A project cost management method and system incorporating long short-term memory networks

Abstract

This invention provides a project cost management method and system combining long short-term memory networks in the interdisciplinary field of artificial intelligence and project management. The method includes: Step S1, extracting time-series features and feature relationships from a constructed project dataset, building a cost feature vector library based on each time-series feature, and initializing a cost prediction model based on each feature relationship; Step S2, monitoring the latest project status data in real time through a cost dynamic monitoring agent, and dynamically updating the cost feature vector library based on the project status data; Step S3, training the project cost management model using the project dataset, and pushing the project cost management model to the project management terminal through a federated gateway; Step S4, inputting real-time cost management requests into the project cost management model at the project management terminal to obtain a real-time project cost prediction report. The advantage of this invention is that it greatly improves the accuracy and reliability of project cost management.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and project management, and in particular to a project cost management method and system that combines long short-term memory networks. Background Technology

[0002] In the increasingly complex and uncertain modern project management environment, Project Cost Management (PCM), as one of the core competency domains of the project management knowledge system, has risen in importance from simple budget execution to a strategic activity concerning project success or failure and core competitiveness of enterprises. Project cost management is not an isolated process, but a systematic process highly coupled with and mutually constraining elements such as project schedule, scope, quality, and resources. It spans the entire project lifecycle (initiation, planning, execution, monitoring, and closure), encompassing three key stages: cost estimation, cost budgeting, and cost control. Its core objective is to minimize project costs or maximize return on investment (ROI) while meeting established project scope, quality, and schedule constraints through precise planning, dynamic monitoring, and intelligent optimization of personnel, materials, equipment, management fees, and other project expenses.

[0003] With the rapid development of information technology, project cost management practices have gradually shifted from traditional manual recording and paper reports to a highly electronic, information-based, and automated model. Various Project Management Information Systems (PMIS) and professional cost management software (such as Primavera P6, Microsoft Project, SAPPS, Oracle PPM, etc.) are widely used across industries, greatly improving data processing efficiency. These software systems typically incorporate multiple mathematical models and methodologies to assist in cost estimation, budget preparation, and variance analysis. However, when addressing the complexity and dynamism of modern projects, existing technological systems are gradually revealing the following significant technical bottlenecks and inherent defects: 1. Static traps and insufficient dynamic adaptability: Most mainstream cost management software currently relies on a pre-set static budget baseline for its core algorithm model. This makes it difficult to make effective and adaptive prediction updates based on project progress (such as delays or advances of the critical path), actual cost of completed work, and real-time changes in environmental variables (such as market supply and demand fluctuations and supply chain disruptions) during the actual execution of a project.

[0004] Once the cost baseline is determined early in the project, it often lacks a flexible adjustment mechanism, or the adjustment process is highly dependent on manual intervention and is significantly delayed. This makes the mid-term and late-term forecasts, especially for completion estimates and estimates still to be completed, lose sensitivity to the actual evolution of the project status. The widening trend of the deviation between the forecast results and the actual cost cannot be captured and corrected in a timely manner.

[0005] 2. Weak ability to model nonlinear features: Project cost data inherently possesses highly non-stationary and non-linear characteristics, such as drastic cost fluctuations caused by sudden events, abrupt changes in resource prices, and non-linear cost growth curves. Existing software commonly employs traditional statistical and econometric models, such as linear / multinomial regression-based analysis or classic time series models (e.g., Autoregressive Integrated Moving Average (ARIMA) and exponential smoothing). These models typically assume that the data exhibits a linear trend or periodicity, and that fluctuations (noise) follow a specific distribution (e.g., Gaussian distribution). When faced with the aforementioned complex and highly non-linear cost evolution patterns, the fitting ability, prediction accuracy, and robustness of these models significantly decrease.

[0006] 3. Insufficient depth of multimodal data fusion and information silos: The factors affecting project costs are diverse, and there are close correlations between data from different sources. Complete cost management requires deep integration and analysis of multimodal and heterogeneous data sources, such as financial data streams, schedule data, resource data, and external environment data. However, existing cost management solutions suffer from serious fragmentation at the data integration level: most systems treat the data from the above-mentioned different sources as independent information silos and lack effective cross-modal feature extraction and correlation analysis capabilities.

[0007] Therefore, how to provide a project cost management method and system that combines long short-term memory networks to improve the accuracy and reliability of project cost management has become an urgent technical problem to be solved. Summary of the Invention

[0008] The technical problem to be solved by this invention is to provide a project cost management method and system that combines long short-term memory networks to improve the accuracy and reliability of project cost management.

[0009] In a first aspect, the present invention provides a project cost management method incorporating a long short-term memory network, comprising the following steps: Step S1: The cost management server acquires a large amount of historical project data, preprocesses the historical project data, and then constructs a project dataset. Step S2: The cost management server extracts time series features and feature relationships from the project dataset using a pre-trained feature extraction model, constructs a cost feature vector library based on each time series feature, and initializes the pre-trained cost prediction model based on each feature relationship. The cost prediction model is constructed based on a feature input layer, a cross-modal interaction layer, and a prediction output layer; The feature input layer is used to extract high-level feature representations from the input structured query features and time series features; the cross-modal interaction layer is used to fuse the high-level feature representations of the structured query features and time series features to obtain a second fused feature, and to perform cross-modal interactive reasoning on the second fused feature through a second long short-term memory network to obtain an enhanced feature; the prediction output layer is used to map the enhanced feature to a candidate cost prediction result with confidence evaluation; Step S3: The cost management server creates a cost dynamic monitoring agent, which monitors the latest project status data in real time. After verifying the project status data, the cost feature vector library is dynamically updated based on the project status data through a sliding window mechanism. Step S4: The cost management server creates a project cost management model based on the multimodal fusion layer, dynamic adaptation layer, and prediction generation layer, and sets the prediction loss function and optimization reward function of the project cost management model. The multimodal fusion layer is used to perform semantic parsing on the input cost management request to obtain structured query features. It simultaneously retrieves time series features from the cost feature vector library through the cost management request. After modally aligning the structured query features with the retrieved time series features, it inputs them into the dynamic adaptation layer. The dynamic adaptation layer is used to call the initialized cost prediction model, perform cross-modal interactive reasoning on the input structured query features and time series features, output candidate cost prediction results with confidence evaluation, and input the candidate cost prediction results into the prediction generation layer; The prediction generation layer is used to generate a project cost prediction report based on the candidate cost prediction results, according to the preset project risk management framework; the report includes project cost estimation results, deviation analysis results, and optimization suggestions. Step S5: The cost management server trains the project cost management model using the project dataset, the prediction loss function, and the optimization reward function. The trained project cost management model is then compressed, and the compressed project cost management model is pushed to the distributed project management terminal through the federated gateway. Step S6: The project management terminal obtains the input cost management request, parses and verifies the cost management request to obtain the real-time cost management request, inputs the real-time cost management request into the deployed project cost management model to obtain the real-time project cost prediction report, and displays the real-time project cost prediction report. Step S7: The project management terminal records cost management logs in real time, including at least the project ID, execution time, real-time cost management requests, real-time project cost forecast reports, and execution feedback. Through a reinforcement learning mechanism, the project cost management model is iteratively optimized using the cost management logs. Step S8: The project management terminal encrypts and performs distributed backup of the cost management log.

[0010] Furthermore, step S1 specifically includes: The cost management server acquires a large amount of historical project data, which includes structured data and unstructured data. The structured data includes at least budget data and schedule data, and the unstructured data includes at least resource data and external environment data. For the structured data in each of the historical project data, preprocessing is performed including at least the following: duplicate data removal, missing value handling, error data correction, data type conversion, standardization, normalization, and data integration. For the text data in the unstructured data, preprocessing is performed including at least the following: useless character removal, text format unification, word segmentation, part-of-speech tagging, and stop word removal. For the image data in the unstructured data, preprocessing is performed including at least the following: image format unification, image size adjustment, and image enhancement. For the audio data in the unstructured data, preprocessing is performed including at least the following: audio format conversion and audio noise reduction. The preprocessed historical project data are labeled with at least the cost management requirements, project costs, deviation values, reasons for deviations, and optimization suggestions. A project dataset is then constructed based on the labeled historical project data.

[0011] Furthermore, in step S2, the feature extraction model is constructed based on a time series processing layer, a correlation capture layer, and a feature output layer; The time series processing layer is constructed based on a time series feature extraction channel, a primary correlation preprocessing channel, and a feature concatenation unit. The time series feature extraction channel is used to extract time-dependent features from the input data through a bidirectional gated recurrent unit. The primary correlation preprocessing channel is used to extract preliminary spatial dependency features from the input data through a one-dimensional convolutional neural network. The feature concatenation unit is used to concatenate the time-dependent features and the preliminary spatial dependency features to obtain time series features. The association capture layer is constructed based on a self-attention feature extraction module and a feature fusion module; the self-attention feature extraction module is used to extract attention-weighted features from time series features through a multi-head self-attention unit; the feature fusion module is used to fuse the attention-weighted features and time series features to obtain a first fused feature, and to filter redundant information from the first fused feature through a first long short-term memory network to obtain the feature association; The feature output layer is used to output the time series features and feature correlations.

[0012] Furthermore, step S3 specifically includes: The cost management server creates a cost dynamic monitoring agent for monitoring project status data, configures the monitoring parameters of the cost dynamic monitoring agent, and monitors the latest project status data in real time through the cost dynamic monitoring agent. The project status data includes at least budget data, schedule data, resource data, and external environment data. After verifying the status data of each project according to the preset verification rules, the corresponding length of data is selected from the status data of each project through the sliding window mechanism and input into the feature extraction model. The cost feature vector library is dynamically updated based on the time series features output by the feature extraction model.

[0013] Furthermore, step S5 specifically includes: The cost management server divides the project dataset into a training set, a validation set, and a test set in an 8:1:1 ratio. The project cost management model is trained using the training set, continuously optimizing its parameters using the prediction loss function and the optimization reward function until a preset early termination condition is met. The prediction accuracy is calculated using the validation set to validate the trained project cost management model. The F1 score and confidence level are calculated using the test set to test the validated model. The tested model is compressed using quantization techniques, and the compressed model is pushed in real-time to the distributed project management terminals via a pre-deployed federated gateway.

[0014] Secondly, this invention provides a project cost management system incorporating a long short-term memory network, comprising the following modules: The project dataset construction module is used by the cost management server to obtain a large amount of historical project data, and then construct the project dataset after preprocessing the historical project data. The initialization module is used by the cost management server to extract time series features and feature relationships from the project dataset through a pre-trained feature extraction model, construct a cost feature vector library based on each time series feature, and initialize the pre-trained cost prediction model based on each feature relationship. The cost prediction model is constructed based on a feature input layer, a cross-modal interaction layer, and a prediction output layer; The feature input layer is used to extract high-level feature representations from the input structured query features and time series features; the cross-modal interaction layer is used to fuse the high-level feature representations of the structured query features and time series features to obtain a second fused feature, and to perform cross-modal interactive reasoning on the second fused feature through a second long short-term memory network to obtain an enhanced feature; the prediction output layer is used to map the enhanced feature to a candidate cost prediction result with confidence evaluation; The cost feature vector library update module is used by the cost management server to create a cost dynamic monitoring agent, which monitors the latest project status data in real time through the cost dynamic monitoring agent, and after verifying each project status data, dynamically updates the cost feature vector library based on the project status data through a sliding window mechanism. The project cost management model creation module is used by the cost management server to create a project cost management model based on a multimodal fusion layer, a dynamic adaptation layer, and a prediction generation layer, and to set the prediction loss function and optimization reward function of the project cost management model. The multimodal fusion layer is used to perform semantic parsing on the input cost management request to obtain structured query features. It simultaneously retrieves time series features from the cost feature vector library through the cost management request. After modally aligning the structured query features with the retrieved time series features, it inputs them into the dynamic adaptation layer. The dynamic adaptation layer is used to call the initialized cost prediction model, perform cross-modal interactive reasoning on the input structured query features and time series features, output candidate cost prediction results with confidence evaluation, and input the candidate cost prediction results into the prediction generation layer; The prediction generation layer is used to generate a project cost prediction report based on the candidate cost prediction results, according to the preset project risk management framework; the report includes project cost estimation results, deviation analysis results, and optimization suggestions. The project cost management model deployment module is used by the cost management server to train the project cost management model using the project dataset, the prediction loss function, and the optimization reward function, compress the trained project cost management model, and push the compressed project cost management model to the distributed project management terminal through the federated gateway. The cost management module is used by the project management terminal to obtain input cost management requests, parse and verify the cost management requests to obtain real-time cost management requirements, input the real-time cost management requirements into the deployed project cost management model to obtain a real-time project cost prediction report, and display the real-time project cost prediction report. The project cost management model optimization module is used to record cost management logs in real time on the project management terminal, including at least the project ID, execution time, real-time cost management requests, real-time project cost prediction reports, and execution feedback. Through a reinforcement learning mechanism, the project cost management model is iteratively optimized using the cost management logs. The cost management log management module is used by the project management terminal to encrypt and distribute the cost management logs.

[0015] Furthermore, the project dataset construction module is specifically used for: The cost management server acquires a large amount of historical project data, which includes structured data and unstructured data. The structured data includes at least budget data and schedule data, and the unstructured data includes at least resource data and external environment data. For the structured data in each of the historical project data, preprocessing is performed including at least the following: duplicate data removal, missing value handling, error data correction, data type conversion, standardization, normalization, and data integration. For the text data in the unstructured data, preprocessing is performed including at least the following: useless character removal, text format unification, word segmentation, part-of-speech tagging, and stop word removal. For the image data in the unstructured data, preprocessing is performed including at least the following: image format unification, image size adjustment, and image enhancement. For the audio data in the unstructured data, preprocessing is performed including at least the following: audio format conversion and audio noise reduction. The preprocessed historical project data are labeled with at least the cost management requirements, project costs, deviation values, reasons for deviations, and optimization suggestions. A project dataset is then constructed based on the labeled historical project data.

[0016] Furthermore, in the initialization module, the feature extraction model is constructed based on a time series processing layer, a correlation capture layer, and a feature output layer; The time series processing layer is constructed based on a time series feature extraction channel, a primary correlation preprocessing channel, and a feature concatenation unit. The time series feature extraction channel is used to extract time-dependent features from the input data through a bidirectional gated recurrent unit. The primary correlation preprocessing channel is used to extract preliminary spatial dependency features from the input data through a one-dimensional convolutional neural network. The feature concatenation unit is used to concatenate the time-dependent features and the preliminary spatial dependency features to obtain time series features. The association capture layer is constructed based on a self-attention feature extraction module and a feature fusion module; the self-attention feature extraction module is used to extract attention-weighted features from time series features through a multi-head self-attention unit; the feature fusion module is used to fuse the attention-weighted features and time series features to obtain a first fused feature, and to filter redundant information from the first fused feature through a first long short-term memory network to obtain the feature association; The feature output layer is used to output the time series features and feature correlations.

[0017] Furthermore, the cost feature vector library update module is specifically used for: The cost management server creates a cost dynamic monitoring agent for monitoring project status data, configures the monitoring parameters of the cost dynamic monitoring agent, and monitors the latest project status data in real time through the cost dynamic monitoring agent. The project status data includes at least budget data, schedule data, resource data, and external environment data. After verifying the status data of each project according to the preset verification rules, the corresponding length of data is selected from the status data of each project through the sliding window mechanism and input into the feature extraction model. The cost feature vector library is dynamically updated based on the time series features output by the feature extraction model.

[0018] Furthermore, the project cost management model deployment module is specifically used for: The cost management server divides the project dataset into a training set, a validation set, and a test set in an 8:1:1 ratio. The project cost management model is trained using the training set, continuously optimizing its parameters using the prediction loss function and the optimization reward function until a preset early termination condition is met. The prediction accuracy is calculated using the validation set to validate the trained project cost management model. The F1 score and confidence level are calculated using the test set to test the validated model. The tested model is compressed using quantization techniques, and the compressed model is pushed in real-time to the distributed project management terminals via a pre-deployed federated gateway.

[0019] The advantages of this invention are: 1. A project dataset is constructed by acquiring a large amount of historical project data through a cost management server. A pre-trained feature extraction model extracts time-series features and feature relationships from the project dataset. A cost feature vector library is built based on these time-series features, and the pre-trained cost prediction model is initialized based on these feature relationships. The cost management server monitors the latest project status data in real time through a cost dynamic monitoring agent. After verifying the project status data, the cost feature vector library is dynamically updated based on the project status data using a sliding window mechanism. The cost management server creates a project cost management model based on a multimodal fusion layer, a dynamic adaptation layer, and a prediction generation layer. The prediction loss function and optimization reward function of the project cost management model are set. The project cost management model is trained using the project dataset, the prediction loss function, and the optimization reward function. The trained project cost management model is compressed and pushed to the distributed project management terminal through a federated gateway. The project management terminal parses and verifies the input cost management request to obtain the real-time cost management request. The real-time cost management request is input into the deployed project cost management model to obtain a real-time project cost prediction report. The real-time project cost prediction report is displayed and records at least the project ID, execution time, real-time cost management request, and real-time cost information in real time. The project cost forecast report and execution feedback cost management logs utilize a reinforcement learning mechanism to iteratively optimize the project cost management model. The cost management logs are encrypted and backed up in a distributed manner. Specifically, by introducing a dynamic cost monitoring agent and sliding window mechanism, project status data is monitored in real time, and the cost feature vector library is dynamically updated. This directly addresses the shortcomings of static baselines, allowing the system to adaptively adjust based on the latest data, reducing manual intervention and lag. The reinforcement learning mechanism uses cost management logs to iteratively optimize the project cost management model, ensuring that the model continuously adapts to project changes. The dynamic adaptation layer calls the cost forecast model output... The confidence assessment of candidate cost predictions, combined with the prediction generation layer to generate real-time project cost prediction reports, enhances the sensitivity of predictions to actual conditions. The project cost management model utilizes a long short-term memory network for cross-modal interactive reasoning, effectively improving the modeling capability for nonlinear features. Input data is processed through neural network models (such as feature input layers and cross-modal interaction layers), replacing traditional statistical models and improving robustness to non-stationary and nonlinear features. By parsing users' natural language descriptions of cost management needs into structured query features through a multimodal fusion layer, and simultaneously retrieving dynamically updated cost feature vector libraries in real time, historical time-series data (time series features) are accurately correlated.Then, through deep feature fusion inference via a long short-term memory network in a cross-modal interaction layer, implicit correlations between textual requests and dynamic time-series data are automatically mined. Finally, under a federated learning architecture, multi-source knowledge accumulation across projects and devices is achieved, completely breaking down data silos and significantly improving the accuracy and reliability of project cost management.

[0020] 2. Extract time-series features and feature relationships from historical project data through a pre-trained feature extraction model, and build a cost feature vector library; LSTM, as a core component, is good at capturing long-term dependencies in time-series data, which is crucial in project cost management (such as cost fluctuations, seasonal changes, etc.), thereby greatly improving the efficiency and accuracy of feature extraction, reducing noise interference, and making the model initialization more robust.

[0021] 3. The cost prediction model consists of a feature input layer, a cross-modal interaction layer, and a prediction output layer. It can align and fuse structured query features and time series features. This multimodal design generates enhanced features through cross-modal interactive reasoning, which solves the limitation of independent processing of data types in traditional methods. This significantly improves the generalization ability and robustness of prediction, and is especially suitable for multi-source data integration in complex project environments.

[0022] 4. By using a cost dynamic monitoring agent and sliding window mechanism, project status data is monitored and verified in real time, and the cost feature vector library is updated dynamically. This gives the technical solution the ability to quickly respond to changes (such as cost overruns or demand adjustments) in project management, avoids model obsolescence, and maintains the timeliness of predictions.

[0023] 5. The prediction output layer not only generates candidate cost prediction results but also provides confidence assessment, enabling users to quantify the credibility of the prediction. In addition, the prediction generation layer directly outputs a complete report based on the project risk management framework, including project cost estimation results, deviation analysis results, and optimization suggestions. This simplifies the decision-making process, reduces manual intervention, improves risk management efficiency, and enhances the interpretability of the solution.

[0024] 6. By using a reinforcement learning mechanism to iteratively optimize the project cost management model using cost management logs, this self-learning design enables the project cost management model to continuously adapt to new data and scenarios in practical applications, improve long-term performance, reduce retraining costs, and enhance the sustainability and accuracy of the technical solution in dynamic project environments.

[0025] 7. After compression, the project cost management model is deployed to the distributed project management terminal through a federated gateway, which reduces the consumption of terminal resources (such as computing and storage) and also realizes a lightweight and scalable distributed architecture. In the project management scenario, it ensures that the system can efficiently handle multiple concurrent requests, support large-scale applications, and protect data privacy through the federation mechanism.

[0026] 8. By combining Long Short-Term Memory (LSTM) networks to efficiently process historical project data, extract time-series features, and construct a cost feature vector library, and innovatively employing a cross-modal interaction mechanism to fuse structured queries and time-series data, the prediction accuracy, robustness, and generalization ability are significantly improved. Simultaneously, by leveraging a cost dynamic monitoring agent and a sliding window mechanism to update data in real time, the model's timeliness and responsiveness are ensured, outputting candidate results with confidence assessments and a complete report including cost estimation, bias analysis, and optimization suggestions, thereby enhancing decision reliability and interpretability. The model is continuously iterated and optimized through reinforcement learning mechanisms, and lightweight, scalable distributed applications are achieved through model compression and federated deployment. Data encryption and distributed backup ensure security, ultimately forming an end-to-end automated process that effectively reduces management burden, improves efficiency, and reduces the risk of project cost overruns.

[0027] 9. By integrating structured data (such as budgets and schedules) with unstructured data (such as resources, text, images, and audio from the external environment), and customizing hierarchical preprocessing workflows for different data types (including data cleaning, format standardization, semantic analysis, and multimedia optimization), and combining professional dimensions from the field of cost management (such as reasons for deviations and optimization suggestions) for fine-grained annotation, a high-quality, multi-dimensional project dataset is finally generated. This not only significantly improves the consistency and usability of the data, but also significantly reduces the cost of manual intervention through automated processes, providing an accurate and reliable data foundation for subsequent cost prediction, deviation analysis, and decision optimization, thereby enhancing the level of intelligence in project cost management.

[0028] 10. By innovatively integrating the time series processing layer, the association capture layer, and the feature output layer, efficient feature extraction is achieved. Its core advantages include the efficient fusion of bidirectional gated recurrent units (GRU) and one-dimensional convolutional neural networks (CNN) to simultaneously capture time-dependent features and preliminary spatial-dependent features, ensuring feature comprehensiveness and data adaptability; furthermore, a multi-head self-attention mechanism and feature fusion module are adopted, combined with a long short-term memory network (LSTM) to filter redundant information, significantly enhancing contextual understanding and the refinement of associations, improving the model's robustness and convergence speed; the modular overall design also gives the solution flexible scalability and high output quality, effectively improving the accuracy and efficiency of feature extraction.

[0029] 11. Real-time dynamic monitoring of project budget, schedule, resources, and external environment data is achieved through a cost dynamic monitoring agent. Combined with automated verification rules, data reliability is ensured. The sliding window mechanism and intelligent feature extraction model are used to adaptively analyze time-series data, thereby significantly improving the accuracy and timeliness of cost management. It can respond to project risks (such as budget overruns or resource fluctuations) in an instant, and achieve self-optimization of predictive capabilities through a continuously updated cost feature vector library. Ultimately, while reducing the cost of manual intervention, it provides high-precision trend insights and resource optimization basis for decision-making.

[0030] 12. Scientific data partitioning, dual-function driven model training, and early stopping mechanisms ensured the accuracy and training efficiency of the prediction model. Combined with multi-index validation testing (accuracy, F1 score, confidence), the reliability of the model was significantly improved. At the same time, quantization technology was used to compress the model to reduce resource consumption, and the compressed model was pushed in real time and securely through a federated gateway. Finally, lightweight, highly responsive, and scalable intelligent project cost management was achieved on the distributed management terminal. This fully automated design not only ensured the robustness and generalization ability of the model, but also greatly optimized the system resource utilization and business response time.

[0031] 13. By innovatively integrating Long Short-Term Memory (LSTM) networks with a multimodal feature fusion mechanism, and constructing a dynamic temporal feature library combined with cross-modal interactive reasoning, high-precision prediction of project costs is achieved. Real-time data verification and sliding window updates are supported through a cost dynamic monitoring agent, solving the static lag problem of traditional models. End-to-end training loops integrate reinforcement learning and optimized reward functions for continuous model iteration, significantly improving prediction robustness. Federated compression deployment reduces resource consumption, and encrypted log backup ensures data security. The final output includes a prediction report containing bias analysis and optimization suggestions, providing actionable intelligent decision support while ensuring efficiency, adaptability, and security. Attached Figure Description

[0032] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0033] Figure 1 This is a flowchart of a project cost management method that incorporates long short-term memory networks according to the present invention.

[0034] Figure 2 This is a schematic diagram of the structure of a project cost management system that incorporates a long short-term memory network according to the present invention. Detailed Implementation

[0035] The overall approach of the technical solution in this application is as follows: By introducing a dynamic cost monitoring agent and a sliding window mechanism, project status data is monitored in real time, and the cost feature vector library is dynamically updated. This directly addresses the shortcomings of static baselines, allowing the system to adaptively adjust based on the latest data, reducing manual intervention and lag. A reinforcement learning mechanism is used to iteratively optimize the project cost management model using cost management logs, ensuring that the project cost management model continuously adapts to project changes. The dynamic adaptation layer calls the cost prediction model to output candidate cost prediction results with confidence assessment, and combines this with the prediction generation layer to generate a real-time project cost prediction report, enhancing the sensitivity of the prediction to the actual situation. The project cost management model is configured to use a long short-term memory network. The system performs cross-modal interactive reasoning, effectively improving the modeling capability of nonlinear features. It processes input data through a neural network model, replacing traditional statistical models and enhancing robustness to non-stationary and nonlinear features. By parsing users' natural language descriptions of cost management requests into structured query features through a multimodal fusion layer, it simultaneously retrieves dynamically updated cost feature vector libraries in real time, accurately associating them with historical time-series data. Furthermore, it uses a long short-term memory network in the cross-modal interaction layer for deep feature fusion reasoning, automatically uncovering implicit correlations between textual requests and dynamic time-series data. Ultimately, under the federated learning architecture, it achieves multi-source knowledge accumulation across projects and terminals, completely breaking down data barriers and thus improving the accuracy and reliability of project cost management.

[0036] Please refer to Figures 1 to 2 As shown, a preferred embodiment of the project cost management method combining long short-term memory networks of the present invention includes the following steps: Step S1: The cost management server acquires a large amount of historical project data, preprocesses the historical project data, and then constructs a project dataset. Step S2: The cost management server extracts time series features and feature relationships from the project dataset using a pre-trained feature extraction model, constructs a cost feature vector library based on each time series feature, and initializes the first long short-term memory network of the pre-trained cost prediction model based on each feature relationship. Initializing a Long Short-Term Memory (LSTM) network based on feature associations mainly involves optimizing the initial parameter settings, input feature selection, or data preprocessing methods of the LSTM by analyzing the correlations between features, thereby enabling the LSTM to better capture key information in the data. The cost prediction model is constructed based on a feature input layer, a cross-modal interaction layer, and a prediction output layer; The feature input layer is used to extract high-level feature representations from the input structured query features and time series features; the cross-modal interaction layer is used to fuse the high-level feature representations of the structured query features and time series features to obtain a second fused feature, and to perform cross-modal interactive reasoning on the second fused feature through a second long short-term memory network to obtain an enhanced feature; the prediction output layer is used to map the enhanced feature to a candidate cost prediction result with confidence evaluation; The feature input layer consists of a structured query processing module and a time series preprocessing module. The structured query processing module extracts semantic representations from the input structured query features. Its network structure includes a fully connected layer, which embeds the structured query features into a low-dimensional vector representation. This layer uses the ReLU activation function to introduce non-linearity and outputs a query embedding vector. The time series preprocessing module extracts time patterns from the input time series features (such as historical cost data sequences). Its network structure includes a fully connected layer, which standardizes and reduces the dimensionality of the original time series features and outputs a time series embedding vector. The two modules form a dual-input channel, where the output of the structured query processing module (query embedding vector) and the output of the time series preprocessing module (time series embedding vector) are concatenated to form a high-level feature representation. The cross-modal interaction layer consists of an LSTM fusion module. This LSTM fusion module performs cross-modal interactive reasoning on high-level feature representations of structured queries and time-series features within the LSTM network. Its network structure is a multi-gated LSTM unit, which takes the fused input features as sequence input. Specifically, the query embedding vector is copied and concatenated to the time-series embedding vector at each time step. The LSTM unit dynamically fuses structured query information and extracts sequence dependency features when processing the time series, with the LSTM hidden state serving as the output. This module includes a sequence input gate, a forget gate, and an output gate, enhancing cross-modal interaction by learning long-term dependencies. The prediction output layer consists of a candidate cost prediction module and a confidence evaluation module. The candidate cost prediction module generates multiple candidate cost predictions based on the enhanced features. Its network structure includes a fully connected layer and a Gaussian Mixture Model Unit (GMM). The GMM maps the enhanced features to a high-dimensional space through the fully connected layer, and then outputs multiple Gaussian distribution parameters (mean, variance), each corresponding to a candidate cost prediction result. The confidence evaluation module outputs a confidence score for each candidate cost prediction result. Its network structure includes a fully connected layer and a sigmoid activation layer. The GMM compresses the enhanced features into a probability space and then outputs a confidence score in the range [0,1] through the sigmoid function, representing the reliability of the prediction result. The pre-trained feature extraction model extracts time-series features and feature relationships from historical project data and builds a cost feature vector library. LSTM, as a core component, is good at capturing long-term dependencies in time-series data, which is crucial in project cost management (such as cost fluctuations, seasonal changes, etc.), thereby greatly improving the efficiency and accuracy of feature extraction, reducing noise interference, and making the model initialization more robust.

[0037] The cost prediction model consists of a feature input layer, a cross-modal interaction layer, and a prediction output layer. It can align and fuse structured query features and time series features. This multimodal design generates enhanced features through cross-modal interactive reasoning, which solves the limitation of independent processing of data types in traditional methods. This significantly improves the generalization ability and robustness of prediction, and is especially suitable for multi-source data integration in complex project environments.

[0038] The prediction output layer not only generates candidate cost prediction results but also provides confidence assessments, enabling users to quantify the credibility of predictions. In addition, the prediction generation layer directly outputs a complete report based on the project risk management framework, including project cost estimation results, deviation analysis results, and optimization suggestions. This simplifies the decision-making process, reduces manual intervention, improves risk management efficiency, and enhances the interpretability of the solution.

[0039] Step S3: The cost management server creates a cost dynamic monitoring agent, which monitors the latest project status data in real time. After verifying the project status data, the cost feature vector library is dynamically updated based on the project status data through a sliding window mechanism. By using a cost dynamic monitoring agent and sliding window mechanism, project status data is monitored and verified in real time, and the cost feature vector library is dynamically updated. This gives the technical solution the ability to quickly respond to changes (such as cost overruns or demand adjustments) in project management, avoids model obsolescence, and maintains the timeliness of predictions.

[0040] Step S4: The cost management server creates a project cost management model based on the multimodal fusion layer, dynamic adaptation layer, and prediction generation layer, and sets the prediction loss function and optimization reward function of the project cost management model. The multimodal fusion layer is used to perform semantic parsing on the input cost management request to obtain structured query features. It simultaneously retrieves time series features from the cost feature vector library through the cost management request. After modally aligning the structured query features with the retrieved time series features, it inputs them into the dynamic adaptation layer. The dynamic adaptation layer is used to call the initialized cost prediction model, perform cross-modal interactive reasoning on the input structured query features and time series features, output candidate cost prediction results with confidence evaluation, and input the candidate cost prediction results into the prediction generation layer; The prediction generation layer is used to generate a project cost prediction report based on the candidate cost prediction results, according to the preset project risk management framework; the report includes project cost estimation results, deviation analysis results, and optimization suggestions. The multimodal fusion layer consists of a semantic parsing module, a feature retrieval module, and a modality alignment module; Semantic parsing module: Using a pre-trained language model (such as BERT) network structure based on Transformer, the module takes a cost management request (text sequence) as input, extracts contextual semantic features through a self-attention mechanism, and outputs structured query features (high-dimensional embedding vectors) to capture the core intent and related attributes of the query (such as project size and time range).

[0041] Feature retrieval module: Uses Approximate Nearest Neighbor (ANN) retrieval network structure (such as the vector index implemented in the Faiss library), inputs cost feature vector library (pre-stored historical cost time series data embedding), and retrieves the k most relevant time series features (represented as sequence embedding vectors) through structured query features to supplement historical trend information and enhance the basis for dynamic prediction.

[0042] Modality alignment module: It uses a cross-modal attention mechanism network structure (such as a multi-layer Transformer encoder), takes structured query features and retrieved time series features as input, learns the alignment between features through attention weights, and outputs a modality alignment feature vector of uniform dimension to eliminate modal ambiguity and realize the fusion of text and sequence data.

[0043] The dynamic adaptation layer consists of a cost prediction module, a cross-modal interaction unit, and a confidence assessment unit. Cost prediction module: Using a pre-trained time series prediction model network structure (such as LSTM or GRU network), the time series part of the modality alignment features is input, long-term dependencies are learned through recurrent units, and the initial cost prediction value (e.g., a numerical sequence of future cost estimates) is output to generate the baseline prediction result.

[0044] Cross-modal interaction unit: Using a cross-attention mechanism network structure (such as the Transformer decoder), the structured query features in the modality-aligned features are input to the output of the cost prediction module. The interaction weights are calculated through multi-head attention, and the feature representation after interaction is output. This is used to fuse text and sequence information and improve the adaptability to complex scenarios.

[0045] Confidence assessment unit: Using a probability output layer network structure (such as a softmax classifier), the output of the cross-modal interaction unit is input, the probability distribution of each prediction result is calculated and a confidence score (range of 0-1) is output to quantify the reliability of the prediction and support risk management decisions.

[0046] The prediction generation layer consists of report generator units; Report Generator Unit: Using a Transformer-based sequence generation model network structure (such as the GPT series models), it takes as input the candidate cost prediction results with confidence and the preset project risk management framework (such as input risk thresholds and constraints), generates a text sequence through the decoder layer and attention mechanism, and outputs a structured report containing project cost estimation results (numerical prediction), deviation analysis results (risk assessment compared with actual data), and optimization suggestions (operational guidelines based on risk analysis) to comprehensively support project decision-making.

[0047] Step S5: The cost management server trains the project cost management model using the project dataset, the prediction loss function, and the optimization reward function. The trained project cost management model is then compressed, and the compressed project cost management model is pushed to the distributed project management terminal through the federated gateway. After compression, the project cost management model is deployed to the distributed project management terminal through a federated gateway, reducing terminal resource consumption (such as computing and storage) and achieving a lightweight and scalable distributed architecture. In project management scenarios, it ensures that the system can efficiently handle multiple concurrent requests, support large-scale applications, and protect data privacy through a federated mechanism.

[0048] Step S6: The project management terminal obtains the input cost management request, parses and verifies the cost management request to obtain the real-time cost management request, inputs the real-time cost management request into the deployed project cost management model to obtain the real-time project cost prediction report, and displays the real-time project cost prediction report. Step S7: The project management terminal records cost management logs in real time, including at least the project ID, execution time, real-time cost management requests, real-time project cost forecast reports, and execution feedback. Through a reinforcement learning mechanism, the project cost management model is iteratively optimized using the cost management logs. By using a reinforcement learning mechanism to iteratively optimize the project cost management model using cost management logs, this self-learning design enables the project cost management model to continuously adapt to new data and scenarios in practical applications, improve long-term performance, reduce retraining costs, and enhance the sustainability and accuracy of the technical solution in dynamic project environments.

[0049] Step S8: The project management terminal encrypts and performs distributed backup of the cost management log.

[0050] By combining Long Short-Term Memory (LSTM) networks to efficiently process historical project data, extract time-series features, and construct a cost feature vector library, and innovatively employing a cross-modal interaction mechanism to fuse structured queries and time-series data, the prediction accuracy, robustness, and generalization ability are significantly improved. Simultaneously, by leveraging a cost dynamic monitoring agent and a sliding window mechanism to update data in real time, the model's timeliness and responsiveness are ensured, outputting candidate results with confidence assessments and a complete report including cost estimates, bias analysis, and optimization suggestions, thereby enhancing decision reliability and interpretability. The model is continuously iterated and optimized through reinforcement learning mechanisms, and lightweight, scalable distributed applications are achieved through model compression and federated deployment. Data encryption and distributed backup ensure security, ultimately forming an end-to-end automated process that effectively reduces management burden, improves efficiency, and mitigates the risk of project cost overruns.

[0051] Step S1 specifically involves: The cost management server acquires a large amount of historical project data, which includes structured data and unstructured data. The structured data includes at least budget data and schedule data, and the unstructured data includes at least resource data and external environment data. For the structured data in each of the historical project data, preprocessing is performed including at least the following: duplicate data removal, missing value handling, error data correction, data type conversion, standardization, normalization, and data integration. For the text data in the unstructured data, preprocessing is performed including at least the following: useless character removal, text format unification, word segmentation, part-of-speech tagging, and stop word removal. For the image data in the unstructured data, preprocessing is performed including at least the following: image format unification, image size adjustment, and image enhancement. For the audio data in the unstructured data, preprocessing is performed including at least the following: audio format conversion and audio noise reduction. The preprocessed historical project data are labeled with at least the cost management requirements, project costs, deviation values, reasons for deviations, and optimization suggestions. A project dataset is then constructed based on the labeled historical project data.

[0052] By integrating structured data (such as budgets and schedules) with unstructured data (such as resources, text, images, and audio from the external environment), and customizing hierarchical preprocessing workflows (including data cleaning, format standardization, semantic analysis, and multimedia optimization) for different data types, and combining professional dimensions from the field of cost management (such as reasons for deviations and optimization suggestions) for fine-grained annotation, a high-quality, multi-dimensional project dataset is ultimately generated. This not only significantly improves the consistency and usability of the data, but also significantly reduces the cost of manual intervention through automated processes, providing an accurate and reliable data foundation for subsequent cost prediction, deviation analysis, and decision optimization, thereby enhancing the level of intelligence in project cost management.

[0053] In step S2, the feature extraction model is constructed based on a time series processing layer, a correlation capture layer, and a feature output layer. The time series processing layer is constructed based on a time series feature extraction channel, a primary correlation preprocessing channel, and a feature concatenation unit. The time series feature extraction channel is used to extract time-dependent features from the input data through a bidirectional gated recurrent unit. The primary correlation preprocessing channel is used to extract preliminary spatial dependency features from the input data through a one-dimensional convolutional neural network. The feature concatenation unit is used to concatenate the time-dependent features and the preliminary spatial dependency features to obtain time series features. The association capture layer is constructed based on a self-attention feature extraction module and a feature fusion module; the self-attention feature extraction module is used to extract attention-weighted features from time series features through a multi-head self-attention unit; the feature fusion module is used to fuse the attention-weighted features and time series features to obtain a first fused feature, and to filter redundant information from the first fused feature through a first long short-term memory network to obtain the feature association; The feature output layer is used to output the time series features and feature correlations.

[0054] The feature output layer consists of a global average pooling module and a fully connected output module. The global average pooling module uses a global average pooling layer to apply average pooling to the output of the correlation capture layer along the time step dimension, outputting dimensionality-reduced features for aggregating sequence information, extracting global feature representations, and reducing the impact of noise. The fully connected output module uses a fully connected layer (FC) and a Tanh activation function layer. The fully connected layer (output dimension 32) is followed by a Tanh activation function layer (to prevent gradient vanishing). The input is the output of the global average pooling module, which is mapped to the final feature vector (time series features, feature correlations) for non-linear mapping of output features. Tanh activation ensures that the feature values ​​are in the range of [-1,1], improving representational power.

[0055] By innovatively integrating a time-series processing layer, a correlation capture layer, and a feature output layer, efficient feature extraction is achieved. Its core advantages include the efficient fusion of a bidirectional gated recurrent unit (GRU) and a one-dimensional convolutional neural network (CNN) to simultaneously capture temporally dependent features and preliminary spatially dependent features, ensuring feature comprehensiveness and data adaptability. Furthermore, the adoption of a multi-head self-attention mechanism and a feature fusion module, combined with a long short-term memory network (LSTM) to filter redundant information, significantly enhances contextual understanding and the refinement of correlations, improving the model's robustness and convergence speed. The modular overall design also endows the solution with flexible scalability and high output quality, effectively improving the accuracy and efficiency of feature extraction.

[0056] Step S3 specifically involves: The cost management server creates a cost dynamic monitoring agent for monitoring project status data, configures the monitoring parameters of the cost dynamic monitoring agent, and monitors the latest project status data in real time through the cost dynamic monitoring agent. The project status data includes at least budget data, schedule data, resource data, and external environment data. After verifying the status data of each project according to the preset verification rules, the corresponding length of data is selected from the status data of each project through the sliding window mechanism and input into the feature extraction model. The cost feature vector library is dynamically updated based on the time series features output by the feature extraction model.

[0057] By using a cost dynamic monitoring agent, real-time dynamic monitoring of project budget, schedule, resources, and external environment data is achieved. Combined with automated verification rules, data reliability is ensured. Furthermore, a sliding window mechanism and intelligent feature extraction model are used to adaptively analyze time-series data, thereby significantly improving the accuracy and timeliness of cost management. This approach can respond to project risks (such as budget overruns or resource fluctuations) in an instant, and can also achieve self-optimization of predictive capabilities through a continuously updated cost feature vector library. Ultimately, while reducing the cost of manual intervention, it provides high-precision trend insights and resource optimization basis for decision-making.

[0058] Step S5 specifically involves: The cost management server divides the project dataset into a training set, a validation set, and a test set in an 8:1:1 ratio. The project cost management model is trained using the training set, continuously optimizing its parameters using the prediction loss function and the optimization reward function until a preset early termination condition is met. The prediction accuracy is calculated using the validation set to validate the trained project cost management model. The F1 score and confidence level are calculated using the test set to test the validated model. The tested model is compressed using quantization techniques, and the compressed model is pushed in real-time to the distributed project management terminals via a pre-deployed federated gateway.

[0059] Scientific data partitioning, dual-function driven model training, and early stopping mechanisms ensured the accuracy and training efficiency of the prediction model. Multi-index validation testing (accuracy, F1 score, confidence) significantly improved model reliability. Simultaneously, quantization technology was used to compress the model to reduce resource consumption, and real-time secure push of the compressed model was achieved through a federated gateway. Ultimately, lightweight, highly responsive, and scalable intelligent project cost management was realized on the distributed management terminal. This fully automated design not only ensured the robustness and generalization ability of the model but also greatly optimized system resource utilization and business response timeliness.

[0060] A preferred embodiment of the project cost management system incorporating a long short-term memory network according to the present invention includes the following modules: The project dataset construction module is used by the cost management server to obtain a large amount of historical project data, and then construct the project dataset after preprocessing the historical project data. An initialization module is used by the cost management server to extract time series features and feature relationships from the project dataset through a pre-trained feature extraction model, construct a cost feature vector library based on each time series feature, and initialize the first long short-term memory network of the pre-trained cost prediction model based on each feature relationship. Initializing a Long Short-Term Memory (LSTM) network based on feature associations mainly involves optimizing the initial parameter settings, input feature selection, or data preprocessing methods of the LSTM by analyzing the correlations between features, thereby enabling the LSTM to better capture key information in the data. The cost prediction model is constructed based on a feature input layer, a cross-modal interaction layer, and a prediction output layer; The feature input layer is used to extract high-level feature representations from the input structured query features and time series features; the cross-modal interaction layer is used to fuse the high-level feature representations of the structured query features and time series features to obtain a second fused feature, and to perform cross-modal interactive reasoning on the second fused feature through a second long short-term memory network to obtain an enhanced feature; the prediction output layer is used to map the enhanced feature to a candidate cost prediction result with confidence evaluation; The feature input layer consists of a structured query processing module and a time series preprocessing module. The structured query processing module extracts semantic representations from the input structured query features. Its network structure includes a fully connected layer, which embeds the structured query features into a low-dimensional vector representation. This layer uses the ReLU activation function to introduce non-linearity and outputs a query embedding vector. The time series preprocessing module extracts time patterns from the input time series features (such as historical cost data sequences). Its network structure includes a fully connected layer, which standardizes and reduces the dimensionality of the original time series features and outputs a time series embedding vector. The two modules form a dual-input channel, where the output of the structured query processing module (query embedding vector) and the output of the time series preprocessing module (time series embedding vector) are concatenated to form a high-level feature representation. The cross-modal interaction layer consists of an LSTM fusion module. This LSTM fusion module performs cross-modal interactive reasoning on high-level feature representations of structured queries and time-series features within the LSTM network. Its network structure is a multi-gated LSTM unit, which takes the fused input features as sequence input. Specifically, the query embedding vector is copied and concatenated to the time-series embedding vector at each time step. The LSTM unit dynamically fuses structured query information and extracts sequence dependency features when processing the time series, with the LSTM hidden state serving as the output. This module includes a sequence input gate, a forget gate, and an output gate, enhancing cross-modal interaction by learning long-term dependencies. The prediction output layer consists of a candidate cost prediction module and a confidence evaluation module. The candidate cost prediction module generates multiple candidate cost predictions based on the enhanced features. Its network structure includes a fully connected layer and a Gaussian Mixture Model Unit (GMM). The GMM maps the enhanced features to a high-dimensional space through the fully connected layer, and then outputs multiple Gaussian distribution parameters (mean, variance), each corresponding to a candidate cost prediction result. The confidence evaluation module outputs a confidence score for each candidate cost prediction result. Its network structure includes a fully connected layer and a sigmoid activation layer. The GMM compresses the enhanced features into a probability space and then outputs a confidence score in the range [0,1] through the sigmoid function, representing the reliability of the prediction result. The pre-trained feature extraction model extracts time-series features and feature relationships from historical project data and builds a cost feature vector library. LSTM, as a core component, is good at capturing long-term dependencies in time-series data, which is crucial in project cost management (such as cost fluctuations, seasonal changes, etc.), thereby greatly improving the efficiency and accuracy of feature extraction, reducing noise interference, and making the model initialization more robust.

[0061] The cost prediction model consists of a feature input layer, a cross-modal interaction layer, and a prediction output layer. It can align and fuse structured query features and time series features. This multimodal design generates enhanced features through cross-modal interactive reasoning, which solves the limitation of independent processing of data types in traditional methods. This significantly improves the generalization ability and robustness of prediction, and is especially suitable for multi-source data integration in complex project environments.

[0062] The prediction output layer not only generates candidate cost prediction results but also provides confidence assessments, enabling users to quantify the credibility of predictions. In addition, the prediction generation layer directly outputs a complete report based on the project risk management framework, including project cost estimation results, deviation analysis results, and optimization suggestions. This simplifies the decision-making process, reduces manual intervention, improves risk management efficiency, and enhances the interpretability of the solution.

[0063] The cost feature vector library update module is used by the cost management server to create a cost dynamic monitoring agent, which monitors the latest project status data in real time through the cost dynamic monitoring agent, and after verifying each project status data, dynamically updates the cost feature vector library based on the project status data through a sliding window mechanism. By using a cost dynamic monitoring agent and sliding window mechanism, project status data is monitored and verified in real time, and the cost feature vector library is dynamically updated. This gives the technical solution the ability to quickly respond to changes (such as cost overruns or demand adjustments) in project management, avoids model obsolescence, and maintains the timeliness of predictions.

[0064] The project cost management model creation module is used by the cost management server to create a project cost management model based on a multimodal fusion layer, a dynamic adaptation layer, and a prediction generation layer, and to set the prediction loss function and optimization reward function of the project cost management model. The multimodal fusion layer is used to perform semantic parsing on the input cost management request to obtain structured query features. It simultaneously retrieves time series features from the cost feature vector library through the cost management request. After modally aligning the structured query features with the retrieved time series features, it inputs them into the dynamic adaptation layer. The dynamic adaptation layer is used to call the initialized cost prediction model, perform cross-modal interactive reasoning on the input structured query features and time series features, output candidate cost prediction results with confidence evaluation, and input the candidate cost prediction results into the prediction generation layer; The prediction generation layer is used to generate a project cost prediction report based on the candidate cost prediction results, according to the preset project risk management framework; the report includes project cost estimation results, deviation analysis results, and optimization suggestions. The multimodal fusion layer consists of a semantic parsing module, a feature retrieval module, and a modality alignment module; Semantic parsing module: Using a pre-trained language model (such as BERT) network structure based on Transformer, the module takes a cost management request (text sequence) as input, extracts contextual semantic features through a self-attention mechanism, and outputs structured query features (high-dimensional embedding vectors) to capture the core intent and related attributes of the query (such as project size and time range).

[0065] Feature retrieval module: Uses Approximate Nearest Neighbor (ANN) retrieval network structure (such as the vector index implemented in the Faiss library), inputs cost feature vector library (pre-stored historical cost time series data embedding), and retrieves the k most relevant time series features (represented as sequence embedding vectors) through structured query features to supplement historical trend information and enhance the basis for dynamic prediction.

[0066] Modality alignment module: It uses a cross-modal attention mechanism network structure (such as a multi-layer Transformer encoder), takes structured query features and retrieved time series features as input, learns the alignment between features through attention weights, and outputs a modality alignment feature vector of uniform dimension to eliminate modal ambiguity and realize the fusion of text and sequence data.

[0067] The dynamic adaptation layer consists of a cost prediction module, a cross-modal interaction unit, and a confidence assessment unit. Cost prediction module: Using a pre-trained time series prediction model network structure (such as LSTM or GRU network), the time series part of the modality alignment features is input, long-term dependencies are learned through recurrent units, and the initial cost prediction value (e.g., a numerical sequence of future cost estimates) is output to generate the baseline prediction result.

[0068] Cross-modal interaction unit: Using a cross-attention mechanism network structure (such as the Transformer decoder), the structured query features in the modality-aligned features are input to the output of the cost prediction module. The interaction weights are calculated through multi-head attention, and the feature representation after interaction is output. This is used to fuse text and sequence information and improve the adaptability to complex scenarios.

[0069] Confidence assessment unit: Using a probability output layer network structure (such as a softmax classifier), the output of the cross-modal interaction unit is input, the probability distribution of each prediction result is calculated and a confidence score (range of 0-1) is output to quantify the reliability of the prediction and support risk management decisions.

[0070] The prediction generation layer consists of report generator units; Report Generator Unit: Using a Transformer-based sequence generation model network structure (such as the GPT series models), it takes as input the candidate cost prediction results with confidence and the preset project risk management framework (such as input risk thresholds and constraints), generates a text sequence through the decoder layer and attention mechanism, and outputs a structured report containing project cost estimation results (numerical prediction), deviation analysis results (risk assessment compared with actual data), and optimization suggestions (operational guidelines based on risk analysis) to comprehensively support project decision-making.

[0071] The project cost management model deployment module is used by the cost management server to train the project cost management model using the project dataset, the prediction loss function, and the optimization reward function, compress the trained project cost management model, and push the compressed project cost management model to the distributed project management terminal through the federated gateway. After compression, the project cost management model is deployed to the distributed project management terminal through a federated gateway, reducing terminal resource consumption (such as computing and storage) and achieving a lightweight and scalable distributed architecture. In project management scenarios, it ensures that the system can efficiently handle multiple concurrent requests, support large-scale applications, and protect data privacy through a federated mechanism.

[0072] The cost management module is used by the project management terminal to obtain input cost management requests, parse and verify the cost management requests to obtain real-time cost management requirements, input the real-time cost management requirements into the deployed project cost management model to obtain a real-time project cost prediction report, and display the real-time project cost prediction report. The project cost management model optimization module is used to record cost management logs in real time on the project management terminal, including at least the project ID, execution time, real-time cost management requests, real-time project cost prediction reports, and execution feedback. Through a reinforcement learning mechanism, the project cost management model is iteratively optimized using the cost management logs. By using a reinforcement learning mechanism to iteratively optimize the project cost management model using cost management logs, this self-learning design enables the project cost management model to continuously adapt to new data and scenarios in practical applications, improve long-term performance, reduce retraining costs, and enhance the sustainability and accuracy of the technical solution in dynamic project environments.

[0073] The cost management log management module is used by the project management terminal to encrypt and distribute the cost management logs.

[0074] By combining Long Short-Term Memory (LSTM) networks to efficiently process historical project data, extract time-series features, and construct a cost feature vector library, and innovatively employing a cross-modal interaction mechanism to fuse structured queries and time-series data, the prediction accuracy, robustness, and generalization ability are significantly improved. Simultaneously, by leveraging a cost dynamic monitoring agent and a sliding window mechanism to update data in real time, the model's timeliness and responsiveness are ensured, outputting candidate results with confidence assessments and a complete report including cost estimates, bias analysis, and optimization suggestions, thereby enhancing decision reliability and interpretability. The model is continuously iterated and optimized through reinforcement learning mechanisms, and lightweight, scalable distributed applications are achieved through model compression and federated deployment. Data encryption and distributed backup ensure security, ultimately forming an end-to-end automated process that effectively reduces management burden, improves efficiency, and mitigates the risk of project cost overruns.

[0075] The project dataset construction module is specifically used for: The cost management server acquires a large amount of historical project data, which includes structured data and unstructured data. The structured data includes at least budget data and schedule data, and the unstructured data includes at least resource data and external environment data. For the structured data in each of the historical project data, preprocessing is performed including at least the following: duplicate data removal, missing value handling, error data correction, data type conversion, standardization, normalization, and data integration. For the text data in the unstructured data, preprocessing is performed including at least the following: useless character removal, text format unification, word segmentation, part-of-speech tagging, and stop word removal. For the image data in the unstructured data, preprocessing is performed including at least the following: image format unification, image size adjustment, and image enhancement. For the audio data in the unstructured data, preprocessing is performed including at least the following: audio format conversion and audio noise reduction. The preprocessed historical project data are labeled with at least the cost management requirements, project costs, deviation values, reasons for deviations, and optimization suggestions. A project dataset is then constructed based on the labeled historical project data.

[0076] By integrating structured data (such as budgets and schedules) with unstructured data (such as resources, text, images, and audio from the external environment), and customizing hierarchical preprocessing workflows (including data cleaning, format standardization, semantic analysis, and multimedia optimization) for different data types, and combining professional dimensions from the field of cost management (such as reasons for deviations and optimization suggestions) for fine-grained annotation, a high-quality, multi-dimensional project dataset is ultimately generated. This not only significantly improves the consistency and usability of the data, but also significantly reduces the cost of manual intervention through automated processes, providing an accurate and reliable data foundation for subsequent cost prediction, deviation analysis, and decision optimization, thereby enhancing the level of intelligence in project cost management.

[0077] In the initialization module, the feature extraction model is constructed based on a time series processing layer, a correlation capture layer, and a feature output layer; The time series processing layer is constructed based on a time series feature extraction channel, a primary correlation preprocessing channel, and a feature concatenation unit. The time series feature extraction channel is used to extract time-dependent features from the input data through a bidirectional gated recurrent unit. The primary correlation preprocessing channel is used to extract preliminary spatial dependency features from the input data through a one-dimensional convolutional neural network. The feature concatenation unit is used to concatenate the time-dependent features and the preliminary spatial dependency features to obtain time series features. The association capture layer is constructed based on a self-attention feature extraction module and a feature fusion module; the self-attention feature extraction module is used to extract attention-weighted features from time series features through a multi-head self-attention unit; the feature fusion module is used to fuse the attention-weighted features and time series features to obtain a first fused feature, and to filter redundant information from the first fused feature through a first long short-term memory network to obtain the feature association; The feature output layer is used to output the time series features and feature correlations.

[0078] The feature output layer consists of a global average pooling module and a fully connected output module. The global average pooling module uses a global average pooling layer to apply average pooling to the output of the correlation capture layer along the time step dimension, outputting dimensionality-reduced features for aggregating sequence information, extracting global feature representations, and reducing the impact of noise. The fully connected output module uses a fully connected layer (FC) and a Tanh activation function layer. The fully connected layer (output dimension 32) is followed by a Tanh activation function layer (to prevent gradient vanishing). The input is the output of the global average pooling module, which is mapped to the final feature vector (time series features, feature correlations) for non-linear mapping of output features. Tanh activation ensures that the feature values ​​are in the range of [-1,1], improving representational power.

[0079] By innovatively integrating a time-series processing layer, a correlation capture layer, and a feature output layer, efficient feature extraction is achieved. Its core advantages include the efficient fusion of a bidirectional gated recurrent unit (GRU) and a one-dimensional convolutional neural network (CNN) to simultaneously capture temporally dependent features and preliminary spatially dependent features, ensuring feature comprehensiveness and data adaptability. Furthermore, the adoption of a multi-head self-attention mechanism and a feature fusion module, combined with a long short-term memory network (LSTM) to filter redundant information, significantly enhances contextual understanding and the refinement of correlations, improving the model's robustness and convergence speed. The modular overall design also endows the solution with flexible scalability and high output quality, effectively improving the accuracy and efficiency of feature extraction.

[0080] The cost feature vector library update module is specifically used for: The cost management server creates a cost dynamic monitoring agent for monitoring project status data, configures the monitoring parameters of the cost dynamic monitoring agent, and monitors the latest project status data in real time through the cost dynamic monitoring agent. The project status data includes at least budget data, schedule data, resource data, and external environment data. After verifying the status data of each project according to the preset verification rules, the corresponding length of data is selected from the status data of each project through the sliding window mechanism and input into the feature extraction model. The cost feature vector library is dynamically updated based on the time series features output by the feature extraction model.

[0081] By using a cost dynamic monitoring agent, real-time dynamic monitoring of project budget, schedule, resources, and external environment data is achieved. Combined with automated verification rules, data reliability is ensured. Furthermore, a sliding window mechanism and intelligent feature extraction model are used to adaptively analyze time-series data, thereby significantly improving the accuracy and timeliness of cost management. This approach can respond to project risks (such as budget overruns or resource fluctuations) in an instant, and can also achieve self-optimization of predictive capabilities through a continuously updated cost feature vector library. Ultimately, while reducing the cost of manual intervention, it provides high-precision trend insights and resource optimization basis for decision-making.

[0082] The project cost management model deployment module is specifically used for: The cost management server divides the project dataset into a training set, a validation set, and a test set in an 8:1:1 ratio. The project cost management model is trained using the training set, continuously optimizing its parameters using the prediction loss function and the optimization reward function until a preset early termination condition is met. The prediction accuracy is calculated using the validation set to validate the trained project cost management model. The F1 score and confidence level are calculated using the test set to test the validated model. The tested model is compressed using quantization techniques, and the compressed model is pushed in real-time to the distributed project management terminals via a pre-deployed federated gateway.

[0083] Scientific data partitioning, dual-function driven model training, and early stopping mechanisms ensured the accuracy and training efficiency of the prediction model. Multi-index validation testing (accuracy, F1 score, confidence) significantly improved model reliability. Simultaneously, quantization technology was used to compress the model to reduce resource consumption, and real-time secure push of the compressed model was achieved through a federated gateway. Ultimately, lightweight, highly responsive, and scalable intelligent project cost management was realized on the distributed management terminal. This fully automated design not only ensured the robustness and generalization ability of the model but also greatly optimized system resource utilization and business response timeliness.

[0084] While specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and not intended to limit the scope of the present invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A project cost management method combined with a long short-term memory network, characterized in that: Includes the following steps: Step S1: The cost management server acquires a large amount of historical project data, preprocesses the historical project data, and then constructs a project dataset. Step S2: The cost management server extracts time series features and feature relationships from the project dataset using a pre-trained feature extraction model, constructs a cost feature vector library based on each time series feature, and initializes the pre-trained cost prediction model based on each feature relationship. The cost prediction model is constructed based on a feature input layer, a cross-modal interaction layer, and a prediction output layer; The feature input layer is used to extract high-level feature representations from the input structured query features and time series features; the cross-modal interaction layer is used to fuse the high-level feature representations of the structured query features and time series features to obtain a second fused feature, and to perform cross-modal interactive reasoning on the second fused feature through a second long short-term memory network to obtain an enhanced feature; the prediction output layer is used to map the enhanced feature to a candidate cost prediction result with confidence evaluation; Step S3: The cost management server creates a cost dynamic monitoring agent, which monitors the latest project status data in real time. After verifying the project status data, the cost feature vector library is dynamically updated based on the project status data through a sliding window mechanism. Step S4: The cost management server creates a project cost management model based on the multimodal fusion layer, dynamic adaptation layer, and prediction generation layer, and sets the prediction loss function and optimization reward function of the project cost management model. The multimodal fusion layer is used to perform semantic parsing on the input cost management request to obtain structured query features. It simultaneously retrieves time series features from the cost feature vector library through the cost management request. After modally aligning the structured query features with the retrieved time series features, it inputs them into the dynamic adaptation layer. The dynamic adaptation layer is used to call the initialized cost prediction model, perform cross-modal interactive reasoning on the input structured query features and time series features, output candidate cost prediction results with confidence evaluation, and input the candidate cost prediction results into the prediction generation layer; The prediction generation layer is used to generate a project cost prediction report based on the candidate cost prediction results, according to the preset project risk management framework; the report includes project cost estimation results, deviation analysis results, and optimization suggestions. Step S5: The cost management server trains the project cost management model using the project dataset, the prediction loss function, and the optimization reward function. The trained project cost management model is then compressed, and the compressed project cost management model is pushed to the distributed project management terminal through the federated gateway. Step S6: The project management terminal obtains the input cost management request, parses and verifies the cost management request to obtain the real-time cost management request, inputs the real-time cost management request into the deployed project cost management model to obtain the real-time project cost prediction report, and displays the real-time project cost prediction report. Step S7: The project management terminal records cost management logs in real time, including at least the project ID, execution time, real-time cost management requests, real-time project cost forecast reports, and execution feedback. Through a reinforcement learning mechanism, the project cost management model is iteratively optimized using the cost management logs. Step S8: The project management terminal encrypts and performs distributed backup of the cost management log. 2.The project cost management method combined with a long short-term memory network according to claim 1, wherein: Step S1 specifically involves: The cost management server acquires a large amount of historical project data, which includes structured data and unstructured data. The structured data includes at least budget data and schedule data, and the unstructured data includes at least resource data and external environment data. The structured data in each of the historical project data shall be preprocessed, including at least the following: deduplication, missing value handling, error correction, data type conversion, standardization, normalization and data integration. The text data in the unstructured data undergoes preprocessing including at least the removal of useless characters, text formatting standardization, word segmentation, part-of-speech tagging, and removal of stop words. The image data in the unstructured data is subjected to preprocessing including at least image format unification, image size adjustment, and image enhancement. The audio data in the unstructured data undergoes preprocessing including at least audio format conversion and audio noise reduction. The preprocessed historical project data are labeled with at least the cost management requirements, project costs, deviation values, reasons for deviations, and optimization suggestions. A project dataset is then constructed based on the labeled historical project data.

3. The project cost management method combining Long Short-Term Memory (LSTM) networks as described in claim 1, characterized in that: In step S2, the feature extraction model is constructed based on a time series processing layer, a correlation capture layer, and a feature output layer. The time series processing layer is constructed based on a time series feature extraction channel, a primary correlation preprocessing channel, and a feature splicing unit; the time series feature extraction channel is used to extract time-dependent features from the input data through a bidirectional gated loop unit. The primary correlation preprocessing channel is used to extract preliminary spatial dependency features from the input data through a one-dimensional convolutional neural network; The feature splicing unit is used to splice time-dependent features and preliminary spatial-dependent features to obtain time-series features; The association capture layer is constructed based on a self-attention feature extraction module and a feature fusion module; the self-attention feature extraction module is used to extract attention-weighted features from time series features through a multi-head self-attention unit; the feature fusion module is used to fuse the attention-weighted features and time series features to obtain a first fused feature, and to filter redundant information from the first fused feature through a first long short-term memory network to obtain the feature association; The feature output layer is used to output the time series features and feature correlations.

4. The project cost management method combining Long Short-Term Memory networks as described in claim 1, characterized in that: Step S3 specifically involves: The cost management server creates a cost dynamic monitoring agent for monitoring project status data, configures the monitoring parameters of the cost dynamic monitoring agent, and monitors the latest project status data in real time through the cost dynamic monitoring agent. The project status data includes at least budget data, schedule data, resource data, and external environment data. After verifying the status data of each project according to the preset verification rules, the corresponding length of data is selected from the status data of each project through the sliding window mechanism and input into the feature extraction model. The cost feature vector library is dynamically updated based on the time series features output by the feature extraction model.

5. The project cost management method combining Long Short-Term Memory (LSTM) networks as described in claim 1, characterized in that: Step S5 specifically involves: The cost management server divides the project dataset into a training set, a validation set, and a test set in an 8:1:1 ratio. The project cost management model is trained using the training set, continuously optimizing its parameters using the prediction loss function and the optimization reward function until a preset early termination condition is met. The prediction accuracy is calculated using the validation set to validate the trained project cost management model. The F1 score and confidence level are calculated using the test set to test the validated model. The tested model is compressed using quantization techniques, and the compressed model is pushed in real-time to the distributed project management terminals via a pre-deployed federated gateway.

6. A project cost management system incorporating a long short-term memory network, characterized in that: Includes the following modules: The project dataset construction module is used by the cost management server to obtain a large amount of historical project data, and then construct the project dataset after preprocessing the historical project data. The initialization module is used by the cost management server to extract time series features and feature relationships from the project dataset through a pre-trained feature extraction model, construct a cost feature vector library based on each time series feature, and initialize the pre-trained cost prediction model based on each feature relationship. The cost prediction model is constructed based on a feature input layer, a cross-modal interaction layer, and a prediction output layer; The feature input layer is used to extract high-level feature representations from the input structured query features and time series features; the cross-modal interaction layer is used to fuse the high-level feature representations of the structured query features and time series features to obtain a second fused feature, and to perform cross-modal interactive reasoning on the second fused feature through a second long short-term memory network to obtain an enhanced feature; the prediction output layer is used to map the enhanced feature to a candidate cost prediction result with confidence evaluation; The cost feature vector library update module is used by the cost management server to create a cost dynamic monitoring agent, which monitors the latest project status data in real time through the cost dynamic monitoring agent, and after verifying each project status data, dynamically updates the cost feature vector library based on the project status data through a sliding window mechanism. The project cost management model creation module is used by the cost management server to create a project cost management model based on a multimodal fusion layer, a dynamic adaptation layer, and a prediction generation layer, and to set the prediction loss function and optimization reward function of the project cost management model. The multimodal fusion layer is used to perform semantic parsing on the input cost management request to obtain structured query features. It simultaneously retrieves time series features from the cost feature vector library through the cost management request. After modally aligning the structured query features with the retrieved time series features, it inputs them into the dynamic adaptation layer. The dynamic adaptation layer is used to call the initialized cost prediction model, perform cross-modal interactive reasoning on the input structured query features and time series features, output candidate cost prediction results with confidence evaluation, and input the candidate cost prediction results into the prediction generation layer; The prediction generation layer is used to generate a project cost prediction report based on the candidate cost prediction results, according to the preset project risk management framework; the report includes project cost estimation results, deviation analysis results, and optimization suggestions. The project cost management model deployment module is used by the cost management server to train the project cost management model using the project dataset, the prediction loss function, and the optimization reward function, compress the trained project cost management model, and push the compressed project cost management model to the distributed project management terminal through the federated gateway. The cost management module is used by the project management terminal to obtain input cost management requests, parse and verify the cost management requests to obtain real-time cost management requirements, input the real-time cost management requirements into the deployed project cost management model to obtain a real-time project cost prediction report, and display the real-time project cost prediction report. The project cost management model optimization module is used to record cost management logs in real time on the project management terminal, including at least the project ID, execution time, real-time cost management requests, real-time project cost prediction reports, and execution feedback. Through a reinforcement learning mechanism, the project cost management model is iteratively optimized using the cost management logs. The cost management log management module is used by the project management terminal to encrypt and distribute the cost management logs.

7. A project cost management system incorporating a long short-term memory network as described in claim 6, characterized in that: The project dataset construction module is specifically used for: The cost management server acquires a large amount of historical project data, which includes structured data and unstructured data. The structured data includes at least budget data and schedule data, and the unstructured data includes at least resource data and external environment data. The structured data in each of the historical project data shall be preprocessed, including at least the following: deduplication, missing value handling, error correction, data type conversion, standardization, normalization and data integration. The text data in the unstructured data undergoes preprocessing including at least the removal of useless characters, text formatting standardization, word segmentation, part-of-speech tagging, and removal of stop words. The image data in the unstructured data is subjected to preprocessing including at least image format unification, image size adjustment, and image enhancement. The audio data in the unstructured data undergoes preprocessing including at least audio format conversion and audio noise reduction. The preprocessed historical project data are labeled with at least the cost management requirements, project costs, deviation values, reasons for deviations, and optimization suggestions. A project dataset is then constructed based on the labeled historical project data.

8. A project cost management system incorporating a long short-term memory network as described in claim 6, characterized in that: In the initialization module, the feature extraction model is constructed based on a time series processing layer, a correlation capture layer, and a feature output layer; The time series processing layer is constructed based on a time series feature extraction channel, a primary correlation preprocessing channel, and a feature splicing unit; the time series feature extraction channel is used to extract time-dependent features from the input data through a bidirectional gated loop unit. The primary correlation preprocessing channel is used to extract preliminary spatial dependency features from the input data through a one-dimensional convolutional neural network; The feature splicing unit is used to splice time-dependent features and preliminary spatial-dependent features to obtain time-series features; The association capture layer is constructed based on a self-attention feature extraction module and a feature fusion module; the self-attention feature extraction module is used to extract attention-weighted features from time series features through a multi-head self-attention unit; the feature fusion module is used to fuse the attention-weighted features and time series features to obtain a first fused feature, and to filter redundant information from the first fused feature through a first long short-term memory network to obtain the feature association; The feature output layer is used to output the time series features and feature correlations.

9. A project cost management system incorporating a long short-term memory network as described in claim 6, characterized in that: The cost feature vector library update module is specifically used for: The cost management server creates a cost dynamic monitoring agent for monitoring project status data, configures the monitoring parameters of the cost dynamic monitoring agent, and monitors the latest project status data in real time through the cost dynamic monitoring agent. The project status data includes at least budget data, schedule data, resource data, and external environment data. After verifying the status data of each project according to the preset verification rules, the corresponding length of data is selected from the status data of each project through the sliding window mechanism and input into the feature extraction model. The cost feature vector library is dynamically updated based on the time series features output by the feature extraction model.

10. A project cost management system incorporating a long short-term memory network as described in claim 6, characterized in that: The project cost management model deployment module is specifically used for: The cost management server divides the project dataset into a training set, a validation set, and a test set in an 8:1:1 ratio. The project cost management model is trained using the training set, continuously optimizing its parameters using the prediction loss function and the optimization reward function until a preset early termination condition is met. The prediction accuracy is calculated using the validation set to validate the trained project cost management model. The F1 score and confidence level are calculated using the test set to test the validated model. The tested model is compressed using quantization techniques, and the compressed model is pushed in real-time to the distributed project management terminals via a pre-deployed federated gateway.

Images