Convolutional neural network and IRLS integrated dynamic prediction method for assembly type construction cost

Abstract

The application discloses a kind of convolutional neural network and IRLS fusion's fabricated construction cost dynamic prediction method, belong to specific computer model technical field, including step construction dataset D, construction including single-day weight generation network, single-day weight optimization unit, weight matrix output unit, least square solution unit and cost prediction unit's cost prediction network;Loss function is constructed;Training cost prediction network obtains cost prediction model, for subsequent cost prediction in unconstructed date.This application is directed to the contribution degree of different construction data in the process of fabricated building construction, effectively identifies and suppresses the noise interference of sudden factors, improves the prediction stability.Only local construction stage data is needed to realize high-precision cost prediction, improve the flexibility and practicality of cost management.It also has the characteristics of strong practicality and reliability, wide industry applicability.

Description

Technical Field

[0001] This invention relates to the field of specific computer modeling technology, and in particular to a method for dynamic prediction of prefabricated construction costs by fusing convolutional neural networks and IRLS. Background Technology

[0002] Prefabricated construction is a modern construction method that shifts much of the traditional on-site work to factories. In the factory, the main components of a building (such as walls, floors, stairs, beams, columns, and even entire bathrooms) are prefabricated as standardized "components" or "modules." Once completed, these components are transported to the construction site and assembled and spliced ​​like building blocks using reliable connection methods (such as bolts, welding, and pouring) to ultimately form a complete building.

[0003] Prefabricated buildings, as an important representative of the industrialization and intelligent transformation of the construction industry, rely heavily on the production, transportation, and on-site assembly of prefabricated components during their construction process. They are characterized by numerous stages, high requirements for coordination, and significant susceptibility to external environmental influences. Cost prediction methods for prefabricated buildings include traditional methods such as regression analysis and case analogy, as well as mathematical optimization methods such as static least squares.

[0004] Traditional forecasting methods lack dynamic adaptability and cannot respond in real time to unforeseen circumstances during construction, such as weather changes, supply chain disruptions, and fluctuations in labor efficiency, leading to significant discrepancies between forecasts and actual results. For example, regression analysis, based on the assumption of linear relationships between variables, cannot effectively handle complex cost variations caused by differences in component standardization, hoisting efficiency fluctuations, and changes in the efficiency of multi-trade collaboration, resulting in a disconnect between the forecasting model and actual complex working conditions. Case analogy methods, on the other hand, rely too heavily on historical experience and subjective judgment, making it difficult to quantify differences in construction environment, management level, and worker skill levels across different projects. They also lack adaptability and flexibility, especially when dealing with unforeseen circumstances (such as severe weather, supply chain disruptions, and policy adjustments). Therefore, it has the following defects: (1) Traditional prediction methods lack dynamic adaptability and cannot respond to sudden situations in construction in real time, resulting in a large deviation between the prediction results and the actual situation; (2) It relies on complete historical data and cannot make accurate predictions based on fragmented construction data (such as partial hoisting stage and progress of some floors) in the middle of the project, which limits its application in dynamic decision-making; (3) Construction data often contains noise and outliers (such as sudden increase in daily working hours and abnormal fluctuations in logistics costs), and traditional methods are difficult to automatically identify and suppress these interferences, affecting the stability and reliability of prediction; (4) Although the prediction model has high accuracy, it lacks interpretability and is computationally complex, making it difficult to achieve real-time deployment and operation on construction sites with limited resources.

[0005] While mathematical optimization methods such as static least squares have improved fitting accuracy to some extent through mathematical means, they still fail to address issues such as fixed weights, sensitivity to outliers, and reliance on complete project data. This limits their application in mid-construction cost control and dynamic decision-making. Therefore, the industry urgently needs a dynamic cost management technology that can respond to construction changes in real time and make high-precision predictions based on only partial data to improve cost control capabilities and management intelligence in prefabricated building projects.

[0006] Definitions: IRLS (Iterative Reweighted Least Squares) is a method for solving linear regression problems. It reduces the impact of outliers on the fitting results by introducing weight functions. The idea is to iteratively adjust the weights and solve the weighted least squares problem to obtain more accurate estimates. Summary of the Invention

[0007] The purpose of this invention is to provide a method for dynamic cost prediction of prefabricated construction by integrating convolutional neural networks and IRLS. This method not only breaks through the dependence of traditional cost prediction methods on complete project cycle data, but also achieves high-precision cost prediction with only local construction stage data, such as the hoisting of prefabricated components and the assembly of floors. Furthermore, it can adaptively adjust the contribution of different construction data to address nonlinear fluctuations, accurately capture the complex changes of various factors, effectively suppress noise interference, and improve prediction stability.

[0008] To achieve the above objectives, the technical solution adopted by this invention is as follows: a method for dynamic prediction of prefabricated construction costs by fusing convolutional neural networks and IRLS, where T1 days have been completed, T2 days have not been completed, and the total construction period is T. 总 =T1+T2, including the following steps; S1, Construct dataset D, including steps S11~S13; S11 divides the construction management level, the skill level of technical workers, and the on-site environmental conditions into several levels from high to low, and assigns a value to each level in the range [0,1]. The higher the level, the larger the value assigned. S12, using a sliding window of length L, acquire data from day t-L+1 to day t, and construct the sample x for day t. t 10-dimensional features x t,1 ~x t,10 x t =(x t,1 ,x t,2 ,⋯,x t,10 ), L≤t≤T1-∆T, where ∆T is the prediction time difference; x t,1x represents the total building area. t,2 To determine the prefabrication rate, x t,3 x represents the total number of lifting operations. t,4 The standard deviation σ of the working hours consumed in L days inside the sliding window. t ; x t,5 Let be the rate of change of the amount completed on day t. ; x t,6 Let be the cost density on day t. ; x t,7 To determine the dominant frequency energy characteristics on day t, the logistics costs for L days within the sliding window are sequenced and transformed using FFT to generate L frequency components arranged in descending order as X(1)~X(L). The results are then calculated using the formula: X(v) is the v-th frequency component in descending order; x t,8 For management efficiency coefficient, ; x t,9 The efficiency coefficient is the ratio of personnel proficiency. ; x t,10 To design the complexity prefabrication rate interaction item, x t,10 = Value assigned to the on-site environmental condition level on day t × Design prefabrication rate; S13, starting from day L of construction, sequentially generate feature matrices for all days already completed, forming dataset D, where the feature matrix for day t is X(t). ; S2. Construct a cost prediction network, including a daily weight generation network, a daily weight optimization unit, a weight matrix output unit, a least squares solution unit, and a cost prediction unit. The daily weight generation network is used as input x t Generate the weight w for day t. t ; The daily weight optimization unit is used to determine x based on preset conditions. t Is it an outlier? Generate the optimized weight w' for day t. t If x t If w' is an outlier, then t =w t ×0.3, otherwise w' t =w t ; The weight matrix output unit is used to calculate the optimized weights w' from day t-L+1 to day t. t-L+1 ~w' t Generate the weight matrix W(t+∆T) for day t+∆T, where W(t+∆T)=diag(w't-L+1 ,w' t-L+2 ,…,w' t ), where diag(∙) is the diag function; The least squares solving unit is used to solve for the characteristic coefficient vector β on day t+∆T according to the following formula. t+∆T ; , Where λ = 1e-5 is the ridge parameter, I is the identity matrix, and T is the transpose operation. ~ β t+∆T The first to the tenth characteristic coefficients; The cost forecasting unit is used to calculate the forecast cost for day t+∆T according to the following formula. ; , In the formula, β0 is the preset baseline cost; S3, construct the loss function Loss for the cost prediction network; , In the formula, i represents one day from day t-L+1 to day t. The optimization weight for day i. , Let be the actual cost and the predicted cost on day i, respectively; α be the L2 regularization strength; and θ be all trainable parameters of the cost prediction network. S4. Train the cost prediction network using dataset D, and adjust θ to minimize the loss until the cost prediction network converges to obtain the cost prediction model. Mark the feature coefficient vector β at this time as the optimal feature coefficient vector β*. S5, for the number of days without construction d, if d-∆T days are within the number of days with construction, then construct a sample of d-∆T and substitute it into the cost prediction unit to calculate the predicted cost for day d using β*.

[0009] Preferably, in S11, the construction management level is divided into 1 to 4 levels, with corresponding values ​​of 1.0, 0.7, 0.4, and 0.1, respectively; the skilled worker proficiency is divided into 1 to 3 levels, with corresponding values ​​of 1.0, 0.6, and 0.2, respectively; and the on-site environmental conditions are divided into 1 to 4 levels, with corresponding values ​​of 1.0, 0.7, 0.4, and 0.1, respectively.

[0010] Preferably, the daily weight generation network includes an input layer, a one-dimensional convolutional layer, a layer normalization layer, a max pooling layer, a feature extraction layer, a feature concatenation layer, a first fully connected layer, and a second fully connected layer arranged in sequence. The input layer is used to process x t Mapped to 16×3 input features ; The one-dimensional convolutional layer is used to generate convolutional features c according to the following formula. t ; , In the formula, W conv (k), b conv (k) represents the weights and biases of the convolution kernel k, respectively, and c t ∈R 16×3 ; The normalization layer and max pooling layer sequentially perform normalization and max pooling operations on c(t), outputting a 16×1 pooling feature. ; The feature extraction layer is used for calculation. Mean, variance, and peak value of pooling features; The feature splicing layer is used to... The concatenation feature F, which is concatenated with the mean, variance, and peak value, is 19×1. t ; The first fully connected layer and the second fully connected layer respectively use the SiLU activation function and the Softplus activation function to convert F t Adjust to 64 dimensions and 1 dimension, output the weight w for day t. t .

[0011] Preferably, the first fully connected layer and the second fully connected layer are generated according to the following formula: w t ; , In the formula, F t (m) is F t The m-th dimension feature, Let b be the weight of the m-th dimension in the first fully connected layer. fc1 For the bias of the first fully connected layer, Let b be the weight of the j-th dimension in the second fully connected layer. fc2 The bias is set for the second fully connected layer, SiLU(∙) is the SiLU activation function, and Softplus(∙) is the Softplus activation function.

[0012] As a preferred option, the daily weight optimization unit judges x. t Whether a value is an outlier includes steps Sa1 to Sa2; Sa1, calculate the mean μ and standard deviation σ of the actual working hours from t-L+1 to day t, and the working hour deviation g on day t. t The increase in logistics costs on day t. t ; , , Sa2, if or Then x t This is an outlier.

[0013] In this invention: 1. Regarding the construction of the dataset in S1: When constructing the dataset, we first define some quantitative and qualitative indicators, and then perform feature engineering based on these indicators to construct samples for each day. Quantitative indicators are those that can be quantified and directly obtained from design or historical records, such as total building area (m²), prefabrication rate (%), total number of hoisting operations, transportation distance (km), number of floors, number of building types (types of prefabricated components used), number of days under construction, and for each day of construction, the planned construction area, actual construction area, daily labor cost, machinery cost, logistics cost, etc. Qualitative indicators include construction management level, worker proficiency, and site environmental conditions, which are graded and assigned values ​​according to actual needs. Then, based on feature engineering, we construct samples for each day; for example, the sample for day t is x. t x t =(x t,1 ,x t,2 ,⋯,x t,10 After obtaining the samples for each day, the feature matrix corresponding to each day is constructed. The feature matrix X(t) for day t is formed by concatenating the samples from day t-L+1 to day t. If t < L+1, then x... t-L+1 ~x t-1 The samples for each day are set as zero vectors. Samples constructed using this method contain the basic feature x. t,1 ~x t,3 Time-domain derived features x t,4 ~x t,6 Frequency domain features x t,7 Cross-feature x t,8 ~x t,10This system features comprehensive coverage of multi-dimensional cost influencing factors. Basic features, time-domain derived features, frequency-domain features, and cross-features each have their own focus and complement each other, fully covering the core dimensions affecting prefabricated building construction costs. Basic features anchor inherent project attributes; time-domain derived features accurately capture the temporal changes in working hours, costs, and workload during construction; frequency-domain features uncover the implicit fluctuations in logistics costs; and cross-features integrate qualitative and quantitative indicators to quantitatively represent subjective factors such as construction management and personnel proficiency. Simultaneously, the physical meaning of the features is clear, and all derived features are designed based on actual prefabricated construction procedures, aligning with the dynamic progression of construction, with no redundant or invalid features. Furthermore, the feature dimension design is highly compatible with neural network modeling. Ten-dimensional daily samples can be seamlessly integrated with convolution and pooling operations after network mapping, effectively supporting the cost prediction network in extracting local features and generating dynamic weights. The unified feature quantification rules provide reliable input for IRLS solving, significantly improving the training efficiency and prediction accuracy of the cost prediction model, laying a solid feature foundation for the dynamic and accurate prediction of prefabricated building construction costs.

[0014] 2. Regarding the cost prediction network of S2, it includes a daily weight generation network, a daily weight optimization unit, a weight matrix output unit, a least squares solution unit, and a cost prediction unit.

[0015] (2.1) Regarding the daily weight generation network: it includes an input layer, a one-dimensional convolutional layer, a layer normalization layer, a max pooling layer, a feature extraction layer, a feature concatenation layer, a first fully connected layer, and a second fully connected layer. t The input layer performs dimensionality adjustment, a one-dimensional convolutional layer extracts local features, layer normalization improves model stability, and then max pooling downsamples the normalized features to retain key information, resulting in pooled features. Then, the feature extraction layer extracts... The mean, variance, and peak value are used as supplementary statistical features and then concatenated by the feature splicing layer to... The 19×1 splicing feature F is obtained. t The purpose is to supplement statistical features and enhance F. t The first fully connected layer has an input dimension of 16 + 3 = 19 dimensions and an output dimension of 64 dimensions. Its core function is to non-linearly map features and enhance expression. The second fully connected layer has an input dimension of 64 dimensions and an output dimension of 1 dimension. Its core function is to output the dynamic weight of a single sample, that is, the weight w on day t. t The weights generated by the daily weight generation network are different for each day.

[0016] (2.2) Regarding the daily weight optimization unit: it is based on "outlier suppression" for w t Optimize and generate the corresponding optimization weights w't Its essence is based on conditions. or Determine x t If the value is an outlier, then trigger the "outlier suppression" weight decay operation via w'. t =w t ×0.3 to w t Perform attenuation.

[0017] (2.3) Regarding the weight matrix output unit, the purpose is to use the optimized weights w' from day t-L+1 to day t. t-L+1 ~w' t Generate the weight matrix W(t+∆T) for day t+∆T, where W(t+∆T)=diag(w' t-L+1 ,w' t-L+2 ,…,w' t This can be expressed in matrix form as follows: .

[0018] (2.4) Regarding the least squares solution unit and the cost prediction unit, W(t+∆T) is substituted into them for least squares solution to calculate β. t+∆T , and then β t+∆T Substitute the values ​​into the cost forecasting unit to solve for the predicted cost on day t+∆T. , and y t+∆T The difference between them is the residual on day t+∆T.

[0019] (2.5) Regarding the generation and optimization of weights, this invention can be divided into three stages: Phase 1: Since the overall parameter adjustment of the "cost prediction network" is based on the loss function Loss, and Loss is related to the residuals, such as the residual on day i = The residual on day i+∆T = Therefore, through a daily weight generation network, samples with large residuals (which may be outliers) are assigned smaller weights, while samples with small residuals are assigned larger weights. Hence, w t The ability to generate different weights based on different samples is the core manifestation of "dynamic weight generation," which replaces the assumption in the traditional least squares method that all samples have a weight of 1.

[0020] Phase 2: Dynamically optimize weights using a daily weight optimization unit. This unit determines whether each day's sample is an outlier based on preset conditions; if so, it optimizes the weights using w'. t =w t The weight is reduced by ×0.3, which can further reduce the interference of outliers on the subsequent coefficient calculation.

[0021] Phase 3: Iterative optimization during training. The residuals calculated in this round are backpropagated through the loss function to optimize the parameters of the cost prediction network, so that the residuals in the next iteration gradually decrease. Each iteration here forms the following sequence: "Previous round residual → dynamic weights W(t+∆T) → feature coefficient vector β". t+∆T →The iterative closed loop of “residual in this round” is continuously optimized until the residual converges to the minimum, ensuring that the model prediction accuracy is continuously improved until convergence, thus obtaining the cost prediction model.

[0022] Compared with the prior art, the advantages of the present invention are as follows: (1) Reasonable sample construction. When constructing the sample for each day, quantitative and qualitative indicators are fully considered, and 10-dimensional features are generated for each day using feature engineering to form the sample for each day. These 10-dimensional features not only include basic features, but also time-domain derived features, frequency-domain features, and cross features. Each dimension of features has its own emphasis and is deeply integrated, comprehensively covering the core influencing factors of prefabricated building construction costs. This effectively makes up for the shortcomings of existing technologies, such as single sample features, one-sided dimensions, and only focusing on a single construction link or static indicators, which cannot adapt to the dynamic changes in construction. Among them, basic features anchor the inherent attributes and core construction parameters of the project, providing a basic benchmark for cost prediction; time-domain derived features capture short-term dynamic fluctuations such as the stability of working hours, changes in the amount of work completed, and unit area cost, accurately adapting to the dynamics of the construction process; frequency-domain features explore the implicit fluctuation patterns of logistics costs, which can capture the potential impact of sudden factors such as supply chain fluctuations in advance; cross features integrate qualitative and quantitative indicators, quantifying subjective influencing factors such as construction management level, personnel proficiency, and on-site environmental conditions, effectively adapting to the construction needs of multi-link collaboration. Meanwhile, the sample construction strictly follows the sliding window design logic, with daily features dynamically updated according to the construction progress. This ensures the timeliness and authenticity of the features while avoiding the introduction of redundant and invalid features. The physical meaning of the features is clear, and the quantification rules are unified, which can seamlessly connect with the modeling needs of the subsequent cost prediction network. This provides reliable data support for the model to accurately capture the nonlinear changes in construction costs, greatly improving the usability and relevance of the sample data. Compared with static samples and single-dimensional samples in existing technologies, this is more in line with the actual scenario of prefabricated building construction, laying a solid foundation for subsequent dynamic weight generation, IRLS solution, and high-precision cost prediction.

[0023] (2) Dynamic Weight Adaptation: Traditional static weight prediction methods struggle to accurately capture the nonlinear cost fluctuations caused by component standardization differences, multi-trade collaboration efficiency fluctuations, and unforeseen factors such as weather delays and supply chain disruptions during prefabricated building construction. This invention designs a cost prediction network that dynamically generates and optimizes weights, and iteratively trains them to adaptively adjust the contribution of different construction data. This effectively identifies and suppresses noise interference from unforeseen factors (such as work stoppages due to heavy rain and supply chain disruptions), enabling the model to remain stable in complex and ever-changing construction environments and improving prediction stability.

[0024] (3) Local Data Prediction. This invention breaks through the dependence of traditional cost prediction methods on complete project cycle data. It only requires data from local construction stages, such as the hoisting of prefabricated components and the assembly of floors, to achieve high-precision cost prediction. By extracting the time series features of segments, such as the fluctuation of working hours within L days, a mapping relationship between segments and overall costs is established, supporting dynamic cost control and decision optimization in the mid-term of the project, which greatly improves the flexibility and practicality of prefabricated building cost management.

[0025] (4) Enhanced practicality and reliability of the model. This invention constructs a prediction model adapted to engineering realities by integrating a dynamic weighting mechanism and a neural network algorithm. Its advantage lies in its ability to intelligently distinguish between normal construction fluctuations and abnormal interferences, achieving stable and reliable prediction outputs without relying on complex parameter tuning or a large amount of complete historical data. The model has strong anti-interference capabilities, maintaining a reasonable prediction trend even under conditions of poor data quality or sudden anomalies, providing a reliable basis for project management decisions and significantly reducing control risks caused by cost misjudgments.

[0026] (5) Industry Applicability. This technology can be applied to the field of prefabricated building construction, including precast concrete (PC) structures and steel structure prefabricated projects such as residential buildings, public buildings, and industrial plants. It can also be applied to large-scale infrastructure construction, such as modular construction scenarios for bridges, tunnels, and subway stations. Through dynamic integration of parameters such as steel prices, welding processes, and high-altitude operation efficiency, it enables specialized cost prediction and monitoring. In modular building and special structural engineering, such as data centers, hospital modules, and emergency buildings, this technology can also predict key cost drivers such as IT equipment procurement, module assembly, and system energy consumption. Furthermore, cost prediction for decoration and finishing projects, especially for batch high-end decoration projects, by floor and by unit type, as well as cost uncertainty management in overseas projects and EPC general contracting projects due to resource scheduling, logistics, policies, and other volatile factors, are also important application directions of this patented technology. The flexibility and strong adaptability of this technology enable its cross-industry promotion, providing reliable dynamic cost prediction support for various modular and prefabricated engineering projects. Attached Figure Description

[0027] Figure 1 This is a flowchart of the present invention; Figure 2 This is a diagram of the cost prediction network structure. Figure 3 Generate a network structure diagram for the daily weights. Detailed Implementation

[0028] The present invention will be further described below with reference to the embodiments and accompanying drawings.

[0029] Example 1: See Figures 1-3 A dynamic prediction method for prefabricated construction costs, integrating convolutional neural networks and IRLS, is proposed, where T1 days have been completed, T2 days have not been completed, and the total construction period is T. 总 =T1+T2, including the following steps; S1, Construct dataset D, including steps S11~S13; S11 divides the construction management level, the skill level of technical workers, and the on-site environmental conditions into several levels from high to low, and assigns a value to each level in the range [0,1]. The higher the level, the larger the value assigned. S12, using a sliding window of length L, acquire data from day t-L+1 to day t, and construct the sample x for day t. t 10-dimensional features x t,1 ~x t,10 x t =(x t,1 ,x t,2 ,⋯,x t,10 ), L≤t≤T1-∆T, where ∆T is the prediction time difference; x t,1 x represents the total building area. t,2 To determine the prefabrication rate, x t,3 Total number of hoisting operations; x t,4 The standard deviation σ of the working hours consumed in L days inside the sliding window. t , x t,5 Let be the rate of change of the amount completed on day t. ; x t,6 Let be the cost density on day t. ; x t,7 To determine the dominant frequency energy characteristics on day t, the logistics costs for L days within the sliding window are sequenced and transformed using FFT to generate L frequency components arranged in descending order as X(1)~X(L). The results are then calculated using the formula: X(v) is the v-th frequency component in descending order; x t,8 For management efficiency coefficient, ; x t,9 The efficiency coefficient is the ratio of personnel proficiency. ; x t,10 To design the complexity prefabrication rate interaction item, x t,10 = Value assigned to the on-site environmental condition level on day t × Design prefabrication rate; S13, starting from day L of construction, sequentially generate feature matrices for all days already completed, forming dataset D, where the feature matrix for day t is X(t). ; S2. Construct a cost prediction network, including a daily weight generation network, a daily weight optimization unit, a weight matrix output unit, a least squares solution unit, and a cost prediction unit. The daily weight generation network is used as input x t Generate the weight w for day t. t ; The daily weight optimization unit is used to determine x based on preset conditions. t Is it an outlier? Generate the optimized weight w' for day t. t If x t If w' is an outlier, then t =w t ×0.3, otherwise w' t =w t ; The weight matrix output unit is used to calculate the optimized weights w' from day t-L+1 to day t. t-L+1 ~w' t Generate the weight matrix W(t+∆T) for day t+∆T, where W(t+∆T)=diag(w' t-L+1 ,w' t-L+2 ,…,w' t ), where diag(∙) is the diag function; The least squares solving unit is used to solve for the characteristic coefficient vector β on day t+∆T according to the following formula. t+∆T ; , Where λ = 1e-5 is the ridge parameter, I is the identity matrix, and T is the transpose operation. ~ β t+∆T The first to the tenth characteristic coefficients; The cost forecasting unit is used to calculate the forecast cost for day t+∆T according to the following formula. ; , In the formula, β0 is the preset baseline cost; S3, construct the loss function Loss for the cost prediction network; , In the formula, i represents one day from day t-L+1 to day t. The optimization weight for day i. , Let be the actual cost and the predicted cost on day i, respectively; α be the L2 regularization strength; and θ be all trainable parameters of the cost prediction network. S4. Train the cost prediction network using dataset D, and adjust θ to minimize the loss until the cost prediction network converges to obtain the cost prediction model. Mark the feature coefficient vector β at this time as the optimal feature coefficient vector β*. S5, for the number of days without construction d, if d-∆T days are within the number of days with construction, then construct a sample of d-∆T and substitute it into the cost prediction unit to calculate the predicted cost for day d using β*.

[0030] In S11, the construction management level is divided into 1 to 4 levels, with corresponding values ​​of 1.0, 0.7, 0.4, and 0.1, respectively; the skill level of the technical workers is divided into 1 to 3 levels, with corresponding values ​​of 1.0, 0.6, and 0.2, respectively; and the on-site environmental conditions are divided into 1 to 4 levels, with corresponding values ​​of 1.0, 0.7, 0.4, and 0.1, respectively.

[0031] The daily weight generation network includes, in sequence, an input layer, a one-dimensional convolutional layer, a layer normalization layer, a max pooling layer, a feature extraction layer, a feature concatenation layer, a first fully connected layer, and a second fully connected layer. The input layer is used to process x t Mapped to 16×3 input features ; The one-dimensional convolutional layer is used to generate convolutional features c according to the following formula. t ; , In the formula, W conv (k), b conv (k) represents the weights and biases of the convolution kernel k, respectively, and c t ∈R 16×3 ; The normalization layer and max pooling layer sequentially perform normalization and max pooling operations on c(t), outputting a 16×1 pooling feature. ; The feature extraction layer is used for calculation. Mean, variance, and peak value of pooling features; The feature splicing layer is used to... The concatenation feature F, which is concatenated with the mean, variance, and peak value, is 19×1. t ; The first fully connected layer and the second fully connected layer respectively use the SiLU activation function and the Softplus activation function to convert F t Adjust to 64 dimensions and 1 dimension, output the weight w for day t. t .

[0032] The first fully connected layer and the second fully connected layer are generated according to the following formula: w t ; , In the formula, F t (m) is F t The m-th dimension feature, Let b be the weight of the m-th dimension in the first fully connected layer. fc1 For the bias of the first fully connected layer, Let b be the weight of the j-th dimension in the second fully connected layer. fc2 The bias is set for the second fully connected layer, SiLU(∙) is the SiLU activation function, and Softplus(∙) is the Softplus activation function.

[0033] Daily weight optimization unit judgment x t Whether a value is an outlier includes steps Sa1 to Sa2; Sa1, calculate the mean μ and standard deviation σ of the actual working hours from t-L+1 to day t, and the working hour deviation g on day t. t The increase in logistics costs on day t. t ; , , Sa2, if or Then x t This is an outlier.

[0034] Example 2: See Figures 1-3 Based on Example 1, in S11, a specific method is given to divide the construction management level, the skill level of technical workers, and the site environmental conditions into several levels from high to low, and to assign a value to each level in the range [0,1]. See Tables 1 to 3.

[0035] Table 1. Classification and Value Table of Construction Management Level grade describe Assignment Typical characteristics Level 1 excellent 1.0 A sound project management system ensures schedule deviations of less than 5%, precise resource allocation, and a response time of less than 2 hours to unforeseen issues. Level 2 good 0.7 The process is standardized but occasional delays occur, with schedule deviations of 5%-10%, and resource utilization ≥80%. Level 3 medium 0.4 There are loopholes in the management process, the schedule deviation is 10%-15%, and the frequency of resource conflicts is high. Level 4 Poor 0.1 The plan was poorly executed, with schedule deviations exceeding 15% and significant resource waste. Table 2. Skill Level Grading and Assignment Table for Technicians grade describe Assignment Typical characteristics Level 1 skilled 1.0 Possessing advanced skill certification, with an operational error rate of <3%, and achieving 120% of standard working hours in hoisting / installation efficiency. Level 2 generally 0.6 Mastering basic operations, with an error rate of 3% to 8%, and achieving 80% to 100% efficiency of standard working hours. Level 3 rusty 0.2 Frequent guidance required, error rate >8%, efficiency less than 80% of standard working hours. Table 3. Classification and Value Assignment Table of On-site Environmental Conditions describe Assignment Typical characteristics Level 1 excellent 1.0 A sound project management system ensures schedule deviations of less than 5%, precise resource allocation, and a response time of less than 2 hours to unforeseen issues. Level 2 good 0.7 The process is standardized but occasional delays occur, with schedule deviations of 5%-10%, and resource utilization ≥80%. Level 3 medium 0.4 There are loopholes in the management process, the schedule deviation is 10%-15%, and the frequency of resource conflicts is high. Level 4 Poor 0.1 The plan was poorly executed, with schedule deviations exceeding 15% and significant resource waste. The above classification, assignment, and typical features can all be adjusted according to the actual situation.

[0036] Regarding the L2 regularization strength α, its value ranges from 0.001 to 0.05, and can be discussed in different cases as follows: For conventional prefabricated building projects: with a building area of ​​10,000 to 50,000 square meters and a construction period of 100 to 300 days, the default value of α is 0.01.

[0037] For large-scale prefabricated building projects: with a building area of ​​>50,000㎡, a construction period of >300 days, and sufficient sample size, α can be reduced to 0.001~0.005 to reduce regularization constraints and fully learn data patterns.

[0038] For small-scale prefabricated building projects: building area < 10,000㎡, construction period < 100 days, and sample size is small, α can be increased to 0.02~0.05 to enhance the overfitting suppression effect, adapt to the characteristics of datasets of different project sizes, and balance the model fitting accuracy and generalization ability.

[0039] Regarding deployment: To achieve efficient and stable operation of the model in the harsh network environment and limited hardware resources of construction sites, this invention adopts a lightweight deployment scheme. First, the trained cost prediction model is deeply optimized based on the NVIDIA TensorRT inference engine: through FP16 half-precision quantization technology, the model weights are compressed from 32-bit floating-point numbers to 16-bit, reducing the model size and memory usage by almost half without sacrificing accuracy; at the same time, adjacent layers such as convolutions and activation functions in the network are merged into a single computational kernel, significantly reducing kernel call overhead and the number of GPU memory read / write operations. The optimized model can be deployed on edge computing terminals at the construction site to achieve localized real-time inference. This terminal directly connects to the sensor and daily manual data at the construction site, without relying on the cloud network, and can dynamically update the cost prediction results daily or according to the construction stage, providing project managers with immediate decision support, truly realizing the "decentralization" and "real-time" of prediction capabilities.

[0040] Example 3: See Figures 1-3 This paper presents a dynamic cost prediction method for prefabricated construction by fusing convolutional neural networks and IRLS, used for daily cost prediction during the construction phase of prefabricated buildings. In this embodiment, the sliding window length L=7, constructing daily 10-dimensional feature samples, and setting the L2 regularization strength α to a range of 0.001~0.05, with a default value of 0.01 for regular projects. The training process uses a modified weighted squared loss function, and the optimization algorithm is Adam, with an initial learning rate of 0.001, dynamically adjusted during training to achieve accurate and efficient prediction of construction costs.

[0041] The experimental design is as follows: 1. Dataset: A typical prefabricated residential community project was selected: total construction area of ​​32,000 square meters, construction period of 240 days, 6 types of prefabricated components, and a total of 1200 hoisting operations. Real-time daily construction data was collected throughout the entire construction process, including core indicators such as total construction area, prefabrication rate, total hoisting operations, daily construction area, daily labor cost, daily machinery cost, daily logistics cost, daily management level, daily skill level of assigned technicians, and daily site environmental condition level. A dataset of 240 daily samples was constructed, divided into training and testing sets in a 7:3 ratio. The data was standardized and used for model training and testing.

[0042] 2. Experimental environment: The hardware environment consists of an NVIDIA RTX 3090 GPU and an Intel Core i9-12900K CPU, while the software environment consists of PyTorch 1.10, Python 3.9, and 64GB of memory.

[0043] 3. Experimental Groups: The experimental group uses the prediction method based on neural network IRLS proposed in this invention; the control group uses two commonly used cost prediction methods in the industry (Control Group 1: traditional BP neural network cost prediction method; Control Group 2: time series cost prediction method based on LSTM). The three groups of models use the same training set, test set and experimental environment to ensure the fairness of the comparison.

[0044] 4. Performance indicators: Commonly used quantitative indicators in the prefabricated building cost prediction industry, including accuracy and efficiency indicators, are selected.

[0045] Accuracy metrics include Mean Absolute Error (MAE), measured in dollars; Mean Relative Error (MAPE), measured as a percentage; and Root Mean Square Error (RMSE), measured in dollars. The lower the accuracy metric value, the higher the prediction accuracy.

[0046] Efficiency metrics include model training time, measured in minutes; and single prediction time, measured in seconds. The shorter the time, the higher the prediction efficiency.

[0047] 5. The experimental results are shown in Table 4: Table 4. Comparison of Experimental Results Model Name MAE / yuan MAPE / % MSE / yuan Model training time / minute Single prediction time / second This invention 1080 3.9 1320 48 0.06 Control group 1 1350 5.2 1680 56 0.09 Control group 2 1220 4.5 1490 60 0.10 6. Data Analysis: Experimental results show that this invention outperforms the two control group models in both accuracy and prediction efficiency. Its advantages are reasonable and relevant to practical application scenarios. The core advantages are as follows: (1) In terms of accuracy: The MAE, MAPE and RMSE of the model of this invention are 1080 yuan, 3.9% and 1320 yuan respectively, which are 19.3%, 25.0% and 21.4% lower than those of control group 1, and 11.5%, 13.3% and 11.4% lower than those of control group 2. This shows that the present invention can capture the nonlinear and dynamic fluctuation of the construction cost of prefabricated buildings more accurately through multi-dimensional feature fusion, CostNet dynamic weight generation and IRLS algorithm optimization, effectively suppress sudden noise interference such as rainstorm shutdown and supply chain fluctuation, and the prediction accuracy is significantly improved compared with the existing technology.

[0048] (2) Efficiency: The training time of the model of the present invention is 48 minutes, which is 14.3% shorter than that of control group 1 and 20.0% shorter than that of control group 2; the single prediction time is 0.06 seconds, which is 33.3% shorter than that of control group 1 and 40.0% shorter than that of control group 2. Thanks to the optimized feature design and model architecture of the present invention, the 10-dimensional feature is adapted to the neural network modeling, reducing redundant data processing links, and significantly improving the model training and prediction efficiency while ensuring prediction accuracy.

[0049] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for dynamic prediction of prefabricated construction costs by fusing convolutional neural networks and IRLS, where T1 days have been completed, T2 days have not been completed, and the total construction period is T. 总 =T1+T2, characterized in that, Includes the following steps; S1, Construct dataset D, including steps S11~S13; S11 divides the construction management level, the skill level of technical workers, and the on-site environmental conditions into several levels from high to low, and assigns a value to each level in the range [0,1]. The higher the level, the larger the value assigned. S12, using a sliding window of length L, acquire data from day t-L+1 to day t, and construct the sample x for day t. t 10-dimensional features x t,1 ~x t,10 x t =(x t,1 ,x t,2 ,⋯,x t,10 ), L≤t≤T1-∆T, where ∆T is the prediction time difference; x t,1 x represents the total building area. t,2 To determine the prefabrication rate, x t,3 x represents the total number of lifting operations. t,4 The standard deviation σ of the working hours consumed in L days inside the sliding window. t ; x t,5 Let be the rate of change of the amount completed on day t. ; x t,6 Let be the cost density on day t. ; x t,7 To determine the dominant frequency energy characteristics on day t, the logistics costs for L days within the sliding window are sequenced and transformed using FFT to generate L frequency components arranged in descending order as X(1)~X(L). The results are then calculated using the formula: X(v) is the v-th frequency component in descending order; x t,8 Management efficiency coefficient ; x t,9 The efficiency coefficient is the ratio of personnel proficiency. ; x t,10 To design the complexity prefabrication rate interaction item, x t,10 = Value assigned to the on-site environmental condition level on day t × Design prefabrication rate; S13, starting from day L of construction, sequentially generate feature matrices for all days already completed, forming dataset D, where the feature matrix for day t is X(t). ; S2, construct a cost prediction network, including a daily weight generation network, a daily weight optimization unit, a weight matrix output unit, a least squares solution unit, and a cost prediction unit; The daily weight generation network is used as input x t Generate the weight w for day t. t ; The daily weight optimization unit is used to determine x based on preset conditions. t Is it an outlier? Generate the optimized weight w' for day t. t If x t If w' is an outlier, then t =w t ×0.3, otherwise w' t =w t ; The weight matrix output unit is used to calculate the optimized weights w' from day t-L+1 to day t. t-L+1 ~w' t Generate the weight matrix W(t+∆T) for day t+∆T, where W(t+∆T)=diag(w' t-L+1 ,w' t-L+2 ,…,w' t ), where diag(∙) is the diag function; The least squares solving unit is used to solve for the characteristic coefficient vector β on day t+∆T according to the following formula. t+∆T ; ， Where λ = 1e-5 is the ridge parameter, I is the identity matrix, and T is the transpose operation. ~ β t+∆T The first to the tenth characteristic coefficients; The cost forecasting unit is used to calculate the forecast cost for day t+∆T according to the following formula. ; ， In the formula, β0 is the preset baseline cost; S3, construct the loss function Loss for the cost prediction network; ， In the formula, i represents one day from day t-L+1 to day t. The optimization weight for day i. , Let be the actual cost and the predicted cost on day i, respectively; α be the L2 regularization strength; and θ be all trainable parameters of the cost prediction network. S4. Train the cost prediction network using dataset D, and adjust θ to minimize the loss until the cost prediction network converges to obtain the cost prediction model. Mark the feature coefficient vector β at this time as the optimal feature coefficient vector β*. S5, for the number of days without construction d, if d-∆T days are within the number of days with construction, then construct a sample of d-∆T and substitute it into the cost prediction unit to calculate the predicted cost for day d using β*.

2. The method for dynamic prediction of prefabricated construction costs by fusing convolutional neural networks and IRLS according to claim 1, characterized in that, In S11, the construction management level is divided into 1 to 4 levels, with corresponding values ​​of 1.0, 0.7, 0.4, and 0.1, respectively; the skill level of the technical workers is divided into 1 to 3 levels, with corresponding values ​​of 1.0, 0.6, and 0.2, respectively; and the on-site environmental conditions are divided into 1 to 4 levels, with corresponding values ​​of 1.0, 0.7, 0.4, and 0.1, respectively.

3. The method for dynamic prediction of prefabricated construction costs by fusing convolutional neural networks and IRLS according to claim 1, characterized in that, The daily weight generation network includes, in sequence, an input layer, a one-dimensional convolutional layer, a layer normalization layer, a max pooling layer, a feature extraction layer, a feature concatenation layer, a first fully connected layer, and a second fully connected layer. The input layer is used to process x t Mapped to 16×3 input features ; The one-dimensional convolutional layer is used to generate convolutional features c according to the following formula. t ; ， In the formula, W conv (k), b conv (k) represents the weights and biases of the convolution kernel k, respectively, and c t ∈R 16×3 ; The normalization layer and max pooling layer sequentially perform normalization and max pooling operations on c(t), outputting a 16×1 pooling feature. ; The feature extraction layer is used for calculation Mean, variance, and peak value of pooling features; The feature splicing layer is used to... The concatenation feature F, which is concatenated with the mean, variance, and peak value, is 19×1. t ; The first fully connected layer and the second fully connected layer respectively use the SiLU activation function and the Softplus activation function to convert F t Adjust to 64 dimensions and 1 dimension, output the weight w for day t. t .

4. The method for dynamic prediction of prefabricated construction costs by fusing convolutional neural networks and IRLS according to claim 3, characterized in that, The first fully connected layer and the second fully connected layer are generated according to the following formula: w t ; ， In the formula, F t (m) is F t The m-th dimension feature, Let b be the weight of the m-th dimension in the first fully connected layer. fc1 For the bias of the first fully connected layer, Let b be the weight of the j-th dimension in the second fully connected layer. fc2 The bias of the second fully connected layer is defined by SiLU(∙), which is the SiLU activation function, and Softplus(∙) is the Softplus activation function.

5. The method for dynamic prediction of prefabricated construction costs by fusing convolutional neural networks and IRLS according to claim 1, characterized in that, Daily weight optimization unit judgment x t Whether a value is an outlier includes steps Sa1 to Sa2; Sa1, calculate the mean μ and standard deviation σ of the actual working hours from t-L+1 to day t, and the working hour deviation g on day t. t The increase in logistics costs on day t. t ; ， ， Sa2, if or Then x t This is an outlier.

Images