Construction project cost prediction method based on artificial intelligence

Abstract

The invention relates to the technical field of business management data processing, in particular to a construction project cost prediction method based on artificial intelligence, which comprises the following steps: step 1, multi-source heterogeneous data collaborative acquisition: acquiring a component geometric attribute set and a historical project completion settlement data set in a building information model in real time, dynamically capturing price data of a building material market, and calculating the price of the building material market; when the design change is detected, the unique identifier set of the modified component is marked; 2, data preprocessing and feature fusion; step 3, performing dual-channel dynamic prediction; and 4, feedback optimization: triggering local re-calculation in real time according to market data update or design change, and adjusting feature extraction weight based on historical engineering data deviation. By introducing multi-source data collaborative acquisition, data preprocessing and feature fusion, dual-channel dynamic prediction and a feedback optimization mechanism, the accuracy, timeliness and flexibility of prediction are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of business management data processing technology, and in particular to an artificial intelligence-based method for predicting the cost of construction projects. Background Technology

[0002] In the field of construction project cost management, AI-based prediction methods have become a core direction for technological upgrading in the industry. However, existing technologies suffer from the following structural defects in practical applications: 1. Fragmented data from multiple sources leads to distorted predictions; Current mainstream forecasting systems (such as simple integration of BIM-based quantity surveying software with ERP systems) can acquire geometric data from building information models, historical project data, and market price data, but these three types of data are processed in separate channels: BIM geometric data is only used for automatically extracting basic parameters such as component volume and surface area; Material price data is stored in a static database, and the update cycle is usually more than 30 days. Engineering change data must be manually entered into the change notification form before calculation is triggered; This fragmented approach results in a lack of coupling for key features: for example, if an increase in the seismic resistance level of a region necessitates higher reinforcement requirements, the existing system cannot automatically transmit the seismic parameter adjustments from the design changes to the reinforcement calculation module, still relying on cost engineers to manually modify the calculation rules. According to industry survey reports, cost deviations caused by this type of data fragmentation average 12.7%.

[0003] 2. The market price response mechanism has failed; Fluctuations in building material futures prices have a significant impact on construction costs (for example, a daily fluctuation of more than 5% in rebar futures will affect the cost of the main structure by more than 3%), but existing technologies suffer from a double lag: Data acquisition lag: More than 80% of construction companies use monthly centralized procurement prices as a benchmark and cannot access the real-time data stream of futures exchanges; Model update lag: Adjusting price parameters requires retraining the entire prediction model, which takes an average of more than 48 hours; This led to the fact that in the 2022 cement price surge, the final settlement price of most projects exceeded the bid price by 15%-25%, mainly because the prediction model did not respond to price fluctuations in real time.

[0004] Therefore, there is an urgent need for an artificial intelligence-based method for predicting the cost of construction projects to solve the above problems. Summary of the Invention

[0005] To achieve the above objectives, this invention provides an artificial intelligence-based method for predicting the cost of construction projects, comprising the following steps: Step 1: Collaborative acquisition of multi-source heterogeneous data: Real-time acquisition of the set of geometric attributes of components in the building information model and the dataset of historical project final accounts, dynamic capture of building material market price data, and marking the set of unique identifiers of modified components when design changes are detected; Step 2: Data Preprocessing and Feature Fusion: The building information model components are merged into priced units according to construction technology, and a set of constraint rules is generated in combination with regional construction standards. A dynamic price weight matrix is ​​generated based on market price fluctuations. When design changes occur, the scope of affected units is located through the component dependency graph and the quantity of work is recalculated. Step 3: Dual-channel dynamic prediction: Construct the analytical geometric features and constraint rules of the static processing channel, construct the analytical price time series features of the dynamic processing channel, and fuse and output the spatiotemporal feature tensor through the cross-channel attention mechanism. Based on the spatiotemporal fusion feature tensor regression, output the comprehensive unit price of the unit project, the coefficient of the measure item fee and the proportion of the risk reserve fund. Step 4: Feedback Optimization: Local recalculation is triggered in real time based on market data updates or design changes, and feature extraction weights are adjusted based on deviations in historical engineering data.

[0006] Preferably, the markers for design changes in step 1 include: Obtain the set of spatial positioning coordinates of the components before and after the change, and calculate the components whose coordinate changes exceed the displacement threshold; Analyze the topological connection relationship of components: extract the force transmission path matrix at the beam-column joint. When the rate of change of the matrix eigenvalues ​​exceeds the structural stability threshold, add the associated column components, beam components and connecting plate components to the identifier set. The displacement threshold is determined by the 90th percentile of the statistical distribution of allowable deformation of components in historical engineering projects, and the structural stability threshold is converted into the critical point of the rate of change of matrix eigenvalues ​​based on the allowable value of structural deformation angle in the local seismic code.

[0007] Preferably, the generation of the quantifiable unit in step 2 includes: Identify rebar binding units: Calculate the rebar distribution density gradient in the beam-column joint area. When the gradient value exceeds the critical value of the binding process, merge the intersecting beam and column members into a single unit. The critical value of the binding process is determined according to the rebar diameter category; the larger the diameter, the lower the critical value. Split formwork support units: Divide the wall length by the local standard formwork size to obtain the theoretical number of units to be split. Then adjust the splitting points according to the location of door and window openings to ensure that each unit contains a complete construction joint.

[0008] Preferably, the transformation of the constraint rule set in step 2 includes: Dynamic correction of stirrup spacing: Obtain the seismic acceleration value and building importance coefficient of the project site, calculate the spacing reduction ratio caused by the increase of seismic resistance level. The reduction ratio is positively correlated with the seismic acceleration value and negatively correlated with the yield strength value of the steel bars. Rebar length compensation: The anchorage length is increased based on the concrete cover thickness specification. The compensation amount is calculated using a piecewise function according to the rebar diameter category. When the diameter is greater than the critical value, the compensation amount increases according to a quadratic curve.

[0009] Preferably, the construction of the component dependency graph in step 2 includes: Using the modified component as the root node, search for three-level associated components along the force transmission path: first-level association consists of directly connected components, second-level association consists of components that indirectly transmit loads through connectors, and third-level association consists of components located in the same force system. When related components are of the same material type, expand the search to the fifth level of related components; The mechanical transmission path is determined by extracting the overlapping area of ​​the stress cloud diagram of the component through finite element analysis.

[0010] Preferably, the cross-channel attention mechanism in step 3 includes: Dynamic weighted similarity: Calculate the cosine similarity between the static channel geometric feature vector and the dynamic channel price feature vector. When the similarity is lower than the correlation threshold, reduce the weight of the price feature. Construction phase adaptation: During the foundation construction phase, the static channel weight is increased according to the proportion of the foundation pit excavation depth; during the decoration phase, the dynamic channel weight is increased according to the fluctuation range of material prices. The correlation threshold is dynamically adjusted based on the statistical quantile of the deviation between material prices and project quantities in historical projects.

[0011] Preferably, the adjustment of feature extraction weights in step 4 includes: Deviation component identification: When the predicted value of a specific type of component deviates from the final settlement value by more than a threshold, the usage distribution of the same type of component in historical projects is retrieved; Weighting adjustment: The feature extraction coefficients are adjusted based on the 75th percentile of the usage distribution. If the actual value is consistently higher than the predicted value, the weight of the geometric feature is increased. If the actual value is dominated by price fluctuations, the weight of the dynamic channel is increased.

[0012] The feature extraction weights are the adjustment coefficients for geometric features and dynamic channels in the deviation feedback, adjusted according to the 75th percentile of the historical usage of similar components. Preferably, the LSTM timing network operation of the dynamic processing channel includes: Input sequence optimization: Divide the dynamic price weight matrix into weekly granularities and calculate the price fluctuation variance of adjacent subsequences; Window adaptation: When the variance exceeds the market volatility threshold, the input sequence time window is shortened to a fixed proportion of the original window length; The market fluctuation threshold is determined based on the median of the difference between the peak and trough of the material's historical price.

[0013] Preferably, the process of generating the risk reserve ratio in step 4 includes: Risk Feature Extraction: Separate a subset of material price dimension features from the spatiotemporal fusion feature tensor; The difference between the maximum and minimum values ​​of this subset within the forecast period is calculated as the price range. Dynamic risk level classification: The price range is matched with the threshold range in the historical engineering risk database, which stores abnormal price fluctuations of various building materials in the past five years and their corresponding actual loss rates. When the price range continues to exceed the first risk threshold for a preset monitoring period, it is marked as a medium risk level. When the price range exceeds the second risk threshold and the volatility growth rate accelerates in adjacent monitoring periods, it is marked as a high-risk level. Tiered adjustment of reserve fund base: Under the medium risk level: the calculation base is increased by multiplying the price range by the basic reserve rate; At the high-risk level: an additional base is added based on the duration of the volatility and the maximum daily volatility. in: The first risk threshold is determined based on the 90th percentile of the historical price fluctuation statistics of the material; The second risk threshold adopts the red alert benchmark value issued by the industry price warning agency; The volatility growth rate is calculated using the absolute value of the first derivative of the price series.

[0014] Preferably, the three-layer fully connected network of the static processing channel includes: Activation function adaptation: The first layer uses the ReLU function to eliminate negative value interference for geometric features, the second layer uses the LeakyReLU function to preserve gradient characteristics for constraint rule features, and the output layer uses the Sigmoid function for normalization; Node number setting: The number of hidden layer nodes increases logarithmically based on the number of types of priced units, and the number of output layer nodes is equal to the number of cost index categories.

[0015] The beneficial effects of this invention are: 1. This invention effectively integrates BIM geometric data, historical engineering data, and market price data by introducing a multi-source data integration and fusion mechanism, and establishes a real-time data processing framework. This technology enables automated coupling and correlation of different data sources within the same system environment, eliminating errors caused by data fragmentation.

[0016] 2. This invention solves the problem of data acquisition lag by introducing a real-time market data stream processing mechanism. The system can access the real-time data stream of the futures exchange and reflect market price fluctuations in the cost prediction model in a timely manner. Simultaneously, this invention also proposes an adaptive model update mechanism that can automatically adjust model parameters when market prices fluctuate, reducing the time required for model training and significantly shortening the model update cycle. This mechanism ensures real-time response to price changes during periods of sharp market price fluctuations, thereby avoiding cost prediction errors caused by model lag.

[0017] 3. By efficiently accessing real-time data streams and dynamically updating the model, this invention enables the cost forecasting model to respond promptly to market changes, significantly improving the timeliness and accuracy of forecasts. When dealing with sudden market fluctuations, the system can adjust quickly and update costs based on the latest data, which is crucial for cost control in construction projects. The system's automated processing reduces manual intervention and the risk of human error, thereby further improving the accuracy of cost forecasting.

[0018] 4. The invention utilizes an automated data transmission mechanism to directly impact the calculation models of design changes on aspects such as steel reinforcement usage and material procurement in real time, eliminating cost discrepancies caused by delayed information transmission. This automated processing not only improves calculation efficiency but also ensures the accuracy of the calculation results.

[0019] 5. Through deep learning and data mining technologies, this invention can continuously learn and optimize the prediction model, maintaining efficient and accurate prediction capabilities in the face of complex and ever-changing market and engineering environments. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the steps of the method of the present invention; Figure 2 This is a flowchart illustrating the steps of converting the constraint rule set in step 2 of the method of the present invention. Figure 3 This is a flowchart of the cross-channel attention mechanism in step 3 of the method of the present invention. Detailed Implementation

[0022] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0023] Please see Figures 1-3 This invention provides an artificial intelligence-based method for predicting the cost of construction projects. In step 1, the system first acquires the geometric attribute data of components (such as volume and surface area) in real time through an interface with a Building Information Modeling (BIM) system. Simultaneously, the system automatically retrieves cost information from past projects from historical project completion settlement datasets, ensuring the prediction model can reference historical data. To maintain the model's timeliness, this invention also includes a module that dynamically captures market price data (such as real-time prices of building materials like steel and concrete). Furthermore, when a design change is detected, the system analyzes the design data in the BIM, automatically marks the modified components, and assigns them unique identifiers, ensuring accurate tracking of affected components in subsequent steps.

[0024] This step effectively integrates multiple data sources, eliminating the data fragmentation problem inherent in traditional methods. Real-time collection and dynamic updating of market price data ensures the accuracy and timeliness of price data during project forecasting.

[0025] In step 2, the system merges components in the Building Information Model (BIM) according to construction process to form quantifiable units. Each unit represents a specific building element and possesses quantifiable attributes. Next, based on local construction standards, a corresponding set of constraint rules is generated for each unit to ensure that the construction plan complies with local standards and regulations. Simultaneously, the system generates a dynamic price weight matrix based on market price fluctuations, reflecting the price fluctuations of different materials over different time periods. In the event of design changes, the system uses a component dependency graph to automatically locate the affected component units and recalculate the quantities, ensuring that budget changes are reflected in the updated cost forecast.

[0026] This step enables the characteristics of different data sources to be standardized and integrated for processing, ensuring the integrity and accuracy of information. In particular, when faced with design changes, it can reflect the impact of changes on the workload and cost in a timely and automated manner.

[0027] Step 3 involves constructing static and dynamic processing channels to analyze the geometric features, constraint rules, and temporal characteristics of market prices for components. In the static channel, the system analyzes the geometric features and constraint rules in the BIM model to generate relevant engineering data. In the dynamic channel, the system processes the time series of market price data to capture the impact of price fluctuations on cost. Next, a cross-channel attention mechanism is introduced, which fuses static and dynamic features to generate a spatiotemporal feature tensor. This tensor comprehensively reflects the various requirements of the project. Based on this, the system outputs the comprehensive unit price of the project, the coefficient of the measure item cost, and the proportion of the risk contingency fund through a regression model, thereby achieving more accurate cost prediction.

[0028] The aforementioned measures project cost coefficient is the proportion of temporary facilities, safe and civilized construction costs, etc., to direct project costs; the aforementioned risk contingency fund ratio is the proportion of the reserve fund for price fluctuations to the total cost, accumulated according to risk level: Medium risk: Base rate × price range ratio (1% × (280 / 1000) = 0.28%) High risk: Additional volatility duration × maximum daily volatility (+10 days × 50 yuan / day = 500 yuan base). By integrating static and dynamic information, the system can comprehensively consider the multi-dimensional characteristics of a project, which not only improves the accuracy of predictions but also ensures that factors such as market fluctuations and design changes can have an immediate impact on cost predictions, thus better adapting to complex project environments.

[0029] It should be noted that the core task of the static processing channel is to transform the geometric data of the BIM model and the construction specification text into a set of digital features that computers can directly understand and use for calculations. Its workflow is as follows: First, the detailed information of the BIM model is automatically read. This includes not only obtaining the size and volume of a component, but also its specific model, its location in the structure, and in-depth information such as how it is connected to other components.

[0030] Next, the built-in "standard knowledge base" is invoked. This knowledge base stores various calculation rules pre-set by experts, such as "if the seismic resistance level is level one, then the stirrups of the beam need to be densified to a spacing of 100 mm," and "if the diameter of the steel bar is greater than 25 mm, then its anchorage length needs to be increased," and so on. The system will, like conducting a questionnaire, take the information read from the BIM model (what kind of component is it? What environment is it used in?) and match it with the corresponding rules in the knowledge base.

[0031] Then, "calculation and fusion" are performed. Based on the matched rules, the system automatically calculates the "compliant" construction requirements and technical parameters for this component. Finally, it packages the component's original geometric data (such as volume) and these newly calculated rule data (such as compliant steel reinforcement usage) together to form a richer and more accurate set of "fusion features." Finally, it should be added that the technical means and analysis process for implementing the above workflow can be selected and adjusted by those skilled in the art. There are corresponding modeling software and analysis tools on the market. This invention does not improve the underlying logic of these software, so it will not be elaborated here.

[0032] In step 4, the system performs timely local recalculation based on real-time market data updates or events triggered by design changes. Through historical engineering data deviation analysis, the system can automatically adjust the weights in the feature extraction process, enabling the model to self-optimize when faced with new data. This mechanism ensures the model's adaptability and continuous improvement, and allows for timely adjustments to respond to market changes and design modifications.

[0033] The feedback optimization mechanism enables the system to learn and adapt on its own, continuously optimizing the prediction model as the market environment changes and project requirements adjust, significantly improving the accuracy and stability of predictions.

[0034] By introducing multi-source data collaborative acquisition, data preprocessing and feature fusion, dual-channel dynamic prediction, and feedback optimization mechanisms, this method successfully solves the problems of data fragmentation and lagging market price response in current construction cost forecasting, significantly improving the accuracy, timeliness, and flexibility of forecasting. This method can efficiently handle design changes and market fluctuations, ensuring the accuracy and timely response of cost forecasts in practical applications, and providing strong technical support for cost control and management of construction projects.

[0035] In one possible implementation, design changes during the construction of a building project often lead to changes in the spatial position of components. A BIM (Building Information Modeling) system can be used to acquire the spatial coordinate data of each component before and after the design change. This coordinate data can include the component's position in three-dimensional space, defined as a set of coordinates for each component. The system then calculates the displacement of each component before and after the change, i.e., it calculates the difference between the old and new coordinates of each component. If this change exceeds a preset displacement threshold, the component is considered to have been affected by the design change and is marked as an object requiring attention.

[0036] The displacement threshold is set based on statistical analysis of historical engineering data. In historical projects, the distribution of allowable deformation of components is statistically analyzed, and the 90th percentile of the deformation is calculated. This value serves as a reference for the displacement threshold. That is, when the displacement of a component exceeds this threshold, the system considers the component to have undergone a significant change and requires attention. This method ensures the system's sensitivity when handling design changes and prevents the omission of critical changes.

[0037] In architectural design, structural components are typically connected into a unified whole through structural elements such as beams, columns, and connecting plates. The force transmission relationships at beam-column joints are crucial for the stability of a building structure. When the design changes, the force transmission paths and the connection relationships between components will change. The system analyzes the topology of the components and extracts the force transmission path matrix of the beam-column joints. This matrix describes the distribution and transmission of forces among different components. If the rate of change of the matrix's eigenvalues ​​exceeds a set structural stability threshold, it indicates that the design change may lead to a significant change in the mechanical performance of the structure. The system will automatically add the relevant column, beam, and connecting plate components to the set of change identifiers.

[0038] The structural stability threshold is set based on local seismic design codes. The seismic design of buildings specifies the maximum permissible deformation angle of the structure. The system converts this standard into a critical point for the rate of change of matrix eigenvalues, ensuring that when the change in the force transmission matrix eigenvalues ​​exceeds this critical point, the system flags potential structural problems. Through this method, the system can identify potential structural safety hazards arising from design changes in advance and provide dynamic early warnings.

[0039] By combining spatial displacement with topological analysis, components affected by design changes are accurately identified. Compared to traditional manual inspection methods, this automated marking method not only improves efficiency but also precisely locates each affected component, ensuring the accuracy of predictions and cost calculations.

[0040] By analyzing changes in the force transmission path matrix and its eigenvalues, structural stability issues that may arise from design changes can be identified in a timely manner. Particularly in seismic design, by incorporating local seismic codes, the system can perform refined risk assessments based on specific regional regulations, preventing structural safety hazards caused by design changes.

[0041] Accurately marking design changes enables more precise budget adjustments and recalculations of quantities, reducing budget deviations caused by design changes. This is crucial for project management and cost control, effectively preventing economic losses due to budget forecasting errors.

[0042] By refining the marking process for design changes and utilizing spatial displacement and structural stability analysis, the impact of design changes on the cost of building projects can be captured comprehensively and accurately. This not only optimizes the cost forecasting process but also ensures the structural safety and budget accuracy of the project.

[0043] In one possible implementation, rebar tying is a crucial process in building structures, particularly in beam-column joint areas. To accurately generate priced units, the rebar distribution density needs to be analyzed. First, the system acquires rebar distribution data from the beam-column joint areas of the building model and calculates the rebar density gradient. This gradient value represents the concentration of rebar within a specific area, reflecting its distribution. When the gradient value exceeds a preset critical value for the tying process, the system automatically merges intersecting beam and column members into a single priced unit.

[0044] The critical value for the binding process is set based on the diameter category of the reinforcing bars. Generally, larger diameter reinforcing bars involve more complex binding processes and require more manual operation. Therefore, the critical value set by the system for larger diameter reinforcing bars is relatively lower. In other words, the larger the diameter of the reinforcing bar, the higher the difficulty and complexity of the binding process, making it easier for the density gradient of the reinforcing bars to exceed the critical value, thereby triggering the rule of merging beam-column members into a single unit.

[0045] By accurately identifying rebar tying units, cost predictions can be effectively refined, especially the rebar tying costs at beam-column joints. This avoids the inaccurate cost estimations of complex components found in traditional methods, providing a more precise basis for pricing.

[0046] Formwork support is a fundamental step in building construction, and its cost calculation also requires precise processing. To this end, the system calculates the wall length and divides it by the standard size of commonly used formwork in the local area to determine the theoretical number of sections to be divided. The purpose of this process is to initially determine the number of units after division based on the standard formwork dimensions. Next, the system further adjusts according to the location of door and window openings in the wall to ensure that each divided unit contains a complete construction joint.

[0047] Door and window openings are a key factor affecting the dismantling of formwork supports, as they alter the installation and support methods of the formwork. The system analyzes the location of these openings and automatically adjusts the dismantling points to ensure that the formwork support units, after dismantling, adequately cover all construction joints, thereby reducing potential uneven formwork support or structural instability during construction.

[0048] Splitting the formwork support units allows for precise calculation of the actual number of formwork units used and installation costs, avoiding additional costs caused by unreasonable splitting or omission of construction joints. Furthermore, adjustments based on door and window openings improve the adaptability of the formwork support scheme, preventing budget deviations caused by neglecting details in traditional methods.

[0049] By precisely identifying and separating rebar tying units and formwork support units, this invention provides a more detailed and accurate basis for cost prediction in construction projects. The separation and merging of rebar and formwork support, combined with multi-dimensional analysis of complex construction processes, significantly improves the accuracy of cost prediction and enables better control of material costs and construction schedules in project management. Ultimately, this method not only optimizes budget management but also improves construction efficiency and reduces waste caused by unreasonable design or inaccurate processes.

[0050] In one possible implementation, stirrups are a crucial element in building structures, providing shear resistance and seismic performance. Based on the seismic acceleration values ​​at the project site and the building's importance coefficient, the system calculates the impact of increasing seismic resistance level on stirrup spacing. When the seismic resistance level increases, design requirements typically lead to a reduction in stirrup spacing to enhance the structure's seismic capacity. The system calculates the percentage reduction in stirrup spacing by collecting and inputting seismic acceleration values ​​(representing the intensity of seismic activity) and the building's importance coefficient (reflecting the building's significance to social functions).

[0051] A higher seismic acceleration indicates more frequent and intense seismic activity, thus requiring greater reinforcement. The stirrup spacing needs to be tighter to improve seismic performance. When the steel reinforcement has higher strength, it can withstand greater seismic loads, therefore the stirrup spacing requirement is relatively lower. The yield strength of the steel reinforcement is negatively correlated with the spacing reduction ratio.

[0052] The reduction ratio is the percentage reduction in stirrup spacing caused by the increase in seismic resistance level. It is positively correlated with seismic acceleration and negatively correlated with the yield strength of steel bars. By dynamically adjusting the stirrup spacing, the system can adjust the design in real time according to different seismic conditions and building requirements to ensure that the building's seismic performance meets standards and avoids non-compliance with specifications during actual construction. At the same time, this dynamic adjustment method can improve the accuracy of cost prediction and avoid additional costs caused by unreasonable design.

[0053] The thickness of the concrete cover for reinforcing bars is a key factor affecting their anchorage effectiveness and tensile strength. Based on the specified thickness of the concrete cover, the system increases the anchorage length of the reinforcing bars to ensure that the anchorage length meets design requirements. To compensate for potential differences in anchorage length due to different bar diameters, the system uses a piecewise function to calculate the compensation. For larger diameter reinforcing bars, the compensation increases according to a quadratic curve, which more reasonably adapts to the actual anchorage needs of reinforcing bars of different diameters.

[0054] The compensation amount is the increased anchorage length of the reinforcing bars (i.e. the length of the reinforcing bars embedded in the concrete) to meet the specifications for the thickness of the concrete cover. Its calculation is based on a piecewise function of the reinforcing bar diameter: it increases linearly when the diameter is small, and increases quadratically when the diameter exceeds the critical value. Concrete cover thickness: refers to the minimum distance from the surface of the reinforcing steel to the surface of the concrete. For larger diameter reinforcing bars, the compensation amount is usually larger because thicker bars require a longer anchorage section to provide sufficient tensile strength. The quadratic growth curve of the compensation amount can more accurately simulate this requirement.

[0055] The compensation amount is the increased anchorage length of the reinforcing bars required due to the thickness of the concrete cover. It is calculated as a piecewise function based on the diameter of the reinforcing bars, and grows on a quadratic curve when the diameter is greater than the critical value. By compensating for rebar length, the system can precisely adjust the anchorage length of the rebar, ensuring a firm connection of the rebar in the concrete and improving the stability and safety of the structure. Furthermore, this compensation method also ensures reasonable cost prediction, avoiding material waste or improper construction caused by inaccurate rebar length calculations.

[0056] By dynamically correcting the stirrup spacing and compensating for the rebar length, this invention provides more precise cost predictions, particularly regarding the seismic resistance of the structure and the anchorage effect of the rebar. These improvements not only enhance building safety and meet higher design and construction standards but also optimize project cost predictions, making construction project budgets more accurate. Furthermore, the system dynamically adjusts rules based on different seismic conditions, building types, and material specifications, effectively improving the operability and execution efficiency of the project. The introduction of this method reduces additional costs incurred during construction due to inapplicable specifications or design errors, thereby improving the overall management level of construction projects.

[0057] In one possible implementation, the "modified component" is first identified as the root node. The modified component typically refers to a component that has undergone changes during the design or construction process. For example, changes in materials, dimensions, or other construction requirements may cause this component to become a critical factor affecting the entire structure.

[0058] Based on the force transfer path, the system searches for components related to the root node level by level along the force transfer path through stress analysis of the building structure. The force transfer path reflects how building components affect the structure through their connections and load transfer.

[0059] First-order connected members are directly connected to the root node and bear the load directly through physical connections (such as beams and columns). Second-order connected members bear the load indirectly through connectors (such as reinforcing bars and welded parts). Their relationship with the root node is more indirect, but they still play an important role in transmitting mechanical influences. Third-order connected members are located within the same force system. Although their direct connection with the root node is farther away, they are still affected by the load of the entire structure.

[0060] When the associated components involved are of the same material type, the system will further extend the search to components at level five. Even if these components are farther from the root node, components of the same material may still influence each other under mechanical forces. Therefore, the component dependency graph will include more extensive structural information to ensure completeness.

[0061] In this embodiment of the invention, finite element analysis (FEM) is used to perform a detailed mechanical analysis of the structure. FEM divides the building structure into multiple finite element elements, and numerical calculations are used to solve the behavior of these elements under stress, thereby obtaining the stress distribution of the entire structure.

[0062] Through finite element analysis, the system generates stress contour maps of the components, representing the stress distribution of different parts under different loads. Overlapping areas in the stress contour maps indicate the interactions and influences between different components. Based on these overlapping areas, the system determines which components are within the same load-bearing system, which helps to accurately capture the dependencies between components and incorporate these dependencies into the component dependency map.

[0063] By accurately constructing component dependency diagrams, the mechanical relationships between each component and its adjacent components in a building project can be analyzed in depth, ensuring the accuracy and safety of cost predictions. Combining finite element analysis and the analysis of overlapping areas in stress cloud diagrams can help building projects better understand and predict the interactions between components, thus providing a more scientific basis for construction and improving the quality and efficiency of building projects.

[0064] In one possible implementation, in cost prediction for a construction project, the static channel includes geometric features related to the project structure, such as building dimensions and material specifications, while the dynamic channel involves features related to market prices, such as material price fluctuations and construction costs. To measure the relationship between the static and dynamic channels, the cosine similarity of the feature vectors of these two channels is first calculated. Cosine similarity is a method for measuring the similarity of two vectors in direction; the closer the value is to 1, the closer the feature relationship between the two channels.

[0065] When the similarity between the geometric features of the static channel and the price features of the dynamic channel is lower than a preset correlation threshold, the system will reduce the weight of the price features in the dynamic channel. This means that if the geometric features and price features are not correlated (for example, building structure and material price fluctuations are not directly related), the influence of price features on the prediction model will be reduced to avoid misleading the final prediction results.

[0066] Dynamic similarity weighting adaptively adjusts the weights between static and dynamic channels, allowing the model to dynamically adjust its calculations based on the correlation between features in each channel, thereby improving the accuracy of prediction results. If the correlation between certain features is weak, the model will automatically reduce their impact on the final prediction, avoiding interference from irrelevant factors.

[0067] During the foundation construction phase, the project involves excavation of the foundation pit and foundation treatment, which have a significant impact on the building structure. The weight of static access routes increases based on the proportion of the excavation depth. Excavation depth directly affects the strength and stability of civil engineering structures; therefore, the geometric characteristics of static access routes (such as the depth and dimensions of the structure) have a significant impact on the predicted cost at this stage.

[0068] During the decoration phase, the focus of construction shifts to material procurement and construction techniques. At this stage, the impact of material price fluctuations on project costs becomes more significant. Therefore, the weight of price characteristics in the dynamic channel will increase during the decoration phase, especially when material prices fluctuate sharply. In this situation, the contribution of price characteristics in the dynamic channel to cost prediction becomes more important.

[0069] Construction phase adaptation enables the model to adaptively adjust weights based on the characteristics of different construction phases, ensuring that the most relevant features are used for prediction at each stage. This allows for a more accurate reflection of cost fluctuations at each stage of construction, improving the flexibility and accuracy of the entire cost prediction process.

[0070] The relevance threshold is not statically set but dynamically adjusted based on historical project data. By statistically analyzing the deviations between material prices and project quantities in historical projects, the system can calculate the statistical quantiles of these deviations and adjust the similarity calculation threshold accordingly. When historical data indicates that the price fluctuations or deviations of a certain type of material are large, the relevance threshold will be increased to reduce the weight of irrelevant features.

[0071] Dynamically adjusting the correlation threshold allows the model to adapt more flexibly to fluctuations in material prices, thereby improving its applicability to different types of projects and market environments. This adaptive adjustment mechanism effectively reduces the discrepancy between historical and current data, ensuring the model's accuracy under various conditions.

[0072] By introducing a cross-channel attention mechanism, particularly dynamic weighting of similarity, adaptation to construction stages, and dynamic adjustment of relevance thresholds, the flexibility and accuracy of construction project cost forecasting can be improved. The feature weights for different stages and components can be dynamically adjusted reasonably according to actual conditions, making the entire cost forecasting process more aligned with the actual construction environment and market conditions. Ultimately, this will help project managers make more accurate decisions in budget control, cost optimization, and construction planning.

[0073] In one possible implementation, during the cost forecasting process for a construction project, the predicted values ​​for certain components may deviate significantly from the actual final settlement values. To ensure the accuracy of the forecast, the system first sets a deviation threshold to identify which components' predicted results deviate from the final settlement values ​​beyond this threshold.

[0074] Once a component with a significant deviation is identified, the system will further retrieve the usage distribution of similar components from historical engineering data. By analyzing similar components in historical projects, the normal usage range and fluctuation patterns of this type of component can be identified.

[0075] This identification mechanism can accurately locate components with large prediction errors, preventing the model from relying too heavily on inaccurate prediction data and thus improving the overall prediction quality. By retrieving historical data, more accurate reference data can be provided for these components with large deviations, preventing error accumulation from affecting the overall prediction results.

[0076] After a component with a deviation is identified, the system will adjust its weights based on the usage distribution of similar components in historical projects. Specifically, the system will calculate the 75th percentile of the usage distribution for that type of component. This means that in all historical project usage data, 75% of the data values ​​are below this percentile, and the remaining 25% are above it. This value represents the upper limit of the normal usage for that component and can serve as a reference for adjusting the prediction weights.

[0077] Based on historical data analysis, if the actual value (final settlement value) consistently exceeds the predicted value, the system will increase the weight of geometric features. This means that in such cases, the geometric features of the building (such as size and shape) have a significant impact on the project cost, and therefore the influence of geometric features should be strengthened.

[0078] If the actual value is primarily driven by price fluctuations, the system will increase the weight of the dynamic channel. In this case, material price fluctuations become the main factor in cost deviations, thus the system will amplify the impact of price fluctuations on the forecast.

[0079] Weight correction dynamically adjusts feature extraction coefficients, enabling the model to flexibly respond to different sources of error based on actual conditions. When the predicted values ​​of biased components deviate significantly, the system can precisely adjust the prediction model by adjusting the weights of geometric features and dynamic channels to reflect the impact of different component characteristics on the construction project cost. This mechanism effectively improves the model's adaptability to special cases and its prediction accuracy.

[0080] Through this adaptive weight adjustment, the system can better simulate complex construction conditions in reality, especially when material prices fluctuate greatly. It can highlight the impact of price factors and avoid underestimating the cost changes caused by price fluctuations.

[0081] By identifying and analyzing the usage distribution of defective components, and combining this with a weighting correction mechanism, the model can dynamically adjust itself when faced with errors in specific components. This method not only improves the accuracy of cost forecasting but also allows the model to maintain high flexibility across different market environments and construction stages. Ultimately, through adaptive adjustment, the cost forecasting for construction projects becomes more accurate and stable, effectively addressing errors in historical data and price fluctuations during actual construction.

[0082] In one possible implementation, the input sequence needs to be optimized to ensure that the LSTM model can effectively capture patterns in price fluctuations. First, the dynamic price weight matrix is ​​divided into weekly granularities, meaning the model will subdivide the price data into multiple subsequences on a weekly basis.

[0083] The dynamic price weight matrix is ​​a weight allocation matrix for building material price fluctuations (behavior is material, column is time), and the normalized variance of weekly price fluctuations is the result. Next, the system calculates the price volatility variance between adjacent subsequences. Volatility variance reflects the degree of price fluctuation over a specific time period and helps the model determine market price volatility.

[0084] By segmenting price fluctuations into weekly granularities and calculating the variance of volatility, the system can more accurately capture price fluctuation trends and changes, avoiding the inability to reflect true market volatility due to overly stable price fluctuations. This helps improve the model's sensitivity to price fluctuations, thereby increasing the accuracy of construction project cost predictions.

[0085] When the calculated price volatility variance exceeds a preset market volatility threshold, the system will adaptively adjust the time window of the input sequence to the LSTM network. Specifically, the system will shorten the time window of the input sequence according to a set ratio. This means that when market prices fluctuate significantly, the model will focus more on short-term price changes rather than long-term trends.

[0086] The market volatility threshold is the limit for abnormal fluctuations in the variance of the price series. If this limit is exceeded, the LSTM input window is shortened.

[0087] Shortening the window helps reduce the interference of long periods of large price fluctuations on model predictions, making the model more sensitive to short-term price fluctuations and helping to improve the real-time performance and accuracy of predictions.

[0088] This adaptive window mechanism dynamically adjusts the length of the input sequence based on actual market fluctuations. This allows the LSTM network to focus more on recent data changes during periods of high price volatility, while smoothing out fluctuations within the time window during periods of low market volatility, thus avoiding over-responding to short-term fluctuations. This enables the model to flexibly adapt to different market environments, enhancing its adaptability and predictive ability in the face of complex market dynamics.

[0089] The market volatility threshold is determined by calculating the median of the peak-to-trough difference in historical price data. The peak-to-trough difference represents the maximum price fluctuation within a certain period. Using the median as the volatility threshold effectively avoids the impact of extreme price fluctuations on the model, making the threshold more stable and representative.

[0090] Using the median of the peak-to-trough difference to determine the market volatility threshold can effectively suppress the impact of extreme price fluctuations on model predictions, preventing the system from overreacting to sudden price volatility. Furthermore, the median represents typical volatility patterns in historical data, enhancing the stability and applicability of the threshold.

[0091] By optimizing the input sequence, window adaptation, and dynamic adjustment of the market volatility threshold, the LSTM time series network can more accurately address the uncertainty of market price fluctuations in construction project cost forecasting. Input sequence optimization provides more accurate market volatility data, window adaptation adjusts the time frame the model focuses on based on market fluctuations, and the determination of the market volatility threshold provides a more stable and effective forecasting basis. Overall, this method improves the real-time performance, accuracy, and stability of construction project cost forecasting, adapts to price fluctuations under different market conditions, and enhances the reliability of the model in practical applications.

[0092] In one possible implementation, the spatiotemporal fusion feature tensor combines multidimensional data from time and space, enabling the system to capture material price trends based on different spatiotemporal information. By extracting a subset of features related to material prices, the system can focus on analyzing core data related to cost fluctuations.

[0093] For each building material, the system calculates the difference between the maximum and minimum prices within the forecast period, obtaining the price range. This difference reflects the fluctuation range of the material price within that period and is an important parameter for measuring the level of risk.

[0094] Extracting a subset of material price features helps remove irrelevant data interference and improves the accuracy of data analysis. Calculating the price range provides a quantitative indicator that clearly reflects the fluctuation range of material prices, providing basic data for subsequent risk assessment.

[0095] The calculated price range is compared with various abnormal price fluctuation events of building materials in a historical engineering risk database. This database stores various price fluctuation events that occurred over the past five years, along with actual loss rates. This matching process allows the system to dynamically assess the risk level of the current project based on historical data.

[0096] When the price range consistently exceeds the first risk threshold and reaches the preset monitoring period, the system marks it as medium risk. The first risk threshold is determined based on the 90th percentile of historical material price fluctuations; therefore, this threshold has high sensitivity and can effectively identify situations where price fluctuations exceed the normal fluctuation range.

[0097] If the price range exceeds the second risk threshold, and the volatility growth rate accelerates in adjacent monitoring periods, it is marked as high risk. The second risk threshold adopts the red alert benchmark value issued by an industry price warning agency, which helps to identify extreme market volatility in a timely manner.

[0098] By matching price range values ​​with historical data, the system can make more accurate risk predictions based on past experience. Dynamic assessments based on price fluctuations can react quickly to rapid changes, ensuring timely adjustments to project risk management strategies.

[0099] At the medium risk level, the system increases the reserve base based on the product of the price range and the basic reserve rate. This means that when price volatility reaches a moderate level, the system will increase the reserve proportionally to cope with potential cost fluctuations.

[0100] At high-risk levels, in addition to increasing the reserve base based on the price range, the system will also add an additional base based on the duration of the fluctuation and the maximum daily fluctuation. This mechanism is designed to provide additional financial protection for the project in the event of extreme volatility.

[0101] This tiered adjustment mechanism allows for more precise and flexible setting of contingency funds, enabling adjustments to capital reserves based on different risk levels. This measure helps project management teams better cope with market uncertainties, especially when dealing with materials with highly volatile prices, ensuring that projects have sufficient funds to manage risks.

[0102] By combining historical risk databases with spatiotemporal characteristics, the potential risks of construction projects can be assessed more accurately. Dynamic risk classification and grading adjustments enable project teams to flexibly respond to different market conditions based on real-time data.

[0103] By accurately predicting risks and allocating contingency funds appropriately, excessive or insufficient fund reserves are avoided. This not only improves the efficiency of fund utilization but also reduces additional costs caused by improper risk management.

[0104] Especially in environments with significant price fluctuations, the system can provide ample financial support for projects, reduce the negative impact of market changes, and enhance the project's resilience to risks.

[0105] By combining artificial intelligence technology and data analysis, accurate cost forecasts are provided for construction projects. Through risk feature extraction, historical data matching, dynamic risk assessment, and tiered contingency fund adjustments, projects can more accurately predict and respond to potential price fluctuation risks. The implementation of this process not only improves the financial management of construction projects but also effectively ensures their smooth progress.

[0106] In one possible implementation, the ReLU activation function is applied to the geometric features. ReLU is a common activation function that avoids the interference of negative values ​​in subsequent calculations by converting negative values ​​to zero. Geometric features (such as building structures, areas, shapes, etc.) often have a certain linear relationship, and ReLU can effectively preserve the positive relationship while eliminating the negative impact of negative values ​​on model training.

[0107] The LeakyReLU activation function is applied to the constraint rule features. LeakyReLU is similar to ReLU, but it assigns a small negative value (e.g., 0.01) when the input is less than zero, thus preserving some gradient information. This is particularly important for constraint rule features, as these features may involve complex boundary conditions or constraint relationships. The introduction of LeakyReLU helps the network better preserve gradient information, avoids information loss, and avoids the "dead neuron" problem that can occur with conventional ReLU.

[0108] The output layer uses the Sigmoid activation function to normalize the output to a value between 0 and 1. The Sigmoid function is suitable for binary classification tasks or scenarios that require normalized predictions. In this way, the predicted value can be standardized, which helps with subsequent decision-making or further analysis.

[0109] The choice of activation function ensures that gradients can be effectively propagated in each layer of the network, avoiding gradient vanishing or exploding, and improving the training effect and stability of the model.

[0110] The ReLU and LeakyReLU functions can be used to process and optimize nonlinear relationships in input data, thereby improving the learning ability of the network.

[0111] The normalized output of the Sigmoid function ensures a reasonable range for the prediction results, making the model more accurate and stable when dealing with cost prediction.

[0112] The number of nodes in the hidden layer grows logarithmically based on the number of quantifiable units. This means that as the number of quantifiable units involved in a project (such as different types of building materials, processes, labor, etc.) increases, the number of nodes in the hidden layer will increase accordingly. This logarithmic growth ensures that the network complexity scales reasonably when dealing with a large number of features, without leading to overfitting or excessively low computational efficiency.

[0113] The number of nodes in the output layer is equal to the number of cost index categories. Each output node represents a specific cost index (such as material costs, labor costs, equipment costs, etc.). This setup ensures that the model can make independent predictions for each cost index, and the prediction results accurately reflect the changes in each index.

[0114] The logarithmic growth of the number of hidden layer nodes ensures that the network is complex enough to adapt to various feature data, while avoiding the waste of computation and training time that may result from too many nodes.

[0115] The number of output layer nodes is consistent with the number of cost index categories, which enables the model to maintain a clear and efficient structure in multi-index prediction and avoids redundant or erroneous output.

[0116] By employing a three-layer fully connected network structure and activation function adaptation, data processing and model training in construction project cost prediction are optimized. The combination of ReLU, LeakyReLU, and Sigmoid functions ensures the network's effective processing of different features, and the number of nodes is reasonably adjusted according to the amount of data and task complexity, thereby improving prediction accuracy, computational efficiency, and network scalability. These designs ensure that the artificial intelligence model has high practical application value and effectiveness in construction project cost prediction.

[0117] This embodiment demonstrates the process of predicting construction project costs using deep learning methods according to the present invention, and showcases its excellent predictive capabilities through specific parameter settings, training processes, and evaluation indicators. Compared with traditional linear regression models, the present invention can more accurately capture complex nonlinear relationships, significantly improve prediction accuracy, and has outstanding practical application value.

[0118] It is understood that, in order to enable those skilled in the art to clearly and completely implement this invention, the specific setting basis and methods of various thresholds, critical values ​​and parameters involved in the technical solution are now explained in detail. These settings are based on recognized standards in the field of construction engineering, historical data statistical analysis methods, and commonly used strategies in the field of machine learning, ensuring the scientific nature, repeatability and adaptability of the method.

[0119] Regarding the threshold setting for design change identification: Displacement Threshold: This threshold is used to determine whether the spatial positional change of a component is significant. It is set based on: collecting deformation data of various components (such as beams, slabs, and columns) within the allowable construction error range from a large number of historical engineering projects to form a statistical sample. The 90th percentile of this sample data is taken as the displacement threshold. This method ensures that the threshold covers the vast majority (90%) of normal construction deformation, only marking abnormal displacements exceeding this range, thus balancing sensitivity and reliability. This threshold can be set separately according to different component types (such as main structural components and decorative components).

[0120] Structural stability threshold: This threshold is used to determine the importance of structural changes through variations in mechanical characteristics. It is set based on the following: First, the limit requirements for inter-story drift angles are obtained from the current seismic design code (such as the "Code for Seismic Design of Buildings" GB50011) for the project location. Then, a standard model is established using finite element analysis software, and the rate of change of eigenvalues ​​of the force transmission path matrix (such as the stiffness matrix) at key beam-column joints under the specified displacement limit is calculated. This rate of change is used as the critical point, i.e., the structural stability threshold. This threshold will be dynamically adjusted as the code is updated and the seismic fortification intensity of the project location varies.

[0121] Regarding the parameter settings for generating quantifiable units: Critical value for rebar binding process: This value is used to determine whether the rebar distribution density has reached a level requiring merging of units. It is set based on different rebar diameter categories (e.g., Φ<12mm, Φ12-25mm, Φ>25mm), referring to the process requirements for rebar binding density in national construction acceptance specifications and construction method standards, and combining expert experience. The basic principle is: the larger the diameter, the stronger the rebar rigidity, and the higher the difficulty of binding construction; therefore, the allowable density gradient critical value is correspondingly reduced. For example, the maximum density gradient value for reliable binding of rebar of different diameters in the joint area can be determined through experiments or historical data statistics.

[0122] Locally Commonly Used Formwork Standard Dimensions: This is a basic data parameter. Its value is directly derived from the specifications of common formwork materials in the construction market of the project implementation area (such as 915mm×1830mm, 1220mm×2440mm, etc.), which can be obtained by surveying the local building materials market or consulting regional standard drawing sets. The system has a built-in library of commonly used dimensions, which can be updated according to market changes.

[0123] Parameters and functions related to constraint rule transformation: Stirrup Spacing Reduction Ratio Function: This function dynamically calculates the amount of stirrup reinforcement required due to an increase in seismic resistance level. The function is a linear or nonlinear empirical formula based on engineering experience. Its key parameters are linked to the design seismic acceleration value of the project site, the building importance coefficient, and the measured or standard yield strength value of the steel reinforcement used in the components. The basic logic is: the larger the seismic acceleration value, the larger the reduction ratio; the higher the yield strength of the steel reinforcement, the smaller the required reduction ratio. Specific coefficients can be obtained through regression analysis of historical engineering data (actual construction drawings under different seismic resistance levels).

[0124] Rebar Length Compensation Piecewise Function: This function calculates the increased anchorage length required due to the thickness of the protective layer. The function is based on the basic anchorage lengths for different rebar diameters specified in national structural design codes (such as the "Code for Design of Concrete Structures" GB50010). The compensation calculation is typically divided into two parts: when the rebar diameter is less than or equal to a certain critical value (e.g., 25mm), the compensation is linearly related to the diameter; when the diameter is greater than this critical value, considering the complexity of anchoring thicker rebars, the compensation increases according to a quadratic curve, with the growth coefficient determined through examples or experimental data in the code's explanatory notes.

[0125] Search strategies for component dependency analysis: Three-level / five-level correlation search strategy: The level setting of this strategy is based on a balance between common sense in engineering mechanics and computational efficiency. In general structures, force transmission significantly attenuates after three levels. Therefore, "three-level correlation" is used as the standard search depth. When the correlated components are of the same material type, their mechanical behavior is more consistent, and the influence transmission may be farther. Therefore, it is extended to "five-level correlation" to more comprehensively assess the scope of influence. This level is an empirical value and can be adjusted through the system configuration interface according to the scale of the project (such as large and complex structures) in practical applications.

[0126] Regarding the parameter settings for the cross-channel attention mechanism: Relevance Threshold: This threshold is used to determine whether the correlation between static geometric features and dynamic price features is sufficient. It is set based on the following: In the historical engineering database, the cosine similarity between the price series of each material and the corresponding engineering quantity series is calculated, and the distribution of these similarity values ​​is statistically analyzed (e.g., the 25th percentile is used). If the feature similarity of the current project is lower than this historical statistical value, the correlation is considered insufficient, and the weight needs to be reduced. This threshold is automatically updated periodically as the historical database expands.

[0127] Construction phase weight adjustment ratio: This is a rule-based setting, not a fixed value. For example, "increase the static channel weight according to the excavation depth ratio" means: the static channel weighting coefficient = 1 + α × (current excavation depth / total excavation depth). Here, α is an amplification factor, which can be set to 0.5 or 1 based on experience. Similarly, "increase the dynamic channel weight according to material price fluctuations" can be expressed as: dynamic channel weighting coefficient = 1 + β × (recent price volatility). α and β are configurable parameters, with default values ​​provided during implementation, which users can fine-tune according to project characteristics.

[0128] Regarding the threshold setting for feedback optimization and risk prediction: Deviation threshold: The lower limit of prediction deviation used to trigger feature weight adjustments. This value is set based on the statistical distribution of historical prediction errors, typically the root mean square error (RMSE) or the 75th percentile of historical prediction errors for similar components. Different deviation thresholds can be set for different types of components (such as concrete, steel reinforcement, and formwork).

[0129] Market volatility threshold: The lower limit of price volatility used to trigger adaptive adjustment of the LSTM input window. It is set based on the following: for each material, calculate the sequence of all peak-to-trough differences in its price data over the past year (or a typical cycle), and take the median of this sequence as the threshold. The median effectively avoids interference from extreme outliers and represents the typical price volatility of the material.

[0130] Risk level classification threshold: First risk threshold (medium risk): The 90th percentile of the historical price volatility (daily or weekly) of the material over the past five years. Exceeding this value indicates that price volatility has entered a relatively high historical range.

[0131] The second risk threshold (high risk): directly adopt the "red price warning" line or similar benchmark value regularly issued by national or industry-recognized building materials price index early warning agencies (such as the Cement Association, the Iron and Steel Industry Association, etc.). This is to maintain consistency with the overall risk assessment of the industry.

[0132] Accelerated volatility growth rate: This is determined by calculating the absolute value of the first derivative of the price series (i.e., the daily change) and judging whether its moving average over a certain time window (e.g., 5 days) continues to rise (i.e., the second derivative is positive).

[0133] Contingency reserve base calculation parameters: Basic contingency rate: It is usually set with reference to industry practice or the company's internal control standards. For example, it is generally taken as 1%-3% of the total project cost as the initial basic rate.

[0134] Additional base calculation: Additional base under high risk = (Number of days of fluctuation - Number of warning days) × (Maximum daily fluctuation amount). "Number of warning days" is a built-in system parameter, representing the number of days during which temporary price anomalies are allowed without triggering additional fees, and can be set according to material characteristics.

[0135] Hidden layer node growth method: Number of nodes = γ × log(1 + N). Where N is the number of quantifiable unit types, γ is a scaling factor (usually a power of 2 such as 32, 64, etc.), and the logarithmic base can be the natural logarithm e or 10. This growth strategy aims to make the network capacity sublinearly related to the problem complexity (unit types), preventing overfitting while ensuring the model's expressive power. The specific value of γ can be determined through cross-validation on historical data.

[0136] Number of output layer nodes: directly equal to the number of cost index categories to be predicted, such as "unit project comprehensive unit price", "measure item cost coefficient", "risk reserve ratio", etc. This is determined by the problem definition and does not require additional setting.

[0137] Example: Application of cost prediction for an office building construction project; I. Project Background and Data Input; This project involves the construction of a new 6-story office building in a certain location, using a reinforced concrete frame structure, with a total construction area of ​​8,600 square meters. The building information model contains 1,200 structural components, including 400 beams, 300 columns, and 500 slabs. The system incorporates final settlement data from 50 similar projects in the region over the past three years as a historical reference.

[0138] The input data consists of three parts: First, geometric data provided by the Building Information Model (BIM), including beam cross-sections of 300 mm x 600 mm, column cross-sections of 600 mm x 600 mm, and concrete strength grade of C30; second, market price data, including the weekly price series of rebar by particle size from January to June 2023, specifically RMB 4200, RMB 4350, RMB 4500, RMB 4800, RMB 5100, and RMB 5300 per ton, and the price of commercial concrete of RMB 380 per cubic meter; third, design change data, including a positional shift of the frame column at the intersection of axis 5 and axis 8 on the 3rd floor, with an 8 mm shift in the X direction and a 6 mm shift in the Y direction.

[0139] II. Data Processing and Feature Fusion; First, a design change marker is created. The displacement of the column member is calculated, and the actual displacement is determined to be 10 mm using the Pythagorean theorem. This is compared to a displacement threshold, which is the 90th percentile of the allowable deformation distribution of members in historical engineering data, and is set at 5 mm. Since 10 mm is greater than 5 mm, the system marks this column member with the number C-3-5-8.

[0140] Structural stability analysis was then performed. The force transmission path matrix at the beam-column joints was extracted using finite element analysis, and the eigenvalue change rate was calculated to be 8.7%. This value exceeds the structural stability threshold of 6.5% calculated according to the local seismic code. Therefore, the associated components were expanded to level three, including two beams and one floor slab.

[0141] Next, a priced unit is generated. The density gradient of reinforcement distribution in the beam-column joint area is calculated. The density in the joint area is 35 kg / m³, and in the non-joint area it is 18 kg / m³, with a gradient value of 34 kg / m³. This is compared to the critical value for the binding process, which is determined based on the reinforcement diameter category. For 25 mm diameter reinforcement, the critical value is 30 kg / m³. Since 34 is greater than 30, the system merges the two intersecting beams and one column into a unified priced unit, numbered U-1.

[0142] The process of breaking down the formwork support unit is as follows: The wall length is measured to be 8.4 meters, which is divided by the local standard formwork size of 1.2 meters to obtain a theoretical number of 7 units. After adjusting according to the location of door and window openings, it is actually broken down into 6 units, ensuring that each unit contains a complete construction joint.

[0143] The application of constraint rules includes two aspects: First, dynamic correction of stirrup spacing. Based on the seismic acceleration value of 0.2g and the building importance coefficient of 1.2 at the project site, combined with the yield strength of 400 MPa of steel bars, the spacing reduction ratio is calculated to be 0.06%, that is, the original spacing of 200 mm is reduced to 188 mm; Second, steel bar length compensation. Based on the concrete cover thickness of 30 mm, a piecewise function is used to calculate the compensation amount of 32.5 mm for 25 mm diameter steel bars.

[0144] III. Dual-channel dynamic prediction; The static processing channel employs a three-layer fully connected network. The input layer has 120 nodes, corresponding to the geometric feature dimension. The number of hidden layer nodes increases logarithmically with the number of quantifiable unit types. The output layer has 3 nodes, corresponding to the comprehensive unit price of the unit project, the coefficient of the measure item cost, and the proportion of the risk reserve fund. The activation functions are configured as follows: the first layer uses the ReLU function to eliminate negative value interference, the second layer uses the LeakyReLU function to preserve gradient characteristics, with a negative slope set to 0.01, and the output layer uses the Sigmoid function for normalization.

[0145] The dynamic processing channel uses an LSTM time series network. First, the variance of the rebar price series is calculated, resulting in 184500. The market fluctuation threshold is taken as the median of the difference between the peak and trough of the material's historical price, which is 280 yuan. Since the square root of the variance, 429.5, is greater than 280, the system automatically shortens the input series time window to 50% of its original length, i.e., adjusts it to 3 weeks.

[0146] The cross-channel attention mechanism performs dynamic weighting based on similarity. The cosine similarity between the static channel's geometric feature vector and the dynamic channel's price feature vector is calculated, yielding 0.73. The relevance threshold is set to the 25th percentile of historical engineering data, which is 0.68. Since 0.73 is greater than 0.68, the dynamic weight coefficient is maintained at 1.0. Simultaneously, weights are adjusted according to the construction stage. Currently, in the main construction stage, the static channel's weight is increased proportionally to the excavation depth, with an enhancement coefficient of 1.5.

[0147] The risk reserve calculation process is as follows: First, the price range is calculated to be 1100 yuan per ton. The first risk threshold is taken as the 90th percentile of the historical price fluctuation statistics for this material, which is 980 yuan per ton. The second risk threshold is taken as the red alert benchmark value of 1200 yuan per ton issued by the industry price warning agency. Since the price range has exceeded the first risk threshold for four consecutive weeks but has not exceeded the second risk threshold, it is marked as a medium risk level. The reserve base is calculated by multiplying the price range by the basic reserve rate, which is 2%, resulting in a calculation of 2.2%.

[0148] IV. Prediction Results and Comparative Analysis; The output results of this invention are as follows: the comprehensive unit price of the project is 1285 yuan per square meter; the coefficient of the measure project cost is 8.7%; the risk reserve ratio is 2.2%; and the total cost forecast is 12.1 million yuan.

[0149] A comparative analysis was conducted with traditional forecasting methods. Regarding design change response, traditional methods require manual input with an average delay of 3 days, while this invention achieves real-time automatic marking, improving timeliness by 100%. In terms of price fluctuation adaptation, traditional methods use monthly updates, while this invention uses weekly adaptive adjustments, improving accuracy by 23%. Regarding risk reserve deviation, traditional methods have a deviation range between +35% and -20%, while this invention controls it within ±5%, improving stability by 40%. The overall forecast deviation decreased from 12.7% of traditional methods to 3.5%, improving accuracy by 72%.

[0150] The specific effects are reflected in three aspects: First, by capturing the price fluctuations of rebar in real time, the weight matrix was adjusted in time when the price rose to 5,300 yuan per ton in May, avoiding a loss of 1.56 million yuan; second, by automatically locating the third-level related components through the component dependency graph, two beam components were found and included, avoiding the omission of 380,000 yuan in construction costs; third, the risk reserve fund ratio was accurately matched with the actual fluctuations, and the final actual price difference was 1,050 yuan per ton, with the reserve fund utilization rate reaching 95%.

[0151] V. Feedback and Optimization Process; The actual final settlement value after project completion was 12.54 million yuan, with a prediction deviation of 3.5%. Deviation analysis revealed that the actual concrete usage exceeded the predicted value by 8%. The system retrieved the usage distribution of similar components from the historical project database and used the 75th percentile value as the adjustment benchmark. The feature extraction weights were adjusted accordingly: the weight of geometric features was increased from 0.6 to 0.65, and the weight of dynamic channels was decreased from 0.4 to 0.35. Through this feedback optimization mechanism, the system achieved self-learning and continuous improvement, providing a more accurate benchmark for subsequent predictions.

[0152] In summary, all thresholds and parameters in this invention are set based on engineering specifications, historical data statistics, best practices in machine learning, or industry consensus. Some parameters provide configurable interfaces to adapt to different application scenarios. Those skilled in the art can fine-tune the above principles and methods in conjunction with the specific project implementation details to obtain the best prediction results.

[0153] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0154] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An artificial intelligence-based construction project cost prediction method, characterized by, The method comprises the following steps: Step 1: Multi-source heterogeneous data collaborative collection: real-time acquisition of component geometric attribute set in building information model and historical engineering completion settlement data set, dynamic capture of building material market price data, and marking of unique identifier set of modified components when design change is detected; Step 2: Data preprocessing and feature fusion: merging building information model components into billable units according to construction technology, generating constraint rule set combining regional construction specifications, and generating dynamic price weight matrix according to market price fluctuation; when design change occurs, affected unit range is located through component dependency graph and engineering quantity is recalculated; Step 3: Double-channel dynamic prediction: constructing a static processing channel to analyze geometric features and constraint rules, and constructing a dynamic processing channel to analyze price time series features, fusing output space-time feature tensor through cross-channel attention mechanism, and outputting unit engineering comprehensive unit price, measure item cost coefficient and risk reserve fee proportion based on space-time fusion feature tensor regression; Step 4: Feedback optimization: real-time triggering of local recalculation according to market data update or design change, and adjusting feature extraction weight based on historical engineering data deviation.

2. The method of claim 1, wherein the method further comprises: The marking of design change in step 1 comprises: Obtaining the spatial positioning coordinate set of the components before and after the change, and calculating the components whose coordinate change exceeds the displacement threshold; Analyzing the topological connection relationship of the components: extracting the force transmission path matrix at the beam-column joint, and when the matrix eigenvalue change rate exceeds the structural stability threshold, the associated column components, beam components and connecting plate components are added to the identifier set; The displacement threshold is determined by the 90th percentile value of the statistical distribution of the allowable deformation of the components in the historical engineering, and the structural stability threshold is converted from the allowable value of the structure deformation angle in the local seismic specification to the critical point of the matrix eigenvalue change rate.

3. The method of claim 1, wherein the method further comprises: The generation of the billable unit in step 2 comprises: Identifying the steel binding unit: calculating the steel distribution density gradient in the beam-column joint area, and when the gradient value exceeds the binding process critical value, the intersecting beam components and column components are merged into a single unit; the binding process critical value is determined according to the diameter category of the steel, and the larger the diameter, the lower the critical value; Splitting the formwork support unit: according to the theoretical splitting number obtained by dividing the wall length value by the local commonly used formwork standard size value, and adjusting the splitting point according to the door and window hole position to ensure that each unit contains a complete construction joint.

4. The method of claim 1, wherein the method further comprises: The conversion of the constraint rule set in step 2 comprises: Dynamic correction of stirrup spacing: obtaining the seismic acceleration value and building importance coefficient of the project location, calculating the spacing reduction ratio caused by the improvement of seismic grade, and the reduction ratio is positively correlated with the seismic acceleration value and negatively correlated with the steel yield strength value; Steel length compensation: increasing the anchoring segment length according to the concrete cover thickness specification value, and the compensation amount is calculated by a segmented function according to the steel diameter category, and when the diameter is greater than the critical value, the compensation amount increases according to a quadratic curve.

5. The method of claim 1, wherein the method further comprises: The construction of the component dependency graph in step 2 comprises: Taking the modified component as the root node, searching for three-level associated components along the mechanical transmission path: first-level association is the directly connected components, second-level association is the components that indirectly transmit load through connecting pieces, and third-level association is the components in the same force system; When the associated components are of the same material type, the search is extended to five-level associated components; The mechanical transmission path is determined by overlapping the stress nephograms of the components extracted by finite element analysis.

6. The method of claim 1, wherein the method further comprises: The cross-channel attention mechanism in step 3 includes: Dynamic similarity weighting: Calculate the cosine similarity between the static channel geometric feature vector and the dynamic channel price feature vector. When the similarity is lower than the correlation threshold, reduce the price feature weight; Construction phase adaptation: Increase the static channel weight by the foundation construction stage according to the foundation excavation depth ratio, and increase the dynamic channel weight by the decoration stage according to the material price fluctuation range; The correlation threshold is dynamically adjusted according to the statistical quantile of the deviation between the material price and the quantity in the historical project.

7. The method of claim 1, wherein the method further comprises: The adjustment of the feature extraction weight in step 4 includes: Deviation component identification: When the predicted value of a specific category component deviates from the final account value by more than a threshold, search for the usage distribution of the same type of component in the historical project; Weight correction: Adjust the feature extraction coefficient according to the 75th percentile value of the usage distribution. If the actual value is consistently higher than the predicted value, increase the geometric feature weight. If the actual value is dominated by price fluctuations, increase the dynamic channel weight.

8. The method of claim 1, wherein the method further comprises: The LSTM time series network operation of the dynamic processing channel includes: Input sequence optimization: Divide the dynamic price weight matrix by week granularity, and calculate the price fluctuation variance value of adjacent sub-sequences; Window self-adaptation: When the variance value exceeds the market fluctuation threshold, shorten the input sequence time window to a fixed proportion of the original window length; The market fluctuation threshold is determined according to the median of the peak-to-valley difference of the historical price of the material.

9. The method of claim 1, wherein the method further comprises: The generation process of the risk preparation fee ratio in step 4 includes: Risk feature extraction: Separate the material price dimension feature subset from the spatio-temporal fusion feature tensor; Calculate the difference between the maximum and minimum values of the subset within the prediction period as the price range value; Dynamic risk level division: Match the price range value with the threshold interval in the historical project risk database, which stores abnormal fluctuation events of various building materials in the past five years and their corresponding actual loss rates; When the price range value continuously exceeds the first risk threshold for a preset monitoring period, it is marked as a moderate risk level; When the price range value exceeds the second risk threshold and the fluctuation growth rate of the adjacent monitoring period accelerates, it is marked as a high risk level; Preparation fee base hierarchical adjustment: Moderate risk level: Increase the calculation base by the product of the price range value and the basic preparation fee rate; High risk level: Additional increase the additional base based on the fluctuation duration and the maximum fluctuation amplitude per day; Where: The first risk threshold is determined according to the 90th percentile value of the historical price fluctuation statistics of the material; The second risk threshold adopts the red alert benchmark value released by the industry price warning agency; The fluctuation growth rate is calculated by the absolute value of the first derivative of the price sequence.

10. The method of claim 1, wherein the method further comprises: The three-layer fully connected network of the static processing channel includes: Activation function adaptation: The first layer uses ReLU function to eliminate negative value interference, the second layer uses LeakyReLU function to retain gradient characteristics, and the output layer uses Sigmoid function for normalization; The number of nodes is set: the number of nodes in the hidden layer increases according to logarithmic relation with the number of kinds of the valuing units, and the number of nodes in the output layer equals the number of the cost index categories.

Images