Engineering change cost prediction and management method based on deep learning

Abstract

The invention discloses an engineering change cost prediction and management method based on deep learning. The method comprises the following steps: S1, constructing a data set; s2, performing data preprocessing on the data set; s3, training a basic deep learning model, and determining an initial model architecture and parameters; s4, continuously collecting real-time change data and actual cost data as dynamic feedback input to obtain real-time feedback data; s5, based on the real-time feedback data, triggering a progressive learning process; s6, performing hybrid optimization in combination with an adaptive differential evolution algorithm and fuzzy logic control, and finally automatically selecting an optimal network layer number, an optimal node number and an optimal activation function; s7, dynamically predicting future change cost of the project, and continuously optimizing a prediction result along with continuous updating of project data; and S8, according to the optimized prediction result, the system automatically generates project management suggestions. According to the method, deep learning and an adaptive optimization method are combined, and dynamic engineering change cost prediction and management are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of engineering project management, and in particular to an engineering change cost prediction and management method based on deep learning. Background Art

[0002] Engineering changes are a common and inevitable phenomenon in modern engineering project management. As the scale and complexity of engineering projects increase, the frequency and magnitude of engineering changes are also increasing. Engineering changes may be caused by a variety of reasons, including design modifications, resource adjustments, unforeseen environmental factors, etc. These changes directly affect the cost, schedule and quality of the project. Therefore, how to effectively predict and manage the cost of engineering changes has become a critical challenge in the field of project management.

[0003] Traditional methods for predicting engineering change costs mainly rely on expert experience and analysis of historical data. These methods usually use statistical models, linear regression analysis, or simple rule-based systems to predict the impact of changes. However, these methods often show certain limitations when faced with complex and dynamically changing project environments. First, traditional methods often assume that the engineering project environment is relatively static and linear, and cannot fully consider the nonlinear changes and dynamic feedback that may occur during the project process. Second, these methods usually rely on past historical data and the subjective judgment of experts, lack sensitivity to real-time data, and have difficulty responding to the impact of engineering changes in a timely manner. In addition, when dealing with complex and multi-dimensional project data, traditional cost forecasting models often find it difficult to capture the deep patterns and potential relationships in the data, resulting in insufficient prediction accuracy, which in turn affects the overall management decision of the project.

[0004] With the rapid development of artificial intelligence technology, deep learning, as a powerful data processing and pattern recognition tool, has gradually been introduced into the field of engineering project management. Deep learning models, especially convolutional neural networks, long short-term memory networks, and transformer models, can process large amounts of unstructured data and time series data, automatically extract complex features from the data, and make accurate predictions. This makes deep learning have great potential in engineering change cost prediction. However, most existing deep learning applications are still in their infancy, usually simply inputting historical data into the model for training, and cannot fully utilize the real-time data generated during the project process for dynamic adjustment. Existing deep learning models usually use fixed network architectures and parameters, which lack flexibility and adaptability when facing dynamic changes in projects, resulting in the model's prediction performance being difficult to achieve the expected results in practical applications.

[0005] In order to improve the practical application effect of deep learning models in engineering change cost prediction, some studies have begun to explore adaptive optimization methods of models, such as adaptive learning rate adjustment and feedback-based dynamic model update. However, most of these methods focus on single-aspect optimization and lack systematic integration and innovation. Especially when faced with complex multi-objective optimization problems, existing technologies still seem to be unable to cope with them. In addition, traditional deep learning models usually use fixed weight coefficients or simple weighted summation methods when dealing with multi-objective optimization problems, which cannot fully consider the complex relationships and mutual influences between various objectives. This processing method may cause the model to sacrifice the performance of other objectives when optimizing a certain objective, thereby affecting the reliability and stability of the overall prediction results.

[0006] Therefore, how to provide an engineering change cost prediction and management method based on deep learning is an urgent problem that technicians in this field need to solve. Summary of the invention

[0007] One purpose of the present invention is to propose a method for predicting and managing engineering change costs based on deep learning. The present invention combines a deep learning model with an adaptive differential evolution algorithm and fuzzy logic control technology to propose a dynamic method for predicting and managing engineering change costs. Through real-time data feedback and progressive learning, the model can automatically optimize the number of network layers, the number of nodes, and the activation function at different stages of the project to ensure high accuracy and strong adaptability of the prediction results. In addition, combined with adversarial training and adaptive dynamic prediction loss functions, the present invention performs well in dealing with complex environments and uncertain factors, and realizes real-time management and precise cost control of engineering projects.

[0008] A method for predicting and managing engineering change costs based on deep learning according to an embodiment of the present invention includes the following steps:

[0009] S1. At the project initiation stage, collect historical project data and initial data of the current project and build a data set;

[0010] S2, preprocess the data set, including cleaning, denoising and feature extraction;

[0011] S3. Use the preprocessed data set to train the basic deep learning model and determine the initial deep learning model architecture and parameters;

[0012] S4. At different stages of the project, real-time change data and actual cost data are continuously collected as dynamic feedback input to obtain real-time feedback data;

[0013] S5. Based on real-time feedback data, the progressive learning process is triggered, and the deep learning model is retrained and optimized according to the new feedback data;

[0014] S6. During the deep learning model retraining process, the adaptive differential evolution algorithm and fuzzy logic control are combined for hybrid optimization, and the optimal number of network layers, number of nodes and activation function are automatically selected;

[0015] S7. Use the deep learning model optimized by the hybrid optimization algorithm to dynamically predict the future change costs of the project, and continuously optimize the prediction results as the project data is continuously updated;

[0016] S8. Based on the optimized forecast results, the system automatically generates project management suggestions including resource allocation, schedule adjustment and risk control, and passes them to the project management system.

[0017] Optionally, the S3 specifically includes:

[0018] S31, define the input data set X = {x1, x2, ..., x n}, where X represents the feature matrix after preprocessing, n represents the total number of samples, and each sample includes m features, denoted by x i ={x i1 ,x i2 ,…,x im};

[0019] S32. Initialize the architecture of the deep learning model. The architecture of the deep learning model includes an input layer, a hidden layer, and an output layer. Define the number of nodes in the input layer as m, the number of hidden layers as Le, and the number of nodes in each layer as h. l , where l∈{1,2,…,Le}, the activation function f(x) uses the adaptive regularized ReLU function:

[0020] f(x)=max(0,x)+λ·Ω(θ);

[0021] Among them, λ represents the regularization parameter, and Ω(θ) represents the complexity penalty term of parameter θ:

[0022]

[0023] Among them, α and β represent weight coefficients, represents the L2 regularization term, ∥θ l ∥1 represents the L1 regularization term;

[0024] S33. Define the loss function Loss, which includes the adversarial training loss L pred and the base prediction loss L adv :

[0025] Loss = L pred +γ·L adv ;

[0026] Among them, γ represents the weight coefficient that controls the impact of adversarial training;

[0027]

[0028] Among them, y i represents the actual value of the i-th sample, represents the model prediction value of the i-th sample;

[0029]

[0030] Among them, δ represents the adversarial perturbation, ∈ represents the maximum norm of the perturbation, represents the model prediction output of the i-th sample under the adversarial perturbation δ, and ||δ|| represents the p-norm of δ;

[0031] S34. Use the adaptive learning rate optimization algorithm to train the deep learning model and update the model parameters:

[0032]

[0033] Among them, θ t represents the model parameter vector of the tth iteration, θ t+1 represents the model parameter vector of the t+1th iteration, η t represents the adaptive learning rate, represents the momentum estimate, represents the weighted average variance estimate, ∈ is used to avoid division by zero error, and λ1 represents the weight decay parameter;

[0034] S35. In each iteration, input the training data set X into the deep learning model and calculate the prediction result of the output layer through forward propagation Calculate the error according to the loss function Loss;

[0035] S36. Calculate the gradient of each parameter according to the error through the back propagation algorithm, and update the model parameters according to the adaptive learning rate optimization algorithm until the model converges and the loss function value no longer decreases significantly;

[0036] S37. Determine the initial model architecture and parameters, evaluate the model performance through the validation set, and save the trained model parameters as the initial configuration of the base model.

[0037] Optionally, the S5 specifically includes:

[0038] S51. Define real-time feedback dataset X new ={x n+1 ,x n+2 ,…,x n+q}, where q represents the number of samples, each sample number includes k features, denoted as xi ={x i1 ,x i2 ,…,x ik}, where i∈{n+1,n+2,…,n+q}, k represents the feature dimension;

[0039] S52, the real-time feedback data set X new With the previous historical data set X hist Combined to form a new comprehensive dataset X combined =X hist ∪X new , and triggers a progressive learning process;

[0040] S53, in the new comprehensive dataset X combined On the above, generate adversarial sample set X adv =

[0041] ″′

[0042] {x n+1 ,x n+2 ,…,x n+q}, the adversarial sample is created by new Add a perturbation δ to each sample in i Generate, the perturbation satisfies ||δ i || p ≤∈1:

[0043] x′ i =x i +δ i ;

[0044] Among them, δ i represents the disturbance vector, ∈1 represents the preset disturbance amplitude, ||δ i || p represents the p-norm of the disturbance;

[0045] S54. In the process of deep learning model retraining, the adversarial sample set X adv With the comprehensive dataset X combined Merge and construct an extended training set X extended =X combined ∪X adv , input into the model for training;

[0046] S55, expand the training set X extended Input into the deep learning model for retraining and optimization.

[0047] Optionally, the S6 specifically includes:

[0048] S61. Define a multi-objective optimization function F(φ), where φ represents the structural parameters of the model, including the number of network layers L, the number of nodes in each layer {n1,n2,…,n L} and the activation function parameter combination {α1,α2,…,α L}, the optimization goal is to minimize the prediction error and model complexity at the same time, and consider the interaction of parameters in each layer:

[0049]

[0050] Among them, m l represents the number of nodes in the lth layer, y j represents the actual value of the jth sample, represents the predicted value of the jth sample, n l represents the number of nodes in the lth layer, α l represents the activation function parameter of the lth layer, λ2, λ3 and λ4 represent weight coefficients, which control the influence of prediction error, complexity and activation function respectively;

[0051] S62, initialize the population of the adaptive differential evolution algorithm, define the population size as Q, and each individual H q Represents a set of model parameter combinations φ q ={L q ,{n q1 ,n q2 ,…,n qL},{α q1 ,α q2 ,…,α qL}}, the population is initially randomly generated, and the population size is Q = {H1,H2,…,H Q};

[0052] S63. In each generation, for each individual H in the population q , through the adaptive combination of multiple mutation strategies to generate mutant individuals V q , mutation operations include differential mutation, random mutation and Gaussian mutation:

[0053] L q '=L q +ΔL q , where ΔL q ~U(-2,2);

[0054] n ql '=n ql +G l ·(n p1l -n p2l}+Z(0,1), where l∈{1,2,,L q};

[0055] α ql '=α ql +Δα ql , where Δα ql ~Z(0,0.05};

[0056] Among them, L q ' represents the number of grid layers after mutation, ΔL q Indicates the variation of the number of layers, n ql ' represents the number of nodes in the lth layer after mutation, G l represents the differential variation factor, Z(0,1) represents the standard normal distribution noise, α ql ' represents the activation function parameters after mutation, Δα ql Represents the normal distribution variation of the activation function parameters;

[0057] S64, using a multi-level fuzzy logic control system to correct the mutant individual, the fuzzy logic controller receives the fitness value F(V q ) and the fitness of the current individual F(H q ) as input and output the modified variant individual V′ q , the fuzzy rules include:

[0058] If F(V q ) <F(H q ), then the mutation effect is enhanced;

[0059] If F(V q )≈F(H q ), then retain or fine-tune the variation effect;

[0060] If F(V q )>F(H q ), then the mutation effect is weakened or cancelled;

[0061] S65, perform a crossover operation, and convert the corrected variant individual V′ q With the current individual H q Perform crossover to generate candidate individuals U q :

[0062] U q =(u q1 ,u q2 ,…,u qn );

[0063]

[0064] Among them, v′ qj Represents the variant individual V′ q The jth parameter of qj Indicates the current individual Hq The jth parameter of j represents uniformly distributed random numbers, C r represents the crossover probability;

[0065] S66, according to the fitness value of the multi-objective optimization function F(φ), selecting individuals with better fitness to enter the next generation until a preset termination condition is met, including reaching a maximum number of iterations or the objective function converges;

[0066] S67. In the process of adaptive differential evolution, the fuzzy logic controller adaptively adjusts the mutation factor and crossover probability to ultimately determine the optimal model architecture parameter combination. And used for optimal configuration of deep learning models.

[0067] Optionally, the S7 specifically includes:

[0068] S71, optimal model architecture parameter combination φ obtained based on S67 opt , build prediction model M opt , and use the current training data set X train ={x1,x2,…,x T}Train the model to obtain the initial prediction results;

[0069] S72. Define the adaptive dynamic prediction loss function L dyn (ψ), this loss function is used to measure the prediction error of the model at different time points and can be automatically adjusted according to data changes:

[0070]

[0071] Among them, ψ represents the current parameters of the model, T represents the number of time steps, and y t represents the actual value at time point t, represents the predicted value at time point t, represents the predicted value at the t-1th time point, Δt represents the time interval, σ t represents the volatility of the predicted value at the current time step, a and b represent the weight coefficients of the loss function, and τ represents the hyperparameter that controls the magnitude of the prediction adjustment;

[0072] S73, retrain the model based on the real-time updated project data, and adjust the current parameter ψ of the model to minimize L dyn (ψ);

[0073] S74. Each time new data arrives, by adjusting the volatility term σ in the loss function t and the time-dependent term Δt, which adapts to the dynamic changes of data;

[0074] S75. Use the updated model to generate change cost forecasts for future time steps And continue to collect data and update models throughout the project life cycle;

[0075] S76. Output the final prediction result

[0076] The beneficial effects of the present invention are:

[0077] First of all, the engineering change cost prediction and management method based on deep learning proposed in the present invention significantly improves the accuracy and real-time performance of change cost prediction in engineering project management, and solves the limitations of existing technologies in complex dynamic environments. By introducing a progressive learning mechanism, the present invention can dynamically update the deep learning model according to real-time feedback data at different stages of the project, so that the model can continuously adapt to changes in the project environment and always maintain a high prediction accuracy. Compared with traditional static prediction methods, the present invention provides a more flexible and efficient dynamic prediction solution.

[0078] In addition, by combining the adaptive differential evolution algorithm and fuzzy logic control technology, the present invention can automatically optimize the network architecture and parameter settings of the deep learning model. The adaptive differential evolution algorithm automatically selects the optimal number of network layers, number of nodes and activation function through the adaptive combination of multiple mutation strategies, so that the model can be adaptively adjusted at different project stages to cope with changes in different data characteristics. The fuzzy logic control system ensures the efficiency and stability of the optimization process by dynamically adjusting the mutation factor and crossover probability, thereby improving the overall performance of the model.

[0079] By introducing adversarial training and adaptive dynamic prediction loss function, the present invention also enhances the robustness of the model to complex environments and uncertain factors. The adversarial sample set generated by adversarial training can effectively improve the model's anti-interference ability when facing abnormal or extreme data; and the design of the adaptive dynamic prediction loss function, by introducing time dependence and volatility control, enables the model to more accurately reflect the dynamic changes of data when making future cost forecasts. The integration of these technologies enables the present invention to not only improve the accuracy of predictions, but also enhance the support capabilities for management decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0081] Figure 1 A flowchart of a method for predicting and managing engineering change costs based on deep learning proposed by the present invention;

[0082] Figure 2 This is a schematic diagram of the incremental learning and deep learning model optimization process of the engineering change cost prediction and management method based on deep learning proposed in the present invention. DETAILED DESCRIPTION

[0083] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0084] refer to Figure 1 and Figure 2 , a method for predicting and managing engineering change costs based on deep learning, comprising the following steps:

[0085] S1. At the project initiation stage, collect historical project data and initial data of the current project and build a data set;

[0086] S2, preprocess the data set, including cleaning, denoising and feature extraction;

[0087] S3. Use the preprocessed data set to train the basic deep learning model and determine the initial deep learning model architecture and parameters;

[0088] S4. At different stages of the project, real-time change data and actual cost data are continuously collected as dynamic feedback input to obtain real-time feedback data;

[0089] S5. Based on real-time feedback data, the progressive learning process is triggered, and the deep learning model is retrained and optimized according to the new feedback data;

[0090] S6. During the deep learning model retraining process, the adaptive differential evolution algorithm and fuzzy logic control are combined for hybrid optimization, and the optimal number of network layers, number of nodes and activation function are automatically selected;

[0091] S7. Use the deep learning model optimized by the hybrid optimization algorithm to dynamically predict the future change costs of the project, and continuously optimize the prediction results as the project data is continuously updated;

[0092] S8. Based on the optimized forecast results, the system automatically generates project management suggestions including resource allocation, schedule adjustment and risk control, and passes them to the project management system.

[0093] In this implementation, S3 specifically includes:

[0094] S31, define the input data set X = {x1, x2, ..., x n}, where X represents the feature matrix after preprocessing, n represents the total number of samples, and each sample includes m features, denoted by x i ={x i1 ,x i2 ,…,x im};

[0095] S32. Initialize the architecture of the deep learning model. The architecture of the deep learning model includes an input layer, a hidden layer, and an output layer. Define the number of nodes in the input layer as m, the number of hidden layers as Le, and the number of nodes in each layer as h. l , where l∈{1,2,…,Le}, the activation function f(x) uses the adaptive regularized ReLU function:

[0096] f(x)=max(0,x)+λ·Ω(θ);

[0097] Among them, λ represents the regularization parameter, and Ω(θ) represents the complexity penalty term of parameter θ:

[0098]

[0099] Among them, α and β represent weight coefficients, represents the L2 regularization term, ∥θ l ∥1 represents the L1 regularization term;

[0100] S33. Define the loss function Loss, which includes the adversarial training loss L pred and the base prediction loss L adv :

[0101] Loss = L pred +γ·L adv ;

[0102] Among them, γ represents the weight coefficient that controls the impact of adversarial training;

[0103]

[0104] Among them, y i represents the actual value of the i-th sample, represents the model prediction value of the i-th sample;

[0105]

[0106] Among them, δ represents the adversarial perturbation, ∈ represents the maximum norm of the perturbation, represents the model prediction output of the i-th sample under the adversarial perturbation δ, and ||δ|| represents the p-norm of δ;

[0107] S34. Use the adaptive learning rate optimization algorithm to train the deep learning model and update the model parameters:

[0108]

[0109] Among them, θ t represents the model parameter vector of the tth iteration, θ t+1 represents the model parameter vector of the t+1th iteration, η t represents the adaptive learning rate, represents the momentum estimate, represents the weighted average variance estimate, ∈ is used to avoid division by zero error, and λ1 represents the weight decay parameter;

[0110] S35. In each iteration, input the training data set X into the deep learning model and calculate the prediction result of the output layer through forward propagation Calculate the error according to the loss function Loss;

[0111] S36. Calculate the gradient of each parameter according to the error through the back propagation algorithm, and update the model parameters according to the adaptive learning rate optimization algorithm until the model converges and the loss function value no longer decreases significantly;

[0112] S37. Determine the initial model architecture and parameters, evaluate the model performance through the validation set, and save the trained model parameters as the initial configuration of the base model.

[0113] In this implementation manner, S5 specifically includes:

[0114] S51. Define real-time feedback dataset X new ={x n+1 ,x n+2 ,…,x n+q}, where q represents the number of samples, each sample number includes k features, denoted as x i ={x i1 ,x i2 ,…,x ik}, where i∈{n+1,n+2,…,n+q}, k represents the feature dimension;

[0115] S52, the real-time feedback data set X new With the previous historical data set X hist Combined to form a new comprehensive dataset X combined =X hist ∪X new , and triggers a progressive learning process;

[0116] S53, in the new comprehensive dataset X combined On the above, generate adversarial sample set X adv ={x′ n+1 ,x′ n+2 ,…,x′n+q}, the adversarial sample is created by new Add a perturbation δ to each sample in i Generate, the perturbation satisfies ||δ i || p ≤∈1:

[0117] x′ i =x i +δ i ;

[0118] Among them, δ i represents the disturbance vector, ∈1 represents the preset disturbance amplitude, ||δ i || p represents the p-norm of the disturbance;

[0119] S54. In the process of deep learning model retraining, the adversarial sample set X adv With the comprehensive dataset X combined Merge and construct an extended training set X extended =X combined ∪X adv , input into the model for training;

[0120] S55, expand the training set X extended Input into the deep learning model for retraining and optimization.

[0121] In this implementation manner, S6 specifically includes:

[0122] S61. Define a multi-objective optimization function F(φ), where φ represents the structural parameters of the model, including the number of network layers L, the number of nodes in each layer {n1,n2,…,n L} and the activation function parameter combination {α1,α2,…,α L}, the optimization goal is to minimize the prediction error and model complexity at the same time, and consider the interaction of parameters in each layer:

[0123]

[0124] Among them, m l represents the number of nodes in the lth layer, y j represents the actual value of the jth sample, represents the predicted value of the jth sample, n l represents the number of nodes in the lth layer, α l represents the activation function parameter of the lth layer, λ2, λ3 and λ4 represent weight coefficients, which control the influence of prediction error, complexity and activation function respectively;

[0125] S62, initialize the population of the adaptive differential evolution algorithm, define the population size as Q, and each individual Hq Represents a set of model parameter combinations φ q ={L q ,{n q1 ,n q2 ,…,n qL},{α q1 ,α q2 ,…,α qL}}, the population is initially randomly generated, and the population size is Q = {H1,H2,…,H Q};

[0126] S63. In each generation, for each individual H in the population q , through the adaptive combination of multiple mutation strategies to generate mutant individuals V q , mutation operations include differential mutation, random mutation and Gaussian mutation:

[0127] L q '=L q +ΔL q , where ΔL q ~U(-2,2);

[0128] n ql '=n ql +G l ·(n p1l -n p2l )+Z(0,1), where l∈{1,2,,L q};

[0129] α ql '=α ql +Δα ql , where Δα ql ~Z(0,0.05);

[0130] Among them, L q ' represents the number of grid layers after mutation, ΔL q Indicates the variation of the number of layers, n ql ' represents the number of nodes in the lth layer after mutation, G l represents the differential variation factor, Z(0,1) represents the standard normal distribution noise, α ql ' represents the activation function parameters after mutation, Δα ql Represents the normal distribution variation of the activation function parameters;

[0131] S64, using a multi-level fuzzy logic control system to correct the mutant individual, the fuzzy logic controller receives the fitness value F(V q ) and the fitness of the current individual F(H q ) as input and output the modified variant individual V′ q, the fuzzy rules include:

[0132] If F(V q ) <F(H q ), then the mutation effect is enhanced;

[0133] If F(V q )≈F(H q ), then retain or fine-tune the variation effect;

[0134] If F(V q )>F(H q ), then the mutation effect is weakened or cancelled;

[0135] S65, perform a crossover operation, and convert the corrected variant individual V′ q With the current individual H q Perform crossover to generate candidate individuals U q :

[0136] U q =(u q1 ,u q2 ,…,u qn );

[0137]

[0138] Among them, v′ qj Represents the variant individual V′ q The jth parameter of qj Indicates the current individual H q The jth parameter of j represents uniformly distributed random numbers, C r represents the crossover probability;

[0139] S66, according to the fitness value of the multi-objective optimization function F(φ), selecting individuals with better fitness to enter the next generation until a preset termination condition is met, including reaching a maximum number of iterations or the objective function converges;

[0140] S67. In the process of adaptive differential evolution, the fuzzy logic controller adaptively adjusts the mutation factor and crossover probability to ultimately determine the optimal model architecture parameter combination. And used for optimal configuration of deep learning models.

[0141] In this implementation manner, the S7 specifically includes:

[0142] S71, optimal model architecture parameter combination φ obtained based on S67 opt , build prediction model M opt , and use the current training data set X train ={x1,x2,…,xT}Train the model to obtain the initial prediction results;

[0143] S72. Define the adaptive dynamic prediction loss function L dyn (ψ), this loss function is used to measure the prediction error of the model at different time points and can be automatically adjusted according to data changes:

[0144]

[0145] Among them, ψ represents the current parameters of the model, T represents the number of time steps, and y t represents the actual value at time point t, represents the predicted value at time point t, represents the predicted value at the t-1th time point, Δt represents the time interval, σ t represents the volatility of the predicted value at the current time step, a and b represent the weight coefficients of the loss function, and τ represents the hyperparameter that controls the magnitude of the prediction adjustment;

[0146] S73, retrain the model based on the real-time updated project data, and adjust the current parameter ψ of the model to minimize L dyn (ψ);

[0147] S74. Each time new data arrives, by adjusting the volatility term σ in the loss function t and the time-dependent term Δt, which adapts to the dynamic changes of data;

[0148] S75. Use the updated model to generate change cost forecasts for future time steps And continue to collect data and update models throughout the project life cycle;

[0149] S76. Output the final prediction result

[0150] Embodiment 1:

[0151] In order to verify the feasibility of the present invention in implementation, the present invention is applied to a large-scale infrastructure construction project. The project changes are frequent and complex, involving design modifications, resource allocation and unforeseen environmental factors. Due to the long project cycle, the large number of participants, and the dynamic changes in external conditions, the traditional change cost forecasting method is difficult to cope with, often leading to the risk of budget overruns and project delays. Therefore, the management team of the project decided to adopt the engineering change cost forecasting and management method based on deep learning proposed in the present invention to improve the accuracy of change cost forecasting and realize dynamic management of the entire project process.

[0152] The project is located in East China, with a scheduled construction period of two years. The main changes involved include design adjustments, material price fluctuations, and changes in on-site construction conditions. To verify the effectiveness of the method of the present invention, the project management team collected data from similar projects in the past ten years at the beginning of the project, including historical records of engineering changes, actual cost data, and related environmental data, as initial training data.

[0153] At the start-up stage of the project, the management team first used the method of the present invention to construct a historical data set and preprocess the data. Through cleaning and denoising, outliers and noise in the data were removed, and features closely related to the change cost were extracted, including the type of change, change time, required resources, and actual expenditure. The processed data set was used to train the basic deep learning model, and a neural network model with three hidden layers was initially established. The activation function used the adaptive regularized ReLU function to ensure the generalization ability of the model.

[0154] As the project progressed, the management team continued to collect real-time data from the project, especially actual costs and resource consumption data from the construction site. These data were combined with historical data to form a new comprehensive data set, triggering a progressive learning process. The system dynamically adjusted model parameters based on new data, and by introducing adversarial sample generation technology, the model's robustness in the face of extreme data or abnormal situations was enhanced. For example, in the third month of the project, due to sudden natural disasters, the conditions on the construction site deteriorated and the actual costs increased significantly. With the addition of adversarial samples, the model was able to quickly adapt to new data patterns, and the prediction results were significantly optimized.

[0155] To further improve the model's predictive performance, the management team combined the adaptive differential evolution algorithm and fuzzy logic control technology. After each data update, the system automatically adjusts the network architecture of the deep learning model, including increasing or decreasing the number of network layers, adjusting the number of nodes in each layer, and optimizing the parameter combination of the activation function. For example, in the sixth month of the project, due to the further complexity of the engineering design, the model adaptively added a hidden layer and reconfigured the number of nodes, allowing the model to more accurately capture the potential patterns in complex data.

[0156] In actual application, the project management team found that after adopting the method of the present invention, the accuracy of the engineering change cost forecast was significantly improved. For example, in the ninth month of the project, the error rate between the change cost predicted by the model and the actual cost was reduced from 12.5% ​​of the original method to 3.2%. In addition, the management suggestions generated by the system, such as resource allocation and schedule adjustment, enable project managers to respond to various emergencies more calmly, effectively reducing the risk of project delays. Through the system's automated management suggestions, the team can take countermeasures in advance before changes occur, so that the project progresses smoothly.

[0157] In order to verify the beneficial effects of the present invention in more detail, the project management team compared the data of this project with two similar projects in the past. In the past projects, using the traditional change cost forecasting method, there were as many as 15 changes in the middle of the project, which eventually led to a 10% budget overrun and a 2-month delay in the construction period. In this project, although it also experienced multiple engineering changes, due to the use of the method of the present invention, the project was completed on schedule within the budget, the budget overrun rate was controlled within 1.5%, and the construction period delay was only 2 days, which was much lower than the level of previous projects.

[0158] Table 1 Comparison of engineering change cost prediction between traditional method and the method of the present invention

[0159]

[0160]

[0161] In Table 1 above, we compare the performance of the traditional method and the method of the present invention in the cost forecast of engineering changes, covering different stages of the project. The table shows that at multiple key nodes of the project, there is a large error between the cost forecast of the traditional method and the actual cost, especially in the 12th month, due to the increase in raw material costs caused by material price fluctuations, the forecast error rate of the traditional method is as high as 10%. In contrast, the method of the present invention significantly improves the accuracy of the forecast by dynamically adjusting the model parameters. In the same scenario, the forecast error rate is only 2.7%.

[0162] Further analysis shows that at the end of the project, the overall budget overrun rate of the traditional method reached 6.7%, while the method of the present invention controlled this ratio to 1.5%, fully demonstrating the advantages of the method of the present invention in budget control. At the same time, in terms of the control of construction delays, the method of the present invention also showed excellent performance, accurately predicting the actual number of days of construction delay, while the prediction error rate of the traditional method was 3.3%.

[0163] These data show that the method of the present invention has stronger adaptability and accuracy when dealing with complex and dynamic project environments. By introducing progressive learning, adaptive differential evolution algorithm and fuzzy logic control, the method of the present invention can respond to changes in real-time data in a timely manner, continuously optimize the model, and ensure the reliability of the prediction results. This capability not only reduces the risk of budget overruns, but also effectively controls project delays, thereby improving the overall execution efficiency and success rate of the project. These data further verify the application value and practical benefits of the method of the present invention in engineering change cost prediction and management.

[0164] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for predicting and managing engineering change costs based on deep learning, characterized in that: The steps include: S1. At the project initiation stage, collect historical project data and initial data of the current project and build a data set; S2, preprocess the data set, including cleaning, denoising and feature extraction; S3. Use the preprocessed data set to train the basic deep learning model and determine the initial deep learning model architecture and parameters; S4. At different stages of the project, real-time change data and actual cost data are continuously collected as dynamic feedback input to obtain real-time feedback data; S5. Based on real-time feedback data, the progressive learning process is triggered, and the deep learning model is retrained and optimized according to the new feedback data; S6. During the deep learning model retraining process, the adaptive differential evolution algorithm and fuzzy logic control are combined for hybrid optimization, and the optimal number of network layers, number of nodes and activation function are automatically selected; S7. Use the deep learning model optimized by the hybrid optimization algorithm to dynamically predict the future change costs of the project, and continuously optimize the prediction results as the project data is continuously updated; S8. Based on the optimized forecast results, the system automatically generates project management suggestions including resource allocation, schedule adjustment and risk control, and passes them to the project management system.

2. The method for predicting and managing engineering change costs based on deep learning according to claim 1, characterized in that: The S3 specifically includes: S31, define the input data set X = {x1, x2, ..., x n }, where X represents the feature matrix after preprocessing, n represents the total number of samples, and each sample includes m features, denoted by x i ={x i1 ,x i2 ,…,x im }; S32. Initialize the architecture of the deep learning model. The architecture of the deep learning model includes an input layer, a hidden layer, and an output layer. Define the number of nodes in the input layer as m, the number of hidden layers as Le, and the number of nodes in each layer as h. l , where l∈{1,2,…,Le}, the activation function f(x) uses the adaptive regularized ReLU function: f(x)=max(0,x)+λ·Ω(θ); Among them, λ represents the regularization parameter, and Ω(θ) represents the complexity penalty term of parameter θ: Among them, α and β represent weight coefficients, represents the L2 regularization term, ∥θ l ∥1 represents the L1 regularization term; S33. Define the loss function Loss, which includes the adversarial training loss L pred and the base prediction loss L adv : Loss=L pred +γ·L adv ; Among them, γ represents the weight coefficient that controls the impact of adversarial training; Among them, y i represents the actual value of the i-th sample, represents the model prediction value of the i-th sample; Among them, δ represents the adversarial perturbation, ∈ represents the maximum norm of the perturbation, represents the model prediction output of the i-th sample under the adversarial perturbation δ, and ||δ|| represents the p-norm of δ; S34. Use the adaptive learning rate optimization algorithm to train the deep learning model and update the model parameters: Among them, θ t represents the model parameter vector of the tth iteration, θ t+1 represents the model parameter vector of the t+1th iteration, η t represents the adaptive learning rate, represents the momentum estimate, represents the weighted average variance estimate, ∈ is used to avoid division by zero error, and λ1 represents the weight decay parameter; S35. In each iteration, input the training data set X into the deep learning model and calculate the prediction result of the output layer through forward propagation Calculate the error according to the loss function Loss; S36. Calculate the gradient of each parameter according to the error through the back propagation algorithm, and update the model parameters according to the adaptive learning rate optimization algorithm until the model converges and the loss function value no longer decreases significantly; S37. Determine the initial model architecture and parameters, evaluate the model performance through the validation set, and save the trained model parameters as the initial configuration of the base model.

3. The method for predicting and managing engineering change costs based on deep learning according to claim 1, characterized in that: The S5 specifically includes: S51. Define real-time feedback dataset X new ={x n+1 ,x n+2 ,…,x n+q }, where q represents the number of samples, each sample number includes k features, denoted as x i ={x i1 ,x i2 ,…,x ik }, where i∈{n+1,n+2,…,n+q}, k represents the feature dimension; S52, the real-time feedback data set X new With the previous historical data set X hist Combined to form a new comprehensive dataset X combined =X hist ∪X new , and triggers a progressive learning process; S53, in the new comprehensive dataset X combined On the above, generate adversarial sample set X adv ={x′ n+1 ,x′ n+2 ,…,x′ n+q }, the adversarial sample is created by new Add a perturbation δ to each sample in i Generate, the perturbation satisfies ∥δ i ∥ p ≤∈1: x′ i =x i +δ i ; Among them, δ i represents the disturbance vector, ∈1 represents the preset disturbance amplitude, ||δ i || p represents the p-norm of the disturbance; S54. In the process of deep learning model retraining, the adversarial sample set X adv With the comprehensive dataset X combined Merge and construct an extended training set X extended =X combined ∪X adv , input into the model for training; S55, expand the training set X extended Input into the deep learning model for retraining and optimization.

4. The method for predicting and managing engineering change costs based on deep learning according to claim 1, characterized in that: The S6 specifically includes: S61. Define a multi-objective optimization function F(φ), where φ represents the structural parameters of the model, including the number of network layers L, the number of nodes in each layer {n1,n2,…,n L } and the activation function parameter combination {α1,α2,…,α L }, the optimization goal is to minimize the prediction error and model complexity at the same time, and consider the interaction of parameters in each layer: Among them, m l represents the number of nodes in the lth layer, y j represents the actual value of the jth sample, represents the predicted value of the jth sample, n l represents the number of nodes in the lth layer, α l represents the activation function parameter of the lth layer, λ2, λ3 and λ4 represent weight coefficients, which control the influence of prediction error, complexity and activation function respectively; S62, initialize the population of the adaptive differential evolution algorithm, define the population size as Q, and each individual H q Represents a set of model parameter combinations φ q ={L q ,{n q1 ,n q2 ,…,n qL },{α q1 ,α q2 ,…,α qL }}, the population is initially randomly generated, and the population size is Q = {H1,H2,…,H Q }; S63. In each generation, for each individual H in the population q , through the adaptive combination of multiple mutation strategies to generate mutant individuals V q , mutation operations include differential mutation, random mutation and Gaussian mutation: L q '=L q +ΔL q , among whichΔL q ~U(-2,2); where l∈{1,2,,L q }; a ql '=a ql +Da ql , of whichDa ql ~Z(0,0.05); Among them, L q ' represents the number of grid layers after mutation, ΔL q Indicates the variation of the number of layers, n ql ' represents the number of nodes in the lth layer after mutation, G l represents the differential variation factor, Z(0,1) represents the standard normal distribution noise, α ql ' represents the activation function parameters after mutation, Δα ql Represents the normal distribution variation of the activation function parameters; S64, using a multi-level fuzzy logic control system to correct the mutant individual, the fuzzy logic controller receives the fitness value F(V q ) and the fitness of the current individual F(H q ) as input and output the modified variant individual V′ q , the fuzzy rules include: If F(V q ) <F(H q ), then the mutation effect is enhanced; If F(V q )≈F(H q ), then retain or fine-tune the variation effect; If F(V q )>F(H q ), then the mutation effect is weakened or cancelled; S65, perform a crossover operation, and convert the corrected variant individual V′ q With the current individual H q Perform crossover to generate candidate individuals U q : IN q =(in q1 ,in q2 ,…,in qn ); Among them, v′ qj Represents the variant individual V′ q The jth parameter of qj Indicates the current individual H q The jth parameter of j represents uniformly distributed random numbers, C r represents the crossover probability; S66, according to the fitness value of the multi-objective optimization function F(φ), selecting individuals with better fitness to enter the next generation until a preset termination condition is met, including reaching a maximum number of iterations or the objective function converges; S67. In the process of adaptive differential evolution, the fuzzy logic controller adaptively adjusts the mutation factor and crossover probability, and finally determines the optimal model architecture parameter combination φ opt ={L opt ,{n 1,opt ,n 2,opt ,…,n Lopt,opt },{α 1,opt ,α 2,opt ,…,α Lopt,opt }} and used for optimal configuration of deep learning models.

5. The method for predicting and managing engineering change costs based on deep learning according to claim 1, characterized in that: The S7 specifically includes: S71, optimal model architecture parameter combination φ obtained based on S67 opt , build prediction model M opt , and use the current training data set X train ={x1,x2,…,x T }Train the model to obtain the initial prediction results; S72. Define the adaptive dynamic prediction loss function L dyn (ψ): Among them, ψ represents the current parameters of the model, T represents the number of time steps, and y t represents the actual value at time point t, represents the predicted value at time point t, represents the predicted value at the t-1th time point, Δt represents the time interval, σ t represents the volatility of the predicted value at the current time step, a and b represent the weight coefficients of the loss function, and τ represents the hyperparameter that controls the magnitude of the prediction adjustment; S73, retrain the model based on the real-time updated project data, and adjust the current parameter ψ of the model to minimize L dyn (ψ); S74. Each time new data arrives, by adjusting the volatility term σ in the loss function t and the time-dependent term Δt, which adapts to the dynamic changes of data; S75. Use the updated model to generate change cost forecasts for future time steps And continue data collection and model updates throughout the project life cycle; S76. Output the final prediction result

Images