Project full-period dynamic financial cost prediction and accurate control method and system

Abstract

The invention discloses a project full-cycle dynamic financial cost prediction and precise control method and system, and the method comprises the following steps: S1, obtaining and preprocessing multi-modal original data in a project full life cycle, the multi-modal original data containing structured financial data and unstructured operation data; and S2, carrying out feature extraction and fusion on the preprocessed multi-modal data by adopting a technical scheme of combining cross-modal feature alignment, attention weighted fusion and feature cascade enhancement. The invention relates to the technical field of project financial cost management. According to the project full-period dynamic financial cost prediction and precise control method and system, multi-modal data are comprehensively utilized, features are efficiently fused, a dynamic cost prediction model is constructed, precise staged prediction is achieved, a dynamic real-time cost control strategy can be generated, the hyperbranched risk is effectively reduced, the system integration degree is high, a one-stop scheme is provided for project management, and the project management efficiency is improved. Popularization and application are facilitated.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of project financial cost management, more particularly, to a project full-cycle dynamic financial cost prediction and precise control method and system. BACKGROUND

[0002] In the practice of project management, effective prediction and precise control of financial cost play a decisive role in the successful promotion of the project. However, the traditional method exposes many serious problems in dealing with project financial cost management. On the one hand, there is great limitation in data utilization. The traditional method excessively relies on structured financial data such as project budget, historical cost, fund flow, etc., but ignores unstructured operation data. Unstructured data such as supply chain fluctuation records, employee behavior logs, market dynamic public opinion, equipment operation status data, and geographic spatial data contain a large amount of key information closely related to cost prediction. Failure to fully exploit and utilize these data makes cost prediction lack comprehensiveness and accuracy, and makes it difficult to truly reflect the cost status of the project.

[0003] On the other hand, the traditional method has obvious deficiencies in feature extraction and fusion, model adaptability, and cost control strategy. Different types of data have unique feature representation and semantic space. The traditional method is difficult to effectively extract and fuse the features of multi-modal data, and cannot generate a unified and effective fusion feature vector, thereby limiting the performance of the cost prediction model. At the same time, the project environment is in dynamic change, while the traditional cost prediction model is mostly a static model, which cannot learn the correlation between external factors and internal financial indicators in real time, cannot dynamically adjust the model parameters according to the changes of the project environment, and leads to a large deviation between the prediction results and the actual situation. In addition, the traditional cost control strategy is usually based on fixed cost threshold and risk level, lacks dynamic adjustment mechanism, and is difficult to optimize in real time according to the actual situation of the project and the prediction results, cannot achieve precise control of the financial cost of the project, and increases the risk of project cost overrun. SUMMARY

[0004] The present application aims to provide a project full-cycle dynamic financial cost prediction and precise control method and system, which solves the problem that the existing cost control strategy is usually based on fixed cost threshold and risk level, lacks dynamic adjustment mechanism, is difficult to optimize in real time according to the actual situation of the project and the prediction results, cannot achieve precise control of the financial cost of the project, and increases the risk of project cost overrun.

[0005] The present application achieves the above-mentioned purpose through the following technical solution: a project full-cycle dynamic financial cost prediction and precise control method, comprising the following steps: S1, acquiring and preprocessing multi-modal raw data in the full life cycle of the project, the multi-modal raw data covering structured financial data and unstructured operation data; S2, a technical scheme combining cross-modal feature alignment, attention weighted fusion and feature cascade enhancement is adopted to perform feature extraction and fusion on the preprocessed multi-modal data, and a unified dimension fusion feature vector is generated, and the feature weight is adaptively determined based on data contribution and cost correlation in the fusion process; S3, a dynamic cost prediction model based on deep reinforcement learning is constructed, the fusion feature vector is taken as input, combined with the dynamic change factors of the project environment, the correlation between external factors and internal financial indicators is learned in real time by the DRL agent, the model parameters are dynamically optimized, and the model parameter adjustment threshold is dynamically set according to the project stage characteristics and prediction accuracy requirements; S4, based on the optimized prediction model, the financial cost prediction results of each stage of the project life cycle are output; S5, according to the cost prediction results, combined with the project preset cost threshold and risk level, a dynamic cost control strategy is generated to realize accurate control and real-time adjustment of project financial cost.

[0006] In an embodiment, in S1, the structured financial data includes: At least one of project budget data, historical cost data, fund flow data, and economic index data, and at least one of unstructured operation data including supply chain fluctuation records, employee behavior logs, risk event texts, market dynamic public opinions, equipment operation state data, and geographic space data; The structured financial data is standardized by an extreme value standardization method to eliminate the dimension effect, the unstructured operation data is preprocessed according to types, the text data is encoded by a pre-trained language model, the time-dependent features of the time series data are extracted, and the geographic space data is encoded and mapped into a high-dimensional feature vector. Finally, a standardized structured data matrix and an encoded unstructured data vector set are obtained.

[0007] In an embodiment, the structured financial data standardization expression is: Wherein, is the i-th data of the j-th sample, is the i-th data sample set, is the i-th data of the j-th sample, is the i-th data sample set, , , is the extreme value of the dimension data respectively; The text data is processed by a BERT series model to output a fixed dimension feature vector; The time series data is modeled by a sliding window combined with a long short-term memory network; The geographic space data is converted by GeoHash encoding.

[0008] In an embodiment, in the S2: A multi-branch feature extraction network is constructed, and dedicated extraction branches are designed for different types of data to output single-modal feature vectors of the same dimension. Each modal feature is mapped to the same semantic space through a modal adaptation layer, and modal differences are eliminated based on feature similarity calibration. The fusion weight is dynamically determined based on the correlation between the feature and the cost prediction target and the data reliability factor. The calibrated features and the weighted fusion features are concatenated, and a unified dimension fusion feature vector is output after dimension compression.

[0009] In an embodiment, in the multi-branch feature extraction network, the structured data branch uses a multi-layer perceptron with an activation function and a dropout layer, the text data branch uses a Transformer encoder, the time series data branch uses a bidirectional long short-term memory network, and the geographic spatial data branch uses a graph convolution network. The output dimensions of each branch can be flexibly configured. The fusion weight is first calculated based on the Pearson correlation coefficient, then a data reliability factor is introduced for dynamic adjustment, and finally normalized.

[0010] In an embodiment, in the S3: The state space of the deep reinforcement learning framework includes the fusion feature vector and the dynamic change factor of the project environment, the action space is a set of model parameter adjustment strategies, and the reward function integrates the cost prediction error and the model stability. The DRL agent learns the state-action value function through a deep Q network or other value network, and dynamically adjusts the core parameters of the model. An adaptive learning mechanism is introduced, and a threshold of environmental change fluctuation and a threshold of parameter adjustment amplitude are set. When the environmental change exceeds the threshold, an emergency update process is triggered.

[0011] In an embodiment, the threshold of environmental change fluctuation is initially set to the maximum value of the historical factor standard deviation multiplied by a preset multiple, and then dynamically adjusted according to the project stage. The threshold of parameter adjustment amplitude is calculated based on the allowed range of prediction error to avoid model oscillation. The emergency update process uses a priority experience replay mechanism to accelerate optimization. The core parameters of the prediction model include: modal fusion weight, inter-layer weight of the feature extraction network, and regression coefficient of the cost prediction head.

[0012] In an embodiment, in the S4: The prediction model is divided into special branches according to the project stage, and each branch is configured with a dedicated prediction head and a modal weight bias. The weights of historical structured and geographic spatial data are reinforced in the planning stage, and the budget interval and cost allocation prediction are output. The execution stage reinforces the timing and weight of the text data, outputs the cost deviation and trend prediction; The final stage balances the weight of each mode, and outputs the final settlement prediction and deviation reason analysis; The prediction result is presented in the form of an interval, which is determined based on the standard deviation of the prediction error and the confidence level coefficient.

[0013] In an embodiment, in S5: The cost warning threshold includes a basic threshold and a dynamic correction threshold, the basic threshold is determined based on historical data or previous budget, and the dynamic correction introduces a cost sensitivity factor; According to the relationship between the predicted value and the threshold, the low, medium and high risk levels are divided, and the maintenance configuration, resource optimization and emergency control strategy are generated correspondingly.

[0014] A project full-cycle dynamic financial cost prediction and accurate control system is applied to the project full-cycle dynamic financial cost prediction and accurate control method described above, and the system comprises: Data acquisition and preprocessing unit, multi-modal data fusion unit, dynamic cost prediction unit, prediction result output unit, prediction result output unit; The data acquisition and preprocessing unit is used for acquiring and preprocessing multi-modal original data in the project full life cycle, and the multi-modal original data covers structured financial data and unstructured operation data; The multi-modal data fusion unit is used for combining the technical solutions of cross-modal feature alignment, attention weighted fusion and feature cascade enhancement to extract and fuse the preprocessed multi-modal data, generate a unified dimension fusion feature vector, and the feature weight is self-adaptive determined based on data contribution and cost correlation in the fusion process; The dynamic cost prediction unit is used for constructing a dynamic cost prediction model based on deep reinforcement learning, taking the fusion feature vector as input, combining the dynamic change factors of the project environment, learning the correlation between external factors and internal financial indicators in real time through the DRL agent, dynamically optimizing the prediction model parameters, and the model parameter adjustment threshold is dynamically set according to the project stage characteristics and prediction accuracy requirements; The prediction result output unit is used for outputting the financial cost prediction result of each stage of the project full life cycle based on the optimized prediction model; The prediction result output unit is used for generating a dynamic cost control strategy according to the cost prediction result, combining the project preset cost threshold and risk level, and realizing the accurate control and real-time adjustment of the project financial cost.

[0015] The beneficial effects of the present application are: 1. Obtain structured financial data and unstructured operation data throughout the project life cycle, and perform comprehensive preprocessing, fully utilizing the information contained in various types of data, providing a rich data basis for accurate cost prediction.

[0016] 2. The technical scheme combines cross-modal feature alignment, attention weighted fusion and feature cascade enhancement, constructs a multi-branch feature extraction network, designs exclusive extraction branches for different types of data, effectively extracts and fuses the features of multi-modal data, generates a unified dimension fusion feature vector, and the feature weight is self-adaptive in the fusion process based on data contribution and cost correlation, which improves the quality and effectiveness of the fusion feature and enhances the input feature representation ability of the cost prediction model.

[0017] 3. A dynamic cost prediction model based on deep reinforcement learning is constructed, the fusion feature vector is taken as the input, the dynamic change factors of the project environment are combined, the correlation between external factors and internal financial indicators is learned in real time by the DRL agent, the model parameters are dynamically optimized, and the model parameter adjustment threshold is dynamically set according to the project stage characteristics and prediction accuracy requirements, so that the model can adapt to the dynamic changes of the project environment, and the accuracy and real-time performance of the cost prediction are improved.

[0018] 4. The prediction model is divided into special branches according to the project stages, and each branch is configured with a special prediction head and a modal weight bias. The weight of different types of data is strengthened according to the characteristics of different stages, and the prediction results meeting the requirements of each stage are output. The weight of historical structured and geographic spatial data is strengthened in the planning stage, the weight of time series and text data is strengthened in the execution stage, and the weight of each modal is balanced in the closing stage. The prediction result is presented in the form of interval, which is determined based on the prediction error standard deviation and the confidence level coefficient. The financial cost of each stage of the project is more accurately reflected.

[0019] 5. According to the cost prediction result, combined with the preset cost threshold and risk level of the project, a dynamic cost control strategy is generated. The cost warning threshold includes a basic threshold and a dynamic correction threshold. According to the relationship between the predicted value and the threshold, the risk level is divided into low, medium and high, and the corresponding maintenance configuration, resource optimization and emergency control strategy is generated. The accurate control and real-time adjustment of the project financial cost are realized, the risk of project cost overrun is effectively reduced, and the economic benefit of the project is improved. BRIEF DESCRIPTION OF DRAWINGS

[0020] The drawings described herein are used to provide further understanding of the present application, and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application, and do not constitute an improper limitation on the present application. In the drawings: Figure 1 The overall method flowchart of the present application; Figure 2 The data preprocessing flowchart of the present application; Figure 3 A dynamic cost prediction flowchart of the present application; Figure 4 A system block diagram of the present application. DETAILED DESCRIPTION

[0021] The present application will be further described in detail below with reference to the accompanying drawings. It is necessary to point out here that the following detailed description is only used to further illustrate the present application and cannot be understood as limiting the scope of protection of the present application. Those skilled in the art can make some non-essential improvements and adjustments to the present application according to the above application content.

[0022] Example 1: Please refer to Figures 1-3 The present application provides a technical solution: a project full-cycle dynamic financial cost prediction and precision control method, which comprises the following steps: S1, obtaining and preprocessing multi-modal original data in the whole life cycle of the project, the multi-modal original data comprising structured financial data and unstructured operation data; Wherein, the project full life cycle refers to the whole process of the project from start-up, planning, execution, monitoring to completion, covering all stages and activities from the beginning to the end of the project; in the multi-modal original data, the structured financial data has a clear data structure and format, and is usually stored in the form of a table for easy computer processing and analysis, such as income, cost, profit and other data in financial statements; the unstructured operation data has no fixed data structure and is relatively free in form, such as text records, images, audio and other data in the project operation process, which may contain various detailed information in the project execution process; the preprocessing is to clean, convert, normalize and other operations on the obtained original data to remove noise data, fill in missing values, unify data format, etc., so that the data is more suitable for subsequent analysis and processing; S2, using the technical means of cross-modal feature alignment, attention weighted fusion and feature cascade enhancement, extracting and fusing the features of the preprocessed multi-modal data to generate a unified dimension fusion feature vector, wherein the feature weight in the fusion process is determined based on the data contribution and cost correlation self-adaptively; Among them, the cross-modal feature alignment is to match and correspond the data features of different modalities, so that they have comparability and consistency in the feature space, so as to facilitate the subsequent fusion operation; the attention weighted fusion is to assign different weights to different data modalities or features according to their importance to the cost prediction, and then to fuse them, so that the data modalities or features with high importance are given greater weight in the fusion process, thereby highlighting their influence on the prediction result; the feature concatenation enhancement is to connect the extracted different features in a certain order to form a longer feature vector, which can integrate the information of different features and enhance the expression ability of the features; the fused feature vector is a vector with unified dimension obtained after feature extraction and fusion, which integrates various feature information of multi-modal original data and can more comprehensively reflect the characteristics and status of the project; the data contribution degree is to measure the contribution degree of different data modalities or features to the final prediction result, and the data with high contribution degree will be given greater weight in the fusion process; the cost correlation refers to the correlation between the data and the project financial cost, and the data with strong cost correlation will be considered in the fusion process; the adaptive determination is to automatically adjust and determine the feature weight according to the characteristics of the data itself and the actual situation of the project, instead of using fixed weight values; S3, a dynamic cost prediction model based on deep reinforcement learning is constructed, the fused feature vector is taken as the input, the dynamic change factors of the project environment are combined, the correlation between external factors and internal financial indicators is learned in real time by the DRL agent, the model parameters are dynamically optimized, and the model parameter adjustment threshold is dynamically set based on the project stage characteristics and the prediction accuracy requirement; Among them, deep reinforcement learning is a machine learning method that combines deep learning and reinforcement learning. Deep learning is used to process high-dimensional input data and extract features, while reinforcement learning learns the optimal strategy through the interaction between the agent and the environment to achieve a specific goal; the dynamic cost prediction model can adjust the prediction results of the cost prediction model in real time according to the changes of the project environment and internal financial indicators to adapt to the changing situation throughout the project life cycle; the dynamic change factor of the project environment is an index or factor that reflects the change of the environment in which the project is located, which will have an impact on the financial cost of the project; the DRL agent is the subject that performs actions and obtains feedback from the environment in the deep reinforcement learning framework. It learns the optimal strategy by trying different actions and learning from the feedback (reward or punishment) of the environment to optimize the cost prediction; the correlation between external factors and internal financial indicators: the mutual influence and connection between external factors and internal financial indicators of the project; the model parameter adjustment threshold is used to control the adjustment range of the model parameters, and when the deviation between the model prediction results and the actual values exceeds the threshold, the model parameters will be adjusted; the project stage characteristics are the different characteristics and needs of the project in different stages, such as the start-up stage may pay more attention to resource investment and planning, and the end stage may focus more on cost control and benefit evaluation; the prediction accuracy requirement is the requirement for the accuracy of the cost prediction results, and different project stages or different application scenarios may have different requirements for the prediction accuracy; S4, output the financial cost prediction results of each stage of the project life cycle based on the optimized prediction model; Among them, the optimized prediction model is a cost prediction model that has been dynamically optimized by the deep reinforcement learning agent, which can more accurately reflect the trend of the project's financial cost; the financial cost prediction result is the numerical value or result obtained by predicting the financial cost of each stage of the project life cycle based on the optimized prediction model, including the cost amount, cost trend, etc. S5, according to the cost prediction results, combining the project preset cost threshold and risk level, generate dynamic cost control strategy, realize the accurate control and real-time adjustment of project financial cost; Among them, the project preset cost threshold is the upper or lower limit of the cost set in advance before the project starts, which is used to monitor whether the project cost exceeds the reasonable range; the risk level is the division of the risk degree of the project according to the risk factors that the project may face, and different risk levels correspond to different countermeasures and cost control strategies; the dynamic cost control strategy is a strategy and method to control the project financial cost generated in real time according to the cost prediction results, preset cost threshold and risk level, to ensure that the project cost is within a controllable range and to maximize the economic benefits of the project; accurate control and real-time adjustment can timely and accurately adjust the cost control strategy according to the actual situation of the project and the change of the cost, to ensure the effectiveness and timeliness of the cost control.

[0023] It should be noted that when using this system, acquiring and preprocessing multimodal raw data can comprehensively cover project information, laying the foundation for accurate analysis. Cross-modal fusion technologies effectively integrate different types of data, generating a unified fusion feature vector to fully explore the value of the data. Furthermore, the adaptive determination of feature weights enhances the rationality and effectiveness of data utilization. Constructing a dynamic model based on deep reinforcement learning can learn the correlation between external and internal factors in real time, dynamically optimize parameters, adapt to changes in the project environment, and improve prediction accuracy. The dynamic setting of parameter adjustment thresholds also enhances the flexibility of the model. Based on the optimized model, the output of prediction results at each stage can provide a reliable basis for cost control. Dynamic strategies can be generated based on the prediction results. Combined with preset thresholds and risk levels, accurate and real-time control of financial costs can be achieved, effectively reducing project cost risks and ensuring the smooth progress and economic benefits of the project.

[0024] In one embodiment, acquiring and preprocessing multimodal raw data throughout the project's entire lifecycle includes: Acquire multimodal raw data at each stage of the project's entire lifecycle. Structured financial data includes project budget data, historical cost data, cash flow data, and economic index data. Unstructured operational data includes supply chain fluctuation records, employee behavior logs, risk event texts, market dynamics and public opinion, equipment operating status data, and geospatial data. The structured financial data is standardized using the following expression: in, For the first Sample No. Item data, For the first Item data sample set, , These represent the minimum and maximum values ​​of the data in this dimension, respectively. Classification and preprocessing of unstructured operational data: For text-based data: The BERT-Base model is used for word segmentation, word embedding, and context feature encoding, outputting a text feature vector with a dimension of 768. For time-series data: Time segment features are extracted using a sliding window, and a long short-term memory network is used to model time-series dependencies, outputting a time-series feature vector; Geospatial data: GeoHash encoding is used to convert spatial coordinates into strings, and then the data is mapped into high-dimensional spatial feature vectors through an embedding layer; Finally, a standardized structured data matrix is ​​obtained. and encoded unstructured data vector set These correspond to text, time series, and geospatial feature vectors, respectively. The number of structured data samples, For structured data feature dimensions.

[0025] This design allows for the acquisition of multimodal data across all stages of the project's lifecycle. Structured financial data is standardized, and unstructured data is preprocessed and categorized. This comprehensive data collection provides rich material for subsequent analysis. Standardized processing of structured data eliminates the influence of units, making the data comparable. Categorized preprocessing of unstructured data transforms different types of data into a unified, processable format, such as encoding text into vectors and extracting features from time-series data. This facilitates subsequent feature extraction and fusion, laying a solid foundation for accurate cost prediction and improving data utilization efficiency and model accuracy.

[0026] In one embodiment, feature extraction and fusion are performed on the preprocessed multimodal data to generate a fused feature vector of uniform dimension, including: Construct a multi-branch feature extraction network and design dedicated extraction branches for different types of data: Structured data branch: A 3-layer multilayer perceptron with ReLU activation function is used, and a dropout layer is employed to prevent overfitting. This is applied to the standardized structured data matrix. Perform deep feature extraction and output structured feature vectors. ; Text data branch: Using a Transformer encoder to process text feature vectors Perform semantic feature enhancement and output text feature vectors. ; Temporal data branch: Bidirectional LSTM is used to process temporal feature vectors. Perform bidirectional temporal dependency capture and output temporal feature vectors. ; Geospatial data branch: Employing graph convolutional networks to process spatial feature vectors Perform spatial correlation modeling and output spatial feature vectors. ; in To unify the feature dimension, it is set to 256; Employing cross-modal feature alignment techniques to eliminate modal differences: A modality adaptation layer is constructed, which maps the features of each modality to the same semantic space through linear transformation. The formula for the adaptation layer is as follows:

[0027] in, Indicates the modal type, For adapting layer parameters; Calculate the cosine similarity between each modal feature and the structured feature, and perform feature calibration based on the similarity. The calibration formula is as follows:

[0028] in, For the first The angle between the modal and the structured mode; The fusion weights of each modality feature are calculated using an attention-weighted fusion mechanism. The rules for determining the fusion weights are as follows: Basic weight calculation: Pearson correlation coefficient based on feature and cost prediction objectives Determine the initial weights and the basic weights of the structured features. Other modal basic weights ,in This represents the average correlation coefficient between the corresponding modal characteristics and the cost objective; Dynamic adjustment rules: Introducing data reliability factors (structured data) Scoring is determined based on data completeness and accuracy; text data Determined based on text sentiment polarity consistency score; time series data Determined based on data sampling frequency and continuity score; geospatial data (Based on spatial data accuracy scoring), the final fusion weight is:

[0029] Feature cascading enhancement technology is used to improve the fusion effect: Weighted fusion features are calculated based on fusion weights. ; The modal-calibrated features are concatenated with the weighted fusion features to obtain the enhanced features. ; A single-layer MLP is used to compress the dimensionality of the cascaded features, outputting a fused feature vector with a unified dimension. ,in To unify the feature dimension to 256.

[0030] This design constructs a multi-branch network to extract features from different modalities, uses cross-modal alignment techniques to eliminate differences, employs attention-weighted fusion to determine weights, and uses cascaded enhancements to improve performance. It fully mines the unique information of each modality, and through adaptation layers and calibration, enables effective interaction of different modalities in a unified semantic space. The attention mechanism rationally allocates weights, highlights important modal features, and cascaded enhancements fuse multiple features to improve information richness. Finally, it outputs a unified dimension vector, which is convenient for model processing, improves the ability to characterize project cost features, and enhances prediction accuracy.

[0031] In one embodiment, a dynamic cost prediction model based on deep reinforcement learning is constructed. Using fused feature vectors as input and incorporating dynamic factors affecting the project environment, a DRL agent learns in real-time the correlation between external factors and internal financial indicators, dynamically optimizing the prediction model parameters. This includes: Construct a deep reinforcement learning framework and define the state space. To fuse feature vectors The project environment dynamics change factors include real-time market volatility coefficient, supply chain risk level, policy adjustment factor, raw material price index, and natural disaster early warning level. Define action space A set of parameter tuning strategies for the prediction model, including model weight update magnitude, feature selection threshold, learning rate adjustment value, and number of attention head activations; Define reward function This is an indicator of the overall cost prediction error and model stability, namely... ,in To predict cost values ​​for the model, This is the actual cost value. For the variance of model parameter fluctuations, This is the balance coefficient, with a value of 0.1. By fusing feature vectors As input to the DRL agent, the state-action value function is learned through a deep Q-network. The core parameters of the prediction model are dynamically adjusted. These core parameters include the fusion weights of each modality, the inter-layer weights of the feature extraction network, and the regression coefficients of the cost prediction head. An adaptive learning mechanism is introduced to set a threshold for fluctuations in project environmental change factors. The rules for determining it are as follows: The standard deviation of each variable factor was calculated based on the project's historical environmental data. , Represents a single environmental factor, initial threshold. Take the maximum value that is 1.5 times the standard deviation of all factors; Dynamically adjust according to the characteristics of each project stage: Planning stage (High environmental stability, relaxed threshold). Execution phase (The environment is highly dynamic, so the threshold should be tightened.) Final stage (The environment stabilizes, and the baseline threshold is restored.) When the fluctuation range of environmental change factors in a project exceeds the threshold When this happens, the emergency update process of the DRL agent is triggered, and the model parameter optimization is accelerated through the priority experience replay mechanism to ensure that the prediction model can quickly adapt to sudden environmental changes. Set the threshold for adjusting model parameters The rules for determining it are as follows: Based on the allowable range of prediction error (70% of the industry standard allowable error), through calculate( For model parameters, (This refers to the maximum partial derivative of the predicted value with respect to the parameter), and the parameter adjustment range must not exceed [a certain value]. To avoid model oscillations.

[0032] This design constructs a deep reinforcement learning framework, defining state, action, and reward spaces. Using fused features as input, it learns and adjusts parameters through DQN, introducing an adaptive mechanism to learn the correlation between external and internal factors in real time, dynamically optimizing parameters to enable the model to adapt to changes in the project environment. The adaptive learning mechanism accelerates optimization during sudden environmental changes based on environmental change thresholds and parameter adjustment thresholds, avoiding model oscillations, ensuring the stability and accuracy of the prediction model, and timely reflecting project cost dynamics, providing a reliable basis for cost control.

[0033] In one embodiment, the financial cost forecast results for each stage of the project's entire lifecycle are output based on the optimized forecasting model, including: The optimized prediction model is divided into three specialized prediction branches: planning, execution, and closing phases. Each branch is configured with a dedicated prediction head and modal weight bias based on the cost characteristics of the corresponding project phase. Planning phase: Strengthen the weight of historical structured data and geospatial data, increase the weight coefficient by 1.2 times, adopt ridge regression for prediction head, and set the regularization coefficient to 0.01. Based on the initial fusion characteristics and historical data of similar projects, output the overall cost budget range of the project and the predicted cost allocation values ​​for each sub-task. Execution phase: The weights of time-series data and text data are strengthened, with the weight coefficient increased by 1.3 times. The prediction head adopts gradient boosting regression tree, and dynamic multimodal data during the project execution process is received in real time. The prediction parameters are continuously optimized through DRL agent, and the phased cost deviation prediction and subsequent cycle cost trend prediction are output. Final stage: Balance the weights of each modality, use multilayer perceptron regression for the prediction head, integrate data from the entire project lifecycle, and output the final project cost settlement prediction value and the analysis results of cost savings / overruns; The prediction results for each stage are output in interval prediction form. The formula for calculating the prediction interval is as follows:

[0034] in The standard deviation of the prediction error. The coefficients are set based on the confidence level (at a 95% confidence level). ).

[0035] This design divides the prediction model into stages, with each stage having its own prediction head and weight bias, outputting interval prediction results. Targeting the cost characteristics of different stages, the dedicated settings more accurately capture the cost patterns of each stage. The planning stage considers historical and geographical factors, the execution stage focuses on time series and textual data, and the closing stage integrates data from the entire cycle. Interval prediction provides a cost range, taking into account uncertainties, making the prediction results more practical and flexible, and providing a comprehensive and accurate reference for cost decisions at each stage of the project.

[0036] In one embodiment, based on cost forecast results and combined with the project's preset cost threshold and risk level, a dynamic cost control strategy is generated to achieve precise control and real-time adjustment of the project's financial costs, including: Preset cost warning thresholds for each stage of the project's entire lifecycle. , in, The threshold for the number of project phases is determined by the following rules: Base threshold The budget is determined based on 1.1 times the historical average cost of similar projects during the planning phase, 1.05 times the budget allocated during the execution phase during the execution phase, and the sum of the accumulated actual costs and remaining budget during the closing phase during the closing phase. Dynamic adjustment: Introducing a cost-sensitivity factor , For the project phase, Corrected threshold (The higher the sensitivity, the more lenient the threshold, to avoid over-warning.) The criteria for classifying risk levels are established, and the rules for determining them are as follows: Low risk: Forecast value ,in Based on the determination of the project's risk tolerance, conservative projects Balanced projects radical projects ; Medium risk: Predicted value ,in Budget baseline; High risk: Forecast value ; When the forecast results are at a low-risk level, a control strategy is generated to maintain the existing resource allocation, with a focus on monitoring the stability of key cost drivers; When the forecast results are at a medium-risk level, resource optimization and adjustment strategies are generated, including supply chain supplier replacement suggestions, personnel allocation optimization plans, and budget reduction ratios for non-core tasks. When the prediction results are at a high-risk level, emergency cost control strategies are generated, including suspending non-essential expenditures, activating standby funding plans, renegotiating contract terms, and rapidly iterating the prediction model through DRL agents to evaluate the effectiveness of the control strategies.

[0037] This design presets cost warning thresholds and sets risk level standards. Based on the forecast results, it generates different control strategies and dynamically adjusts the thresholds according to the characteristics of different project stages. This more reasonably reflects cost risks, clarifies risk level classifications, and makes cost control more targeted. Corresponding strategies are generated for different risk levels: low risk maintains the status quo, medium risk optimizes resources, and high risk requires emergency management. This enables precise control and real-time adjustment of project financial costs, effectively reduces the risk of cost overruns, and ensures the achievement of project financial goals.

[0038] Example 2 Please see Figure 4 A dynamic financial cost forecasting and precise control system for the entire project lifecycle, applied to the aforementioned dynamic financial cost forecasting and precise control method for the entire project lifecycle, the system includes: Data acquisition and preprocessing unit, multimodal data fusion unit, dynamic cost prediction unit, prediction result output unit, prediction result output unit; The data acquisition and preprocessing unit is used to acquire and preprocess multimodal raw data throughout the entire project lifecycle. The multimodal raw data includes structured financial data and unstructured operational data. The multimodal data fusion unit is used to perform feature extraction and fusion on preprocessed multimodal data by adopting a technical solution that combines cross-modal feature alignment, attention-weighted fusion and feature concatenation enhancement, and generates a fusion feature vector of a unified dimension. During the fusion process, the feature weights are adaptively determined based on the correlation between data contribution and cost. The dynamic cost prediction unit is used to build a dynamic cost prediction model based on deep reinforcement learning. It takes the fused feature vector as input, combines the dynamic change factors of the project environment, and learns the correlation between external factors and internal financial indicators in real time through the DRL agent to dynamically optimize the prediction model parameters. The model parameter adjustment threshold is dynamically set according to the characteristics of the project stage and the prediction accuracy requirements. The prediction result output unit is used to output the financial cost prediction results for each stage of the project's entire life cycle based on the optimized prediction model. The prediction result output unit is used to generate a dynamic cost control strategy based on the cost prediction results, combined with the project's preset cost threshold and risk level, so as to achieve precise control and real-time adjustment of the project's financial costs.

[0039] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0040] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for dynamic financial cost forecasting and precise control throughout the entire project lifecycle, characterized in that, Includes the following steps: S1. Obtain and preprocess multimodal raw data throughout the entire project lifecycle, wherein the multimodal raw data includes structured financial data and unstructured operational data; S2. A technical solution combining cross-modal feature alignment, attention-weighted fusion, and feature concatenation enhancement is adopted to extract and fuse features from preprocessed multimodal data, generating a unified dimension fused feature vector. During the fusion process, the feature weights are adaptively determined based on the correlation between data contribution and cost. S3. Construct a dynamic cost prediction model based on deep reinforcement learning. Take the fused feature vector as input, combine it with the dynamic change factors of the project environment, and learn the correlation between external factors and internal financial indicators in real time through the DRL agent. Dynamically optimize the prediction model parameters. The model parameter adjustment threshold is dynamically set according to the characteristics of the project stage and the prediction accuracy requirements. S4. Based on the optimized prediction model, output the financial cost prediction results for each stage of the project's entire life cycle; S5. Based on the cost forecast results, combined with the project's preset cost threshold and risk level, a dynamic cost control strategy is generated to achieve precise control and real-time adjustment of the project's financial costs.

2. The method for dynamic financial cost forecasting and precise control throughout the entire project lifecycle as described in claim 1, characterized in that, In S1, the structured financial data includes: At least one of the following: project budget data, historical cost data, cash flow data, and economic index data; unstructured operational data includes at least one of the following: supply chain fluctuation records, employee behavior logs, risk event texts, market dynamics and public opinion, equipment operating status data, and geospatial data. The extreme value standardization method is used to eliminate the influence of dimensions for structured financial data, and unstructured operational data is preprocessed according to type. Text data is encoded through a pre-trained language model, time-dependent features are extracted from time-series data, and geospatial data is encoded and mapped into high-dimensional feature vectors. Finally, a standardized structured data matrix and an encoded unstructured data vector set are obtained.

3. The method for dynamic financial cost forecasting and precise control throughout the entire project lifecycle as described in claim 2, characterized in that: The standardized expression for structured financial data is: in, For the first Sample No. Item data, For the first Item data sample set, , These are the extreme values ​​of the data in this dimension; Text data is processed using BERT series models, which output fixed-dimensional feature vectors. Time-series data are modeled using a sliding window combined with a long short-term memory network; Geospatial data is converted using GeoHash encoding.

4. The method for dynamic financial cost forecasting and precise control throughout the entire project lifecycle as described in claim 1, characterized in that, In S2: Construct a multi-branch feature extraction network, design dedicated extraction branches for different types of data, and output single-modal feature vectors of the same dimension; The modality adaptation layer maps the features of each modality to the same semantic space, and eliminates modality differences based on feature similarity calibration; The fusion weights are dynamically determined based on the correlation between features and cost prediction targets and data reliability factors. The calibrated features are concatenated with the weighted fused features, and a unified dimensional fused feature vector is output after dimensional compression.

5. The method for dynamic financial cost forecasting and precise control throughout the entire project lifecycle as described in claim 4, characterized in that: In the multi-branch feature extraction network, the structured data branch uses a multilayer perceptron with activation functions and dropout layers, the text data branch uses a Transformer encoder, the time series data branch uses a bidirectional long short-term memory network, and the geospatial data branch uses a graph convolutional network. The output dimensions of each branch can be flexibly configured. The fusion weights are first calculated based on the Pearson correlation coefficient, then dynamically adjusted by introducing a data reliability factor, and finally obtained through normalization.

6. The method for dynamic financial cost forecasting and precise control throughout the entire project lifecycle as described in claim 1, characterized in that, In S3: The state space of a deep reinforcement learning framework includes a fusion of feature vectors and dynamic factors of the project environment; the action space is a set of model parameter adjustment strategies; and the reward function integrates cost prediction error and model stability. DRL agents learn state-action value functions through deep Q-networks or other value networks and dynamically adjust the core parameters of the model. An adaptive learning mechanism is introduced, and thresholds for environmental change fluctuations and parameter adjustment magnitudes are set. When environmental changes exceed the thresholds, an emergency update process is triggered.

7. The method for dynamic financial cost forecasting and precise control throughout the entire project lifecycle as described in claim 6, characterized in that: The initial threshold for environmental change fluctuations is set at the maximum value of a preset multiple of the historical factor standard deviation, and then dynamically adjusted according to the project stage. The parameter adjustment threshold is calculated based on the allowable range of prediction error to avoid model oscillation. The emergency update process employs a priority experience replay mechanism to accelerate optimization; The core parameters of the prediction model include: modality fusion weights, inter-layer weights of the feature extraction network, and cost prediction head regression coefficients.

8. The method for dynamic financial cost forecasting and precise control throughout the entire project lifecycle as described in claim 1, characterized in that, In S4: The prediction model is divided into specialized branches according to project phases, and each branch is configured with a dedicated prediction head and modal weight bias. During the planning phase, historical structured and geospatial data weights are strengthened to output budget ranges and cost allocation forecasts. During the execution phase, the weighting of time-series and text data is strengthened, and cost deviation and trend prediction are output. In the final stage, the weights of each modality are balanced, and the final settlement prediction and deviation analysis are output. The prediction results are presented in interval form and are determined based on the standard deviation of the prediction error and the confidence level coefficient.

9. The method for dynamic financial cost forecasting and precise control throughout the entire project lifecycle as described in claim 1, characterized in that, In S5: The cost warning threshold includes a basic threshold and a dynamic adjustment threshold. The basic threshold is determined based on historical data or previous budgets, while the dynamic adjustment introduces a cost sensitivity factor. Based on the relationship between predicted values ​​and thresholds, risk levels are divided into low, medium, and high, and corresponding maintenance, resource optimization, and emergency control strategies are generated.

10. A dynamic financial cost forecasting and precise control system for the entire project lifecycle, characterized in that, The system, applied to the method for dynamic financial cost forecasting and precise control throughout the project lifecycle as described in any one of claims 1-9, comprises: Data acquisition and preprocessing unit, multimodal data fusion unit, dynamic cost prediction unit, prediction result output unit, prediction result output unit; The data acquisition and preprocessing unit is used to acquire and preprocess multimodal raw data throughout the entire project lifecycle. The multimodal raw data includes structured financial data and unstructured operational data. The multimodal data fusion unit is used to perform feature extraction and fusion on preprocessed multimodal data by adopting a technical solution that combines cross-modal feature alignment, attention-weighted fusion and feature concatenation enhancement, and generates a fusion feature vector of a unified dimension. During the fusion process, the feature weights are adaptively determined based on the correlation between data contribution and cost. The dynamic cost prediction unit is used to construct a dynamic cost prediction model based on deep reinforcement learning. It takes the fused feature vector as input, combines the dynamic change factors of the project environment, and learns the correlation between external factors and internal financial indicators in real time through the DRL agent to dynamically optimize the prediction model parameters. The model parameter adjustment threshold is dynamically set according to the characteristics of the project stage and the prediction accuracy requirements. The prediction result output unit is used to output the financial cost prediction results for each stage of the project's entire life cycle based on the optimized prediction model. The prediction result output unit is used to generate a dynamic cost control strategy based on the cost prediction results, combined with the project's preset cost threshold and risk level, so as to achieve precise control and real-time adjustment of the project's financial costs.

Images