Dynamic project cost intelligent prediction method, system and equipment

Abstract

The invention relates to a dynamic project cost intelligent prediction method, system and equipment, and the method comprises the steps: anchoring a prediction dimension through a cost motivation map through a closed-loop process of cost motivation quantification, multi-source data structuring, feature engineering and hybrid model dynamic prediction, and carrying out the directional mapping from multi-source data to a cost motivation dimension, performing corresponding feature processing according to feature types on the basis of feature classification defined by the cost motivation quantization atlas to form a structured feature set for model training; project data of different stages are input into a trained mixed cost prediction model, dynamic project cost prediction is achieved, and the mixed cost prediction model comprises an XGBoost model used for extracting static features and an LSTM model based on extracted historical time series data features. Compared with the prior art, the method has the advantages of realizing more accurate and more adaptive project full-stage dynamic cost prediction and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of project cost prediction, in particular to a dynamic project cost intelligent prediction method, system and device. BACKGROUND

[0002] Rail transit signal system (such as CBTC, FAO full automatic operation system) is a core component of urban rail transit, and its cost accounts for a large proportion of the total investment of the line, and is significantly affected by technical iteration and supply chain fluctuations.

[0003] The traditional rail transit signal system project cost prediction only focuses on construction cost, and has not covered operation and maintenance cost (according to statistics, signal system maintenance cost accounts for 15%-20% of the total operation cost), and there is a lack of mature solutions for whole life cycle cost prediction.

[0004] At present, the existing project cost prediction methods mainly include: 1) Historical analogy method: reference similar historical project cost data for estimation, low precision, and unable to cope with the uniqueness of new projects. For example: reference signal system cost of similar lines (such as subway line 1 and line 2), scale according to line length, station quantity, etc., but unable to adapt to the technical differences of intelligent systems (such as FAO system cost is 20% higher than traditional).

[0005] 2) Parameter model method: based on single parameters such as "cost per square meter" to establish linear model, difficult to handle complex nonlinear relationships covering technology upgrade, localization rate, etc.

[0006] 3) Expert judgment method: relies on project manager's personal experience, subjective, and difficult to quantify and inherit.

[0007] 4) Method based on traditional software (such as Excel): although it can perform basic calculations, the data processing capability is limited, lacks intelligent analysis capability, and cannot respond to project changes in real time.

[0008] The existing methods cannot adapt to the characteristics of signal system technology complexity and strong dynamics, and mainly have the following shortcomings: Data island and insufficient fusion: cost data, progress data, procurement data, human data, etc. exist in different systems (such as ERP, CRM, BIM, IoT platform), forming a data island, and cannot be analyzed.

[0009] Static prediction cannot cope with dynamic risks: prediction is usually carried out before the project starts, and it is difficult to dynamically update and re-predict according to actual progress, market fluctuations (such as changes in raw material prices), risk events, etc. during project execution.

[0010] Weak non-linear factor capturing ability and insufficient prediction accuracy: traditional models cannot effectively capture and learn complex and non-linear factors affecting costs (such as the impact of weather on project duration, team efficiency fluctuations, supply chain delays, etc.), resulting in large prediction bias.

[0011] Lack of full life cycle foresight: Most predictions only cover the construction phase and do not consider the dynamic changes in operation and maintenance costs: for example, the initial construction cost of an intelligent signal system is 15% higher, but the operation and maintenance cost can be reduced by 30%. Traditional models cannot quantify such long-term benefits, leading to conservative investment decisions. In addition, most are post-recorded and analyzed, lacking early warning and root cause analysis capabilities for potential cost overrun risks.

[0012] Slow response: When project changes occur, cost prediction adjustments require a lot of manual recalculation, which is inefficient.

[0013] After searching, the Chinese invention patent application publication No. CN120542881A discloses a project cost intelligent management and control method and system based on dynamic multi-dimensional modeling, which constructs an end-to-end intelligent management and control system deeply integrating multi-source heterogeneous data (progress, resources, processes, risks, costs) and integrating advanced artificial intelligence technology, i.e. accurately depicting project dynamics through multi-modal encoding (Bi-LSTM, GCN, attention mechanism, etc.) and spatio-temporal feature extraction (spatio-temporal convolution and multi-head self-attention); innovatively using a gated graph neural network to dynamically model risk transmission relationships, realizing risk quantification and path tracking; combining residual networks and difference calculation to accurately predict costs and deviations, and based on Top-k path search and Transformer decoder to automatically generate actionable optimization strategies and natural language reports. The existing patent application has the problems of not disclosing cost driver quantification and poor prediction accuracy.

[0014] How to achieve accurate prediction of dynamic project costs becomes a technical problem to be solved. SUMMARY

[0015] The purpose of the present application is to overcome the defects of the prior art and provide a dynamic project cost intelligent prediction method, system and device.

[0016] The purpose of the present application can be achieved by the following technical solutions: According to one aspect of the present application, a dynamic project cost intelligent prediction method is provided, which comprises: Cost driver deep analysis and quantification: using a combination of work breakdown structure and failure mode and effects analysis to determine the cost drivers of the project, and quantitatively processing the cost drivers according to their categories to form a cost driver quantification map; Multi-source data collection and structured processing: extracting key information from the collected multi-source historical project data, quantifying the graph based on the cost drivers, mapping the key information to the corresponding dimensions of the graph, and storing it in a structured form in a dedicated cost database; Data cleaning and feature engineering: cleaning the historical project data of the dedicated cost database, and based on the feature classification defined by the cost driver quantification graph, processing the corresponding features according to the feature type to form a structured feature set for model training, including static features and historical time series data features; Model training and dynamic prediction: training the constructed hybrid cost prediction model, inputting project data of different stages into the trained hybrid cost prediction model to realize dynamic project cost prediction; wherein the hybrid cost prediction model includes an XGBoost model for extracting static features, an LSTM model based on extracting historical time series data features, and a fully connected neural network for adaptively weighting and integrating the XGBoost model output and the LSTM model output.

[0017] As a preferred technical solution, the hybrid cost prediction model adopts a two-stage hybrid prediction method, including: In the first stage, the XGBoost model extracts static features from the cost driver quantification graph, learns the complex mapping relationship between the static features and the baseline cost, and outputs a baseline cost prediction value; at the same time, the LSTM model predicts the cost consumption trend in the future time window based on the historical time series data of the project; In the second stage, the baseline cost prediction value output by the XGBoost model and the time series feature vector extracted by the LSTM model are spliced and input into a fully connected neural network for weighted integration to generate the final dynamic cost prediction value.

[0018] As a preferred technical solution, the process of training the constructed hybrid cost prediction model includes: taking the structured feature set of historical projects as input and taking the actual manual cost as output to construct a model training data set; using a 5-fold cross-validation method to train the hybrid cost prediction model, and optimizing the model hyperparameters through grid search.

[0019] As a preferred technical solution, for R&D projects, the cost drivers of the project include: Requirement dimension, including number of function points and operation maturity, number and complexity of external interfaces, and business rule density; Technical dimension, including technical preparation degree, number of real-time constraints, microsecond task proportion, and security authentication target; Scale dimension, including number of software code lines, number of hardware bill of materials, number of FPGA logic units, and number of reusable modules; Resource dimensions, including average experience years of project team, number of interdisciplinary cooperation, availability of experimental equipment; and management dimensions, including supplier rating level, demand change risk, and construction period requirement.

[0020] As a preferred technical solution, for engineering projects, the cost drivers of the project include: functional scope, including reuse functional scope and special functional scope; design complexity, including network scheme type and whether to establish user center; scale dimension, including number of test tracks, number of centralized stations, number of non-centralized stations, number of trains, number of control centers, and number of turnouts; certification constraints, including whether SIL4 safety certification is required and quality assurance requirement level; and management dimensions, including project cycle, number of interface interlocking meetings, and training duration.

[0021] As a preferred technical solution, the quantification of the cost drivers according to categories includes: classification and labeling of category-type drivers, direct collection of numerical values of continuous-type drivers, and quantification and scoring of drivers that are difficult to directly quantify by domain experts according to preset scoring standards.

[0022] As a preferred technical solution, the method further includes: for interface control documents, automatically calculating the number of interfaces through statistical analysis, and automatically matching the preset interface complexity weight according to the interface protocol type, so that unstructured interface data is converted into structured parameters for cost prediction.

[0023] As a preferred technical solution, the process of processing corresponding features according to feature types includes: for category-type features, using target coding technology to convert category features into numerical features with economic significance, with the average labor cost of historical projects as the target variable; for continuous-type features, using standardization processing to eliminate dimension differences; for features with interactive relationships, constructing feature cross terms to capture the synergistic effects between features, and forming a structured feature set for model training; wherein the features with interactive relationships include R&D CAP and technology complexity score, and line reconstruction type and night construction loss.

[0024] According to another aspect of the present application, a dynamic project cost intelligent prediction system is provided, which includes: Data acquisition and processing module: for interfacing with external systems, realizing the automatic collection and import of multi-source historical project data; extracting key information from the collected multi-source historical project data and mapping the key information to the cost driver quantification graph, and storing it in the special cost database in the form of structured data; Data cleaning and feature engineering module: cleaning the historical project data of the special cost database, and processing corresponding features according to feature types to form a structured feature set for model training; Cost driver management module: using the method of work breakdown structure and failure mode and effects analysis to determine the cost drivers of the project, and quantifying the cost drivers according to categories to form a cost driver quantification graph; Database module: store structured historical project data, cost driver parameters, feature sets and model parameters to form a special cost database; Model training and prediction module: train the constructed hybrid cost prediction model, input the project data of different stages into the trained hybrid cost prediction model, and realize dynamic project cost prediction; wherein the hybrid cost prediction model includes an XGBoost model based on static features, an LSTM model based on historical time series data, and a fully connected neural network that adaptively integrates the outputs of the XGBoost model and the LSTM model.

[0025] As a preferred technical solution, the hybrid cost prediction model adopts a two-stage hybrid prediction method, which includes: In the first stage, the XGBoost model extracts static features from the cost driver quantification graph, learns the complex mapping relationship between the static features and the baseline cost, and outputs a baseline cost prediction value; at the same time, the LSTM model predicts the cost consumption trend of the future time window based on the historical time series data of the project; In the second stage, the baseline cost prediction value output by the XGBoost model and the time series feature vector extracted by the LSTM model are spliced and input into a fully connected neural network for weighted integration to generate the final dynamic cost prediction value.

[0026] According to a third aspect of the present application, an electronic device is provided, comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to realize the method.

[0027] According to a fourth aspect of the present application, a computer readable storage medium is provided, which stores a computer program, and the program is executed by a processor to realize the method.

[0028] Compared with the prior art, the present application has the following beneficial effects: 1) The dynamic project cost intelligent prediction method of the application solves the problem of fuzzy cost factors in traditional methods by a closed-loop process of cost driver quantification, multi-source data structuring, feature engineering, and hybrid model prediction, anchors the prediction dimension with a cost driver map, makes the prediction more in line with the actual project, breaks the data silos through directional mapping of multi-source data to the cost driver dimension, improves the cost correlation and reusability of historical data, uses a hybrid model of "XGBoost (static) + LSTM (time series) + fully connected network (integration)", which covers both the basic cost logic of the project and the dynamic fluctuations, and finally realizes more accurate and adaptive dynamic cost prediction for the whole stage of the project. The strong correlation of each link in the process also improves the actual feasibility of the scheme.

[0029] 2) The application innovatively designs a hybrid cost prediction model that cooperates XGBoost model and LSTM model, realizes the separate deep mining and effective fusion of project static inherent attributes and execution dynamic time series information, and fundamentally solves the common technical problems of insufficient prediction accuracy and poor adaptability of traditional single models in dealing with long-cycle, high-complexity, and dynamic projects such as rail transit signal systems.

[0030] 3) The cost driver quantification scheme of the application refines the cost driver dimension by project type, and adapts the quantification method (classification marking / numerical collection / expert scoring) according to the type, which not only realizes the accurate disassembly of complex factors such as function, technology, and resources of R&D projects, but also covers the design complexity, scale, and certification constraints of engineering projects. Let the cost influencing factors of different types of projects change from fuzzy description to structured and calculable quantitative indicators, which improves the scene adaptability of cost drivers and provides accurate dimensional basis for subsequent data mapping and feature processing, and finally supports more accurate dynamic cost prediction that is closer to the actual project.

[0031] 4) The feature processing scheme of the application adapts to different types of features and uses different processing methods: for category type features, target coding is used to convert them into numerical values with economic significance, solving the problem that category information cannot participate in cost calculation; for continuous type features, standardization is done to eliminate dimensional interference, improving the comparability between features; for features with interactive relationship, cross terms are constructed to capture the rules of feature synergy affecting cost, and finally a high-quality structured feature set is formed, which meets the requirements of model training for different types of features and excavates the potential association between features, solving the problem of single feature processing method in traditional feature processing and ignoring feature interaction, effectively improving the learning efficiency and prediction accuracy of the subsequent cost prediction model. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 Key parameter diagram for rail transit R&D project cost in the application; Figure 2 A schematic diagram of a key process affecting the cost of a rail transit project in the present application; Figure 3 A schematic diagram of the process of the project cost intelligent prediction method in the present application. DETAILED DESCRIPTION

[0033] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should fall within the scope of protection of the present application.

[0034] The present embodiment relates to a dynamic project cost intelligent prediction method based on multi-source data fusion and intelligent algorithm, as shown in Figure 3 , comprising the following steps: Step one: cost driver deep analysis and quantification (signal system special) Purpose: For high-tech and high-complexity signal system integration projects, identify and quantify their unique technical and management driving factors.

[0035] 1-1, form an analysis team composed of signal system experts (including chief system architect, safety certification engineer, project delivery director), use the work breakdown structure (WBS) and failure mode and effects analysis (FMEA) combined method to specifically disassemble the cost influencing factors of research and development projects and engineering projects, and determine the cost drivers.

[0036] For research and development projects, determine the cost drivers, including: (a) requirement dimension (number of function points and operation maturity, number and complexity of external interfaces, business rule density); (b) technical dimension (technical preparation degree, number of real-time constraints, proportion of microsecond (us) tasks, safety certification target); (c) scale dimension (software code lines (kLOC), bill of materials (BOM) quantity, field programmable gate array (FPGA) logic unit number (kLUTs), number of reusable modules); (d) resource dimension (average experience of project team, number of interdisciplinary cooperation, availability of experimental equipment); (e) management dimension (supplier rating level, demand change risk, project duration requirement).

[0037] As Figure 1The key cost driver parameters and their quantification methods for R&D products (especially software and hardware combined systems) are listed in a structured manner. The abstract R&D complexity is decomposed into measurable and scorable specific parameters, providing a structured data source for feature engineering, and ensuring that the model can learn the inherent cost rules of R&D activities. The quantification methods use interval range division (high / medium / low), weighted calculation, risk superposition scoring (1-10 points), and other quantification means.

[0038] For engineering projects, the cost drivers are determined, including functional scope (reusable functional scope, such as whether to include point backup systems, centralized line microcomputer monitoring systems, PSDC systems, special functional scope, such as whether to include unmanned monitoring and remote restart functions, ATS TLV new interface), design complexity (network scheme type (SDH, LTE, industrial Ethernet), whether to establish a user center), scale dimension (test track number, centralized station number, non-centralized station number, train number, control center number, turnout number), certification constraints (whether to require SIL4 safety certification, quality assurance requirement level), management dimension (project cycle, interface interlocking meeting frequency, training duration). As shown in Figure 2 The key factors affecting construction cost are analyzed from the project execution level, with emphasis on the interaction between key processes, human resources, and schedule pressure. Optimization of key processes (usually involving core technology or safety certification) is constrained by project progress; unreasonable schedule targets can lead to unbalanced human resource allocation, insufficient skills, or increased communication costs, directly driving up project costs.

[0039] 1-2, Quantification of the above cost drivers by category, including: Classify and label category-type drivers, such as train operation mode (CBTC / FAO), train-ground communication method (LTE-M / Wi-Fi), line type (new / addition / extension), project management mode (general construction contracting / one-number society collaboration), whether to include R&D CAP (customized application platform), whether to include scenario management, whether to include user service center, whether to include off-site test track; Directly collect numerical values for continuous-type drivers, such as train number, centralized station number, test track distance, interface number, design liaison meeting frequency, training duration; For drivers that are difficult to quantify directly (such as interface complexity, technical complexity, risk coefficient), domain experts use pre-set scoring standards (1-10 points) to quantify and score, ultimately forming the "Rail Transit Signal System Project Cost Driver Quantification Atlas".

[0040] Step two: Historical project data collection and structured processing 2-1, Multi-source data collection: Obtain the basic cost data of historical projects (such as labor compensation, overtime pay, and travel expenses) from enterprise resource planning (ERP) systems, customer relationship management (CRM) systems; Obtain project versions, nodes, and requirement change records from technical management platforms (requirement management system RVS, project data management system PDM system); Extract technical parameters (such as train-ground communication methods, security authentication levels, and interface protocol types) from design documents (system requirement specification, interface control document); Extract implementation data (such as commissioning hours, integration delay records, and test track usage) from test reports and completion documents; 2-2, Structured storage: Build a cost-specific database for rail transit signal systems, and store the collected historical project data in a hierarchical structure of "project basic information - cost driver parameters - actual labor cost - hours record"; Then automatically extract key information from design documents (such as using LTE-M technology for train-ground communication and supporting FAO full-automatic operation) through text analysis technology, and map it to the corresponding dimensions in the cost driver quantization atlas; Automatically calculate the number of interfaces in the interface control document through statistical analysis, and automatically match the preset interface complexity weight according to the interface protocol type

preset interface protocol type - interface complexity weight corresponding rule

[0041] Step three: Data cleaning and feature engineering 3-1, Data cleaning: For missing values in historical project data, use the mean value of the same type of project to fill in (such as the mean value of project hours for projects of the same line type and same technology standard); For outliers (such as work records exceeding 8 hours per day), manually verify them in combination with project logs, and remove invalid data caused by extreme events (such as natural disasters causing project delay), to ensure data quality; 3-2, Feature engineering: Through feature processing, the model can capture key rules in the data (such as the impact of different technology types on cost), express complex information with simpler features, help the model learn more efficiently, and avoid redundant calculations. Including: For categorical features (such as project management mode, train-ground communication method, and interface protocol type), use target encoding (Target Encoding) technology to convert categorical features into numerical features with economic significance, with the mean value of labor cost of historical projects as the target variable; For continuous features (such as the number of trains and the number of centralized stations), use standardization processing (Z-Score standardization) to eliminate dimension differences; For features with interaction (such as R&D CAP and technical complexity score, line reconstruction type and night construction loss), cross-feature items (such as R&D CAP x technical complexity score, reconstruction line x night construction coefficient) are constructed to capture the synergistic effect between features.

[0042] The above various features ultimately form a structured feature set for model training, including: Static features: namely non-time series features, such as project size, interface complexity, technology type and other relatively stable features, corresponding to the static dimension in the cost driver; Historical time series data features: namely time series features, such as work hour records at different stages, cost consumption progress and other features that change over time, corresponding to the dynamic dimension in the cost driver.

[0043] Step four: prediction model development, training and dynamic prediction 4-1, model selection and training.

[0044] The cost of rail transit signal system project is not only affected by the complex nonlinearity of many static technical parameters (such as SIL4 level, communication system), but also by the continuous impact of dynamic factors such as progress delay, demand change and supply chain fluctuation in a period of several years. The dominant factors affecting the cost of the project may be different in the early stage (such as the design stage) and the later stage (such as the debugging stage), and the data distribution changes over time (concept drift). A single model is difficult to model these two heterogeneous information at the same time. The present application uses the division and cooperation of XGBoost model and LSTM model to model static complexity and dynamic time series respectively, and then fuses them to realize a more accurate description of the coupling influence mechanism.

[0045] Therefore, the present application adopts a hybrid cost prediction model of static and dynamic feature fusion, which cooperatively utilizes the advantages of XGBoost model and LSTM model to solve the prediction problem of the dual driving of static complex features and dynamic time series evolution in the cost of rail transit signal system project.

[0046] According to the complex interaction of the characteristics of rail transit signal system projects, the historical project data of a single enterprise may be limited (small sample), but the cost driver dimension is high. Therefore, the gradient boosting tree model (XGBoost) is selected as the core prediction model for static feature analysis, mainly dealing with high-dimensional structured static features (such as technical complexity score, interface number, equipment scale, etc.) describing the inherent properties of the project extracted from the cost driver quantification atlas. Its strong nonlinear fitting ability and feature importance sorting function can accurately learn the complex mapping relationship between static features and baseline cost. The XGBoost model effectively prevents overfitting through regularization and pruning strategies, and robustly learns the influence weight of key static features from small samples; the LSTM model efficiently utilizes limited time series data to learn general patterns by sharing time step parameters, improving the model's generalization ability in data-scarce scenarios and ensuring the effectiveness of the prediction model throughout the project life cycle.

[0047] For projects containing long-term work time series data (such as monthly cost consumption), the long short-term memory network (LSTM model) is introduced as a dynamic time series pattern learner, combined with time dimension features (such as schedule progress, seasonal factors) to optimize prediction accuracy. The LSTM model is used to specifically process time series data streams generated during project execution (such as monthly actual work time consumption, cumulative change frequency, key node completion rate, market price index sequence, etc.), and its gated recurrent unit can effectively capture the long-term dependence of cost consumption, periodic fluctuations, and trend changes, ensuring the effectiveness of the prediction model throughout the project life cycle.

[0048] With the structured feature set of historical projects as input and actual labor cost (man-month or total working hours) as output, the model training data set is constructed; The 5-fold cross-validation method is used to train the model, and the grid search (Grid Search) is used to optimize the model hyperparameters (such as learning rate, tree depth, leaf node number), so that the prediction error (root mean square error RMSE) of the model is controlled within 10%.

[0049] The hybrid cost prediction model adopts a two-stage hybrid prediction method. In the first stage, an XGBoost model outputs a baseline cost prediction value based on static features. Meanwhile, an LSTM model predicts the cost consumption trend in the future time window based on historical time series data. In the second stage, a feature fusion layer is designed to splice the baseline cost prediction value (or its derived features) output by the XGBoost model and the time series feature vector (such as the final hidden state) extracted by the LSTM model, and input them together into a fully connected neural network for weighted integration to generate the final dynamic cost prediction value. Each layer of the fully connected neural network performs linear transformation on the features through a weight matrix, adds a bias term, and finally introduces nonlinearity through an activation function (such as ReLU). Throughout the process, the fully connected neural network propagates backward through the loss function (such as the mean squared error of the predicted cost and the actual cost) to continuously optimize the weights of all layers, achieving adaptive weighted integration of XGBoost static features and LSTM time series features.

[0050] This design enables the model to deeply understand the static cost driving factors of the project itself and respond sensitively to the dynamic influences brought by the project process and external environment.

[0051] The application innovatively designs a hybrid cost prediction model in which the XGBoost model and the LSTM model work together, rather than simply stacking them. This hybrid cost prediction model achieves separate deep mining and effective fusion of project static inherent properties and execution dynamic time series information, fundamentally solving the common technical problems of insufficient prediction accuracy and poor adaptability of traditional single models in dealing with long-cycle, high-complexity, and dynamic projects such as rail transit signal systems. This architecture is the core technical guarantee for achieving high-precision dynamic prediction.

[0052] 4-2, Dynamic prediction application, covering the whole project process, achieving cost prediction and dynamic adjustment in different stages: In the pre-sales stage, the sales engineer inputs the preliminary technical solution of the new project (such as train operation mode, train-ground communication method, number of trains, and whether to include user service center), the system calls the hybrid cost prediction model, and outputs the labor cost estimation results (including total man-months, and stage-by-stage working hours (design / installation / commissioning / acceptance)) within 5 minutes, and generates a cost composition detail and sensitivity analysis report (such as a technical complexity score that increases by 1 point, the proportion of labor cost increase), providing a basis for bid pricing; Design phase: system architects refine the design scheme, update the cost driver parameters (such as the final number of interfaces is 25, and the SIL4 safety certification requirement is added), the system updates the feature set in real time and re-invokes the hybrid cost prediction model, outputs the updated labor cost prediction results, verifies the cost feasibility of the design scheme, if the predicted cost exceeds the budget threshold, automatically prompts the optimization direction (such as increasing the proportion of reusable modules); Execution phase: project managers regularly input project actual data (such as actual time consumption of design liaison meeting, interface commissioning delay time, labor input situation), the system compares the deviation between actual data and predicted data, if the deviation exceeds 10%, automatically triggers an early warning, and updates the model input features based on the latest actual data, dynamically adjusts the labor cost prediction results in the subsequent phase, and guides project resource allocation (such as increasing the number of debugging personnel, optimizing the construction plan).

[0053] The embodiment also relates to a dynamic project cost intelligent prediction system based on multi-source data fusion and intelligent algorithm, which comprises: A data acquisition and processing module is used for connecting to external systems such as ERP, CRM, RVS and PDM, realizing automatic acquisition and import of multi-source data of historical project data, processing the acquired historical project data, including: extracting key information of design files by text analysis, counting the number of interfaces and matching the complexity weight, realizing the conversion of unstructured data into structured parameters. A data cleaning and feature engineering module is used for processing missing values and abnormal values of historical project data to ensure data quality: missing values are filled with the mean value of the same type of project, and abnormal values are removed by combining log verification. For different feature types, appropriate methods are used for processing: high-base category features (category features) are converted by target encoding, continuous features are standardized, and feature cross items are constructed to form a training set. A cost driver management module is used for assembling an expert team, decomposing the research and development and engineering scene drivers by WBS and FMEA, including 5 dimensions such as demand, technology and scale for research and development, and 5 dimensions such as function range and design complexity for engineering; the category type driver is classified and labeled, the continuous type driver is valued, and the difficult to quantify driver is scored according to 1-10 points to form a quantitative atlas. It also provides a visual editing function of the cost driver quantitative atlas, supports experts to add and modify cost driver dimensions and quantitative standards; A database module is used for storing structured historical project data, cost driver parameters, feature sets and model parameters; Model training and prediction module: integrate XGBoost, LSTM model, support model training, parameter optimization and dynamic prediction. Select XGBoost / LightGBM as the core model, 5-fold cross-validation + grid search optimization, error <=8%; Time series data introduces LSTM model. The estimated value is calculated within 5 minutes before the sale, and the real-time update is verified during the design stage. The deviation is more than 10% in the execution stage, and the prediction is adjusted.

[0054] Visual output module: output prediction results in the form of table (cost composition details), chart (cost proportion pie chart, stage trend line), report (sensitivity analysis, early warning report), support data export and printing.

[0055] The cost driver quantification management module, data cleaning and feature engineering module of the application can convert the influencing factors of the signal system project cost into structured features (such as scores, weights), so that the prediction model can depict the deep influence of these management and human factors on the cost, thereby realizing more accurate dynamic prediction and early warning.

[0056] The electronic device includes a central processing unit (CPU) that can perform various appropriate actions and processes according to computer program instructions stored in a read-only memory (ROM) or loaded into a random access memory (RAM) from a storage unit. In the RAM, various programs and data required for device operation can also be stored. The CPU, ROM, and RAM are connected to each other via a bus. An input / output (I / O) interface is also connected to the bus.

[0057] A plurality of components in the device are connected to the I / O interface, including: an input unit such as a keyboard, a mouse, etc.; an output unit such as various types of displays, a speaker, etc.; a storage unit such as a magnetic disk, an optical disk, etc.; and a communication unit such as a network card, a modem, a wireless communication transceiver, etc. The communication unit allows the device to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunications networks.

[0058] The processing unit performs the various methods and processes described above. For example, in some embodiments, the method can be implemented as a computer software program tangibly embodied in a machine-readable medium, such as the storage unit. In some embodiments, part or all of the computer program can be loaded and / or installed on the device via the ROM and / or the communication unit. When the computer program is loaded into the RAM and executed by the CPU, one or more steps of the method described above can be performed. Alternatively, in other embodiments, the CPU can be configured to perform the method by any other appropriate means (e.g., with the help of firmware).

[0059] The functionality described above in this document can be performed, at least in part, by one or more hardware logic components. For example, and without limitation, illustrative types of hardware logic components that can be used include Field-programmable Gate Arrays (FPGAs), Program- specific Integrated Circuits (ASICs), Application-specific Standard Products (ASSPs), System-on-a-chip systems (SOCs), Complex Programmable Logic Devices (CPLDs), etc.

[0060] Program code for carrying out the methods of the present application can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the program code, when executed by the processor or controller, produces the functions / operations specified in the flowcharts and / or block diagrams. The program code can be executed entirely on a machine, partially on a machine, partially on a machine and partially on a remote machine or entirely on a remote machine or server.

[0061] In the context of the present application, a machine-readable medium can be a tangible medium that can contain or store program for use by or in connection with an instruction execution system, apparatus, or device. The machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable storage medium can include, without limitation, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of the machine-readable storage medium will include one or more lines of electrical connections, portable computer disks, hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or Flash memory), optical fibers, portable compact disc read-only memories, optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0062] The above description is only specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present application, and these modifications or replacements should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A dynamic intelligent project cost prediction method, characterized in that, The method includes: In-depth analysis and quantification of cost drivers: Using a combination of work breakdown structure and failure mode and impact analysis, the cost drivers of the project are identified and quantified according to category to form a cost driver quantification map; Multi-source data collection and structured processing: Extract key information from the collected multi-source historical project data, map the key information to the corresponding dimension of the cost driver quantification map, and store it in a structured form in a dedicated cost database. Data cleaning and feature engineering: historical project data in the dedicated cost database is cleaned, and feature classification is defined based on the cost driver quantification map. According to the feature type, corresponding feature processing is performed to form a structured feature set for model training, including static features and historical time series data features. Model Training and Dynamic Prediction: The constructed hybrid cost prediction model is trained by inputting project data from different stages into the trained hybrid cost prediction model to achieve dynamic project cost prediction. The hybrid cost prediction model includes an XGBoost model for extracting static features, an LSTM model based on extracting features from historical time-series data, and a fully connected neural network for adaptively weighting and integrating the outputs of the XGBoost model and the LSTM model.

2. The dynamic project cost intelligent prediction method according to claim 1, characterized in that, The hybrid cost prediction model employs a two-stage hybrid prediction method, including: In the first stage, the XGBoost model, based on static features extracted from the cost driver quantification map, learns the complex mapping relationship between static features and baseline cost to output a baseline cost prediction value; at the same time, the LSTM model predicts the cost consumption trend of future time windows based on the project's historical time series data. In the second stage, the baseline cost prediction output by the XGBoost model is concatenated with the temporal feature vector extracted by the LSTM model and input into a fully connected neural network for weighted integration to generate the final dynamic cost prediction.

3. The dynamic project cost intelligent prediction method according to claim 1, characterized in that, The process of training the constructed hybrid cost prediction model includes: constructing a model training dataset by taking the structured feature set of historical projects as input and the actual labor cost as output; training the hybrid cost prediction model using the 5-fold cross-validation method; and optimizing the model hyperparameters through grid search.

4. The dynamic project cost intelligent prediction method according to claim 1, characterized in that, For R&D projects, the cost drivers include: The requirements dimension includes the number of functionalities and their operational maturity, the number and complexity of external interfaces, and the density of business rules; Technical dimensions include technical readiness, number of real-time constraints, proportion of microsecond-level tasks, and security certification objectives; Scale dimensions include the number of lines of software code, the number of hardware bill of materials, the number of FPGA logic units, and the number of reusable modules; Resource dimensions include the average years of experience of the project team, the number of interdisciplinary collaborations, and the availability of experimental equipment; And management dimensions, including supplier rating levels, risk of requirement changes, and timeline requirements.

5. The dynamic intelligent project cost prediction method according to claim 1, characterized in that, For engineering projects, the cost drivers include: Functional scope, including reused functional scope and special functional scope; Design complexity includes the type of network solution and whether a user center should be established. In terms of scale, this includes the number of test tracks, centralized stations, non-centralized stations, trains, control centers, and turnouts. Certification constraints, including whether SIL4 safety certification is required and the level of quality assurance requirements; And management dimensions, including project cycle, number of interface interlocking meetings, and training duration.

6. The dynamic project cost intelligent prediction method according to claim 1, characterized in that, The aforementioned quantification of cost drivers by category includes: classifying and labeling categorical drivers, directly collecting numerical values ​​for continuous drivers, and having domain experts quantify and score drivers that are difficult to quantify directly according to preset scoring standards.

7. The dynamic project cost intelligent prediction method according to claim 1, characterized in that, The method further includes: for interface control documents, statistical analysis is performed to automatically calculate the number of interfaces, and preset interface complexity weights are automatically matched according to the interface protocol type, so that unstructured interface data is converted into structured parameters for cost prediction.

8. The dynamic project cost intelligent prediction method according to claim 1, characterized in that, The process of processing corresponding features according to feature type includes: For categorical features, target coding technique is used, with the average labor cost of historical projects as the target variable, to transform categorical features into economically meaningful numerical features; For continuous features, standardization is used to eliminate dimensional differences; For features with interactive relationships, feature cross terms are constructed to capture the synergistic effects between features, forming a structured feature set for model training; among them, features with interactive relationships include R&D CAP and technical complexity scores, line renovation types and nighttime construction losses.

9. A system utilizing the dynamic project cost intelligent prediction method according to any one of claims 1 to 8, characterized in that, The system includes: Data acquisition and processing module: used to interface with external systems to realize the automatic acquisition and import of historical project data from multiple sources; extract key information from the acquired historical project data from multiple sources, map the key information to the corresponding dimension of the cost driver quantification map based on the aforementioned cost driver quantification map, and store it in a structured form in a dedicated cost database; Data cleaning and feature engineering module: Cleans historical project data in the dedicated cost database, and performs corresponding feature processing according to the feature classification defined by the cost driver quantification map to form a structured feature set for model training; Cost Driver Management Module: Using a combination of work breakdown structure and failure mode and impact analysis, the module identifies the cost drivers of the project and quantifies them according to category to form a cost driver quantification map. Database module: Stores structured historical project data, cost driver parameters, feature sets and model parameters to form a dedicated cost database; Model training and prediction module: Train the constructed hybrid cost prediction model, input project data from different stages into the trained hybrid cost prediction model to achieve dynamic project cost prediction; wherein the hybrid cost prediction model includes an XGBoost model for extracting static features, an LSTM model based on extracting features from historical time series data, and a fully connected neural network for adaptively weighting and integrating the outputs of the XGBoost model and the LSTM model.

10. The system according to claim 9, characterized in that, The hybrid cost prediction model employs a two-stage hybrid prediction method, including: In the first stage, the XGBoost model, based on static features extracted from the cost driver quantification map, learns the complex mapping relationship between static features and baseline cost to output a baseline cost prediction value; at the same time, the LSTM model predicts the cost consumption trend of future time windows based on the project's historical time series data. In the second stage, the baseline cost prediction output by the XGBoost model is concatenated with the temporal feature vector extracted by the LSTM model and input into a fully connected neural network for weighted integration to generate the final dynamic cost prediction.

11. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 8.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 8.

Images