction data and corresponding progress results, avoiding the inefficiency caused by retraining all parameters and enabling the model to adapt to the gradual changes in the construction process in real time. Simultaneously, dual trigger conditions are set to initiate model optimization. When preset events occur, such as major design changes, batch failures of major construction equipment, or continuous extreme weather that may significantly affect progress, optimization is automatically triggered. Similarly, when the deviation between the model's predicted progress and the actual progress exceeds a set threshold, optimization is also initiated, adjusting model parameters through retraining to ensure that the model's prediction accuracy remains within a reasonable range.

[0042] A multi-level risk early warning system is constructed to achieve precise control of schedule risks. Multiple early warning thresholds are set by combining schedule deviation indicators and risk propagation probability. The deviation indicator is the ratio of the difference between actual and planned progress to the planned progress, while the risk propagation probability is the likelihood that a single risk factor will cause delays in related processes. The combination of these two indicators comprehensively reflects the severity and spread trend of schedule risks, and different early warning levels are established to differentiate the urgency of risks. When monitored data reaches the corresponding early warning threshold, the system triggers the appropriate level of early warning.

[0043] Meanwhile, by analyzing the contribution of each influencing factor, the core factors leading to schedule deviations or risks are identified, such as material supply delays and reduced efficiency of key processes. Based on practical management experience in building construction, specific corrective suggestions are given for different core factors, such as adjusting resource scheduling plans and optimizing process connections. This provides managers with precise decision support and enables a shift from passive response to proactive control.

[0044] Please refer to Figure 2 Based on the methods described in this embodiment, this embodiment 1 uses the SF1201-57 plot project of the Yantan Renovation Project in the Shibei District of a certain city as an application case. The project obtained a construction permit on July 25, 2025, and is scheduled to be completed on November 30, 2026. The progress dynamic prediction method of this embodiment is applied to its full-cycle management: First, multi-source data of 11 unit projects of the project are collected, including the actual time consumption of the work process, the connection interval, the material supply cycle, equipment failure rate and other resource data of 10 residential buildings, underground garages, etc. Environmental data such as daily precipitation and frequency of extreme weather, as well as management data such as design changes and supervisor rectification opinions, are regularized with "week" as the unified time granularity. A feature system of process association, resource constraints, and risk quantification is constructed. Based on the differences in stages such as "underground parking garage construction (August 2025 - December 2025), main construction of high-rise residential buildings (October 2025 - June 2026), and decoration of public facilities (July 2026 - October 2026)," the feature weights are dynamically adjusted to achieve feature-scenario adaptation.

[0045] Subsequently, a hybrid model combining LSTM and random forest was built: LSTM was used to capture the temporal dependencies between "underground engineering → main structure → decoration and accessories," and an attention mechanism was used to enhance feature extraction for key processes such as "construction of the 26-story residential building and excavation of the two-story underground garage." Random forest was used to fit the nonlinear relationships of multiple factors such as "material supply delay + low winter temperature" and "equipment failure + manpower shortage," and the optimal configuration was determined through hyperparameter optimization. During project implementation, the model was updated incrementally with a two-week time window. Automatic optimization was performed when "the main construction of Building #5 was delayed by more than 3 days" or the prediction error exceeded 5%. An early warning threshold was set by combining the deviation index of "parallel construction of resettlement housing and commercial housing" and the probability of risk propagation. When "delay in concrete supply to the underground garage" was detected, the cause was located and suggestions were given to "adjust the work rhythm of Building #3 and coordinate backup suppliers." Ultimately, this helped control the project schedule deviation within 3%, and the project was completed as scheduled on November 30, 2026, avoiding three potential delay risks.

[0046] The dynamic progress prediction method for building construction projects presented in this embodiment has broad application prospects in various types of building construction projects, including residential, commercial complexes, and public buildings. For construction companies, it can proactively avoid resource waste issues such as material backlog and idle manpower through accurate progress prediction and dynamic update mechanisms. Simultaneously, it leverages a risk warning system to promptly manage progress deviations, effectively reducing the risk of project delays and enhancing project cost control and quality assurance capabilities. For project construction and supervision units, the real-time progress data, risk warning information, and deviation correction suggestions provided by this method can strengthen the control over the entire project lifecycle, simplify progress verification and risk assessment processes, and provide reliable basis for key decisions such as project acceptance and payment. It is particularly suitable for the refined management needs of large and complex building construction projects.

[0047] At the industry application level, this method can drive the transformation of building construction project management from traditional experience-driven to data-intelligent driven. Through standardized data processing workflows, scenario-adaptive feature systems, and dynamically optimized model architecture, it can adapt to building construction project scenarios of different scales and regions, solving common problems in the industry such as low progress prediction accuracy, poor dynamic adaptability, and passive risk management. With the accumulation of application cases, the model can be continuously optimized through incremental learning, forming exclusive prediction capabilities adapted to different construction companies and different project types. This further expands its application scenarios in new building construction projects such as prefabricated buildings and green buildings, providing technical support for improving the overall efficiency and intelligence level of progress management in the industry.

[0048] Example 2 Please refer to Figure 3 This embodiment 2 provides a machine learning modeling system for dynamic prediction of the progress of building construction projects, including: The multi-source data acquisition and preprocessing unit is used to collect various types of related data that affect the progress of the building construction project throughout its entire lifecycle and remove invalid data. It uses a time-series unification method to organize the time dimension of the multi-source data and adopts standardization and dimension reduction methods to optimize data quality and reduce processing complexity. The building construction feature system adaptation unit is used to construct a feature system that includes process associations, resource constraints, and risk quantification. Based on the differences in scenarios at different construction stages of the project, the influence weight of each feature is adaptively optimized using a dynamic weight adjustment method to achieve real-time adaptation between features and construction scenarios. A hybrid model construction optimization unit is used to build a cascaded hybrid architecture of a time series prediction model and an ensemble learning model. The time series prediction model captures the time series dependencies of progress, and the ensemble learning model fits the nonlinear correlation of multiple factors. An attention mechanism is introduced into the time series prediction model to enhance the extraction of key factor features in the building construction scenario. Hyperparameter optimization combined with cross-validation is used to determine the optimal configuration of the model. The dynamic risk warning unit is used to update local parameters of the model using incremental learning based on a preset time window; when preset event conditions are met or the prediction error exceeds the threshold, the model optimization is automatically initiated; a multi-level risk warning system is constructed, and the warning threshold is set by combining deviation indicators and risk propagation probability, and the cause of deviation is located by the contribution of influencing factors and correction suggestions are given.

[0049] Example 3 This embodiment 3 also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can implement any step of a machine learning modeling method for dynamic prediction of the progress of a building construction project.

[0050] The computer-readable storage medium may include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0051] For a description of the computer-readable storage medium provided in this application, please refer to the above method embodiments; further details will not be repeated here.

[0052] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A machine learning modeling method for dynamic prediction of the progress of building construction projects, characterized in that, include: S1. Collect various types of related data that affect the progress of the building construction project throughout its entire lifecycle and remove invalid data. Regularize the time dimension of the multi-source data through a time series unification method, and optimize data quality and reduce processing complexity by using standardization and dimension reduction methods. S2. Construct a feature system that includes process associations, resource constraints, and risk quantification; based on the differences in scenarios at different construction stages of the project, adopt a dynamic weight adjustment method to adaptively optimize the influence weight of each feature, and complete the real-time adaptation of features to construction scenarios; S3. Construct a hybrid architecture that combines a time-series prediction model and an ensemble learning model. The time-series prediction model captures the time-series dependencies of the project, while the ensemble learning model fits the nonlinear relationships of multiple factors. An attention mechanism is introduced into the time-series prediction model to enhance the extraction of key feature extraction in the building construction scenario. The optimal model configuration is determined by combining hyperparameter optimization with cross-validation. S4. Based on a preset time window, incremental learning is used to update the local parameters of the model; when the preset event conditions are met or the prediction error exceeds the threshold, the model optimization is automatically started; a multi-level risk warning system is constructed, and the warning threshold is set by combining the deviation index and the risk propagation probability. The cause of the deviation is located by the contribution of the influencing factors and correction suggestions are given. The real-time adaptation process between the features and the construction scenario in S2 is as follows: The expression for its construction scenario adaptation function is: ， in, The feature output values ​​are adapted to the current construction scenario; The actual execution intensity of the core process in the current construction phase is defined as the ratio of the actual progress rate of the core process in the current phase to the standard progress rate. The resource consumption rate at the current construction stage is defined as the ratio of the actual consumption to the planned consumption of core resources per unit time at the current stage. The probability of risk occurrence in the current construction phase is defined as the ratio of the actual frequency of risk events occurring within the preset period of the current phase to the average frequency of risk events occurring in similar historical phases. This refers to the total number of construction stages included in the entire lifecycle of a building construction project. For the first The standard execution intensity of core procedures in each construction phase For the first Standard resource consumption rate for each construction phase For the first The standard risk probability for each construction phase is obtained based on statistical data from similar historical building construction projects. These are the basic feature values ​​in the feature system, namely the initial quantification values ​​of the process-related feature subset, the resource constraint feature subset, and the risk quantification feature subset; The specific method for capturing the time series dependencies of progress through the time series prediction model in S3 is as follows: Long Short-Term Memory (LSTM) network is selected as the core model for time series prediction. Based on multi-source associated data with regularized time dimensions, a progress time series sample set is constructed. With a preset time granularity as the step size, feature data within consecutive time periods are associated with the progress completion status of the corresponding time periods, forming a mapping relationship between input sequences and output labels. The intrinsic correlation between time series is quantified through a time dependency capture function, the expression of which is: ， in, For the first The progress time series dependency of a time period is used to characterize the strength of the correlation between historical time period data and the progress of the current time period; For the first The actual progress completed during the time period is derived from the pre-processed data related to the execution of the work process. For the first Time period and the The temporal correlation degree of the time period is defined as the ratio of the overlap of the core processes of two time periods to the time interval. The overlap of the core processes is determined based on the process correlation characteristics, and the time interval is the time span of the two time periods. The data comes from the process correlation characteristics in the feature system. The total number of historical time periods participating in the time-series dependency calculation is determined based on the cyclical characteristics of the construction project's work processes; For the first The actual progress completed during the time period The first output of the time series prediction model Forecasted progress for the specified time period; The hidden layers of the LSTM model calculate temporal dependencies on a time-by-time basis. The study analyzes the temporal correlation of process connections between adjacent time periods, the correlation of the progress patterns of similar processes in different cycles, and the temporal transmission correlation of critical path processes to the overall progress. During model training, time-series dependency To constrain this, the parameters of the model's gating units are iteratively optimized to minimize the deviation between the predicted and actual progress in the time dimension. This enhances the model's ability to capture time series dependencies in long-cycle, multi-process scenarios, ensuring that the model accurately extracts the inherent correlation patterns of the progress of building construction projects over time.

2. The machine learning modeling method for dynamic prediction of construction project progress according to claim 1, characterized in that, The multi-type associated data in S1 specifically includes: Process execution data: at least covering the actual time spent on each sub-item, the interval between process connections, the quality acceptance results of the completed process, and the progress of the critical path process; Resource allocation related data: at least covering material supply cycle, material inventory level, equipment failure rate, equipment maintenance cycle, number of personnel, distribution of personnel skill levels, and resource scheduling response time; External environmental data should include at least daily precipitation, average daily temperature, wind speed, geological disaster early warning information, lighting conditions at the construction site, and frequency of extreme weather events. Project management related data: at least including design change records, supervisor rectification opinions, safety accident records, progress inspection reports, contract period adjustment documents, and minutes of on-site coordination meetings.

3. The machine learning modeling method for dynamic prediction of building construction project progress according to claim 1, characterized in that, The dimensionality reduction method in S1 is as follows: Principal Component Analysis (PCA) or Linear Discriminant Analysis (LDA) is used to extract key influencing factors from multi-source data and remove redundant features, thereby reducing data dimensionality and the computational complexity of subsequent models while retaining core progress-related information.

4. The machine learning modeling method for dynamic prediction of construction project progress according to claim 1, characterized in that, The feature system in S2 is constructed based on the process characteristics of building construction projects, resource constraint rules, and risk transmission logic. The specific process is as follows: Based on the process characteristics of building construction projects, the connection coefficients, cross-influence factors and parallel operation efficiency of different sub-projects are extracted to form a subset of process-related features. Based on the resource constraint patterns of building construction projects, we quantify the impact weight of the construction period corresponding to the material inventory early warning threshold, the transmission coefficient of equipment failure repair time on process delays, and the correlation between the proportion of manpower shortage and construction efficiency, thus forming a subset of resource constraint features. Based on the risk transmission logic of building construction projects, a mapping relationship and quantitative indicators between risk factors, including weather, geology, and design changes, and construction progress are established to form a subset of risk quantitative characteristics. By integrating the aforementioned process-related feature subset, resource constraint feature subset, and risk quantification feature subset, a complete feature system for predicting the progress of building construction projects is formed.

5. The machine learning modeling method for dynamic prediction of construction project progress according to claim 1, characterized in that, The specific process of fitting multi-factor nonlinear associations using an ensemble learning model in S3 is as follows: Random forest is selected as the core model for ensemble learning. The feature output values ​​adapted to the construction scenario are used as input data to construct a multi-factor association training sample set. Each sample in the set contains dynamic quantified values ​​of process association, resource constraints, and risk quantification-related features, along with the actual progress deviation value for the corresponding time period. The impact of the synergistic effect of multiple factors on the progress is quantified through a nonlinear association fitting function. The expression of the nonlinear association fitting function is as follows: ， in, For the first The nonlinear correlation impact value of multiple factors on the progress during a time period is used to characterize the comprehensive effect of multiple factors on the progress of the current time period under synergistic effect; For the first The output values ​​of the class features after being adapted to the current construction scenario cover the class features of process association, resource constraints, and risk quantification. The data comes from the adaptation results of features and construction scenarios. For the first Class features in the first The scene relevance of a time period is defined as the first time period. The ratio of the actual occurrence frequency of the influencing factors corresponding to the class feature in the current period to the average occurrence frequency in the entire construction cycle is obtained from preprocessed multi-source correlation data. This refers to the total number of feature categories included in the feature system, namely, three types of features: process association, resource constraints, and risk quantification. For the first The frequency of multi-factor interactions during a time period is defined as the number of times that influencing factors corresponding to different categories of features occur simultaneously within the current time period. The data is obtained by statistically analyzing the occurrence records of various features in the feature system. Random forest models learn from the training sample set in parallel by constructing multiple decision trees, with each decision tree based on the nonlinear association influence value. The local nonlinear relationship between a single factor and progress is analyzed, and then the outputs of all decision trees are merged through an integrated voting mechanism to fit the global nonlinear correlation under the synergistic effect of multiple factors. During model training, the influence value of the nonlinear correlation is used. The difference between the actual value of the schedule deviation and the actual value is used as the optimization target. The branch nodes and depth of the decision tree are iteratively adjusted to improve the fitting accuracy of complex nonlinear relationships of multiple factors.

6. The machine learning modeling method for dynamic prediction of construction project progress according to claim 1, characterized in that, The specific process of determining the optimal model configuration using hyperparameter optimization combined with cross-validation in S3 is as follows: First, a hyperparameter candidate space is constructed, which includes the network structure parameters and training process parameters of the time series prediction model, as well as the basic structure parameters and ensemble rule parameters of the ensemble learning model. The hyperparameters of the time-series prediction model include the number of hidden layer nodes and the training step size; the hyperparameters of the ensemble learning model include the number of base learners and the base learner decision boundary splitting threshold. The model training dataset is constructed based on the feature data adapted to the construction scenario. The training dataset is divided into several mutually exclusive subsets using a stratified sampling method. The number of subsets is determined based on the total number of construction stages of the building project. Each set of hyperparameter configurations in the hyperparameter candidate space is quantitatively evaluated using a hyperparameter configuration evaluation function, which is: ， in, The comprehensive evaluation value for the hyperparameter configuration is used to characterize the predictive performance and adaptability of the model under the evaluation function expression configuration; This is the number of folds in the cross-validation, i.e., the number of subsets. For the first The actual progress values ​​corresponding to each subset of data are derived from preprocessed process execution data. For the first The predicted progress values ​​output by the model for each subset of the dataset under the current hyperparameter configuration; For the first The planned schedule values ​​corresponding to each subset of data are derived from project management data. For the first The scenario complexity coefficient of the subset is defined as the _th _th__. The ratio of the number of construction stage types in each subset of data to the total number of construction stages in the entire project is used to correct evaluation biases under different scenario complexities. Each subset of the dataset is used as the validation set, and the remaining subset is used as the training set. A hybrid machine learning model is trained based on the current hyperparameter configuration, and the comprehensive evaluation value of the current hyperparameter configuration on the validation set is calculated through the hyperparameter configuration evaluation function. Traverse all configurations in the hyperparameter candidate space, repeat the cross-validation process, and record the comprehensive evaluation value corresponding to each configuration. Select the hyperparameter configuration with the largest comprehensive evaluation value as the optimal configuration of the model. Under the optimal configuration of the model, the model can achieve the best balance between prediction performance and adaptability in each construction stage scenario.

7. A machine learning modeling system for dynamic prediction of construction project progress, characterized in that, A machine learning modeling method for implementing the dynamic prediction of construction project progress as described in claim 1 includes: The multi-source data acquisition and preprocessing unit is used to collect various types of related data that affect the progress of the building construction project throughout its entire lifecycle and remove invalid data. It uses a time-series unification method to organize the time dimension of the multi-source data and adopts standardization and dimension reduction methods to optimize data quality and reduce processing complexity. The building construction feature system adaptation unit is used to construct a feature system that includes process associations, resource constraints, and risk quantification. Based on the differences in scenarios at different construction stages of the project, the influence weight of each feature is adaptively optimized by dynamically adjusting the weights to achieve real-time adaptation between features and construction scenarios. A hybrid model construction optimization unit is used to build a cascaded hybrid architecture of a time series prediction model and an ensemble learning model. The time series prediction model captures the time series dependencies of progress, and the ensemble learning model fits the nonlinear correlation of multiple factors. An attention mechanism is introduced into the time series prediction model to enhance the extraction of key factor features in the building construction scenario. Hyperparameter optimization combined with cross-validation is used to determine the optimal configuration of the model. The dynamic risk warning unit is used to update local parameters of the model using incremental learning based on a preset time window; when preset event conditions are met or the prediction error exceeds the threshold, the model optimization is automatically initiated; a multi-level risk warning system is constructed, and the warning threshold is set by combining deviation indicators and risk propagation probability, and the cause of deviation is located by the contribution of influencing factors and correction suggestions are given.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is executed by a processor as described in any one of claims 1-6: a machine learning modeling method for dynamic prediction of the progress of a building construction project.

Images