AI-based EPC enterprise financial ticket tax management method, apparatus and device, and storage medium

Abstract

The invention relates to an EPC enterprise financial ticket tax management method and device based on AI, equipment and a storage medium. The method comprises the steps of obtaining applicable laws and regulations, applicable tax rates and applicable deduction conditions of an EPC enterprise according to current tax policies and regulations and enterprise conditions; extracting structured data, semi-structured data and non-structured data from a multi-source data source in the EPC enterprise, and performing cleaning and normalization preprocessing to obtain preprocessed data; training a machine learning model according to the applicable laws and regulations, the applicable tax rates, the applicable deduction conditions and the preprocessed data; performing multi-dimensional evaluation on the machine learning model; and when new project data is obtained, tax payable is obtained through the evaluated machine learning model, and a tax planning scheme is generated. According to the method, intelligent advancement of full-life-cycle tax management of the EPC project is realized through deep fusion machine learning; the key tax management problems of large tax cost prediction deviation, difficult real-time tax calculation, incomplete economic activity analysis and the like are effectively solved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence and tax management technology, and in particular to an AI-based EPC enterprise invoice tax management method, device, equipment and storage medium. Background Art

[0002] In recent years, the country has vigorously promoted the Engineering Procurement Construction (EPC) model. Most industrial EPC general contractors have been transformed from design companies, and construction and municipal projects are also dominated by design companies. These companies have improved their comprehensive capabilities and effectively managed projects by leveraging their advantages in technology, construction period, and cost management to achieve goals such as shortened construction period, quality assurance, and investment control.

[0003] As the EPC model is widely used in the engineering construction field, these companies face many challenges in tax management:

[0004] Difficulty in data integration: Multi-source heterogeneous data (such as ERP and financial systems) lack unified standards, and incompatible formats lead to data loss and poor consistency;

[0005] Inaccurate tax forecasting: Traditional manual forecasting relies on empirical formulas, which results in large deviations and makes it difficult to dynamically respond to policy changes.

[0006] Inefficient processing of unstructured data: Unstructured data such as contract texts and invoice images requires manual processing, which is inefficient and prone to errors.

[0007] Complex application of tax rules: Dynamic policies across regions and multiple tax types are difficult to match in real time, hindering tax planning and compliance.

[0008] In summary, although the existing EPC enterprise financial invoice and tax management system attempts to optimize tax management through data extraction and simple analysis, it lacks intelligent means and cannot meet the scale and dynamic needs of EPC projects. Summary of the Invention

[0009] Based on this, it is necessary to provide an AI-based EPC enterprise invoice tax management method, device, equipment and storage medium to address the above technical problems, which can intelligently realize efficient integration and analysis of data, precise tax forecasting and planning, and accurate application of complex tax rules.

[0010] An AI-based EPC enterprise invoice tax management method, the method comprising:

[0011] Obtain the applicable regulations, applicable tax rates and applicable exemption conditions for EPC companies based on current tax policies and regulations and the company's circumstances;

[0012] Extracting structured data, semi-structured data, and unstructured data from multiple data sources within the EPC enterprise, and performing cleaning and normalization preprocessing to obtain preprocessed data;

[0013] Train machine learning models based on the applicable regulations, applicable tax rates, applicable tax exemptions and reductions, and pre-processed data;

[0014] Performing a multi-dimensional evaluation of the machine learning model;

[0015] When new project data is monitored, the tax cost reference value is obtained through the evaluated machine learning model, and a tax planning plan is generated.

[0016] In one embodiment, obtaining the applicable regulations, applicable tax rates, and applicable exemption conditions of the EPC enterprise based on current tax policies and regulations and the enterprise's situation includes:

[0017] Use the pre-trained ALBERT-xxlarge model to analyze current tax policies and regulations, extracting policy terms, tax rates, geographical conditions, exemption conditions, and applicable entities;

[0018] A relational model based on the attention mechanism identifies relationships between nodes and constructs a Neo4j-based tax policy knowledge graph. This tax policy knowledge graph has a multi-layered architecture: the top layer represents the tax legal system, the middle layer is indexed by the three dimensions of "industry, region, and taxpayer type," and the bottom layer contains metadata related to specific policy clauses, including effective date, document number, and revision history. Node attributes innovatively incorporate dynamically calculated fields, and relationship weights are quantified using the TF-IDF algorithm to quantify the strength of clause associations.

[0019] According to the enterprise situation, the applicable regulations, applicable tax rates and applicable exemption conditions of the EPC enterprise are obtained from the tax policy knowledge graph.

[0020] In one embodiment, the method further comprises:

[0021] Receive new tax policies and regulations in real time through Apache Kafka and update the tax policy knowledge graph.

[0022] In one embodiment, performing cleaning and normalization preprocessing to obtain preprocessed data includes:

[0023] The structured data is automatically processed using the Python pandas library. Specifically, the mean and standard deviation of numerical features are calculated, the range of outliers is determined according to the 3σ principle, and explainable outliers are corrected by mean filling or linear interpolation. Unexplained outliers are directly eliminated.

[0024] For the structured data that needs to retain the original distribution characteristics, MinMaxScaler is used to linearly map it to the interval [0,1]. For the structured data that needs to eliminate the dimension effect, StandardScaler is used to convert it into a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0025] Extract key information from the semi-structured and unstructured data through natural language processing and OCR technology;

[0026] Dynamic matching of tax rules is achieved based on the tax policy knowledge graph and rule engine.

[0027] In one embodiment, training a machine learning model based on the applicable regulations, applicable tax rates, and pre-processed data includes:

[0028] Based on the scikit-learn library: for data with significant linear relationships, use Ridge regression to build a linear regression model; or for data with significant nonlinear characteristics, use the ExtraTreesRegressor class to build a decision tree model; and

[0029] Perform integrated learning, including XGBoost and LightGBM:

[0030] XGBoost method: End-to-end training is performed by defining the tree structure, setting the loss function, and optimizing the model. Precise tree structure parameters and a business-adaptive loss function are defined, along with an adaptive optimization strategy. The feature engineering phase innovatively introduces third-order interaction features and temporal dynamic features, and employs hierarchical feature importance screening. The training process utilizes distributed incremental learning, ensuring model robustness through early stopping and adversarial validation. A model interpreter outputs feature contribution heatmaps in real time.

[0031] LightGBM method: Flexible training is performed by defining model parameters, setting loss functions and optimizers; during model training, a stratified cross-validation method is used to evaluate model performance, and Bayesian optimization is used to search for optimal hyperparameters and business-oriented loss function design, combined with an adaptive optimization strategy; in the feature processing stage, dynamic feature encoding is innovatively implemented, and a temporal attention mechanism is introduced to weightedly process historical tax data; the training process uses GPU-accelerated histogram construction and distributed data parallelism, and incremental model updates are achieved through elastic weight solidification technology.

[0032] In one embodiment, the pre-processed data is stored in the enterprise-level message queue Kafka by using the Hadoop and Spark frameworks to achieve efficient data storage and parallel computing; the new project data includes contract execution data, invoicing and collection data, and cross-region prepayment data, and is packaged in JSON format and sent to the Kafka queue.

[0033] In one embodiment, performing a multi-dimensional evaluation on the machine learning model includes:

[0034] Regression task evaluation, using the XGBoost framework to calculate the mean squared error (MSE) and compare the deviation between the predicted value and the true value to evaluate the prediction accuracy of the machine learning model; or / and

[0035] For classification task evaluation, the AUC-ROC curve is used as the evaluation metric. By balancing the true positive rate and false positive rate, the ROC curve is first constructed based on the risk probability output of the test set samples. The optimal decision threshold is selected by maximizing the Youden index. The model is deployed when the AUC value is > 0.92. SHAP value analysis is then used to identify high-contribution feature combinations and convert them into Drools monitoring rules and derived features. FGSM is used to generate adversarial samples for model reinforcement training. SMOTE oversampling is used for missed samples to balance the data distribution. Finally, the uncertain samples of the model are pushed for manual annotation to form a closed-loop feedback loop; or / and

[0036] Named Entity Recognition evaluation: Building an evaluation system based on the BERT pre-trained model to evaluate entity recognition performance through F1 score and entity boundary accuracy; or / and

[0037] Text classification evaluation involves constructing a multi-dimensional feature fusion engine to structure input text features. These features are first processed through a multi-source data collection layer, including PDF / OCR parsing of unstructured text such as tax documents, local regulations, and corporate contracts. All text is then domain-adaptively segmented to generate character-level features. At the word embedding level, the tax professional version of Word2Vec is combined with dynamic context embeddings generated by RoBERTa-wwm-ext to achieve adaptive fusion of static and dynamic word embeddings. A sentence-level representation incorporates a tax clause structure-aware encoder to identify the three-part logical structure of "policy subject-applicable conditions-implementation standards" in the text, embedding the clause type label as an additional position. Document-level features are cross-modally fused with metadata specific to tax policies. Training is performed in stages based on text complexity using difficult samples generated through adversarial training. The training process begins with training the underlying feature extractor using basic policy text, using a class-weighted cross-entropy loss function. Difficult samples such as complex local regulations and cross-border tax agreements are then introduced, using focal loss to enhance edge case learning. Adversarial training is also initiated to enhance robustness. Mixed-precision training is then used to optimize the model structure.

[0038] Data processing efficiency evaluation, by calculating the throughput of data preprocessing, feature engineering, and model inference to evaluate the processing efficiency of the distributed computing framework; or / and

[0039] The data consistency and integrity assessment establishes a data quality assessment system based on the data governance framework, and ensures the reliability and consistency of data in all links through data traceability and data verification rules.

[0040] In one embodiment, the method further comprises:

[0041] Based on the scoring results of the machine learning model, the customer base is divided into three levels: high value, medium value and low value through the K-Means clustering algorithm. Each level is subdivided in combination with business rules, and differentiated tiered management strategies are implemented for customers with different scores.

[0042] In one embodiment, the method further comprises:

[0043] The tax cost reference value is monitored in real time by a monitoring model to track model performance.

[0044] In one embodiment, the method further comprises:

[0045] The machine learning model is retrained at every preset period or when the tax policy knowledge graph is updated or the model performance indicators do not meet the standards.

[0046] In one embodiment, the method further comprises:

[0047] Generate project economic activity analysis report.

[0048] An AI-based EPC enterprise invoice tax management device, comprising:

[0049] The regulations acquisition module is used to obtain the applicable regulations, applicable tax rates and applicable exemption conditions of EPC companies based on current tax policies and regulations and corporate conditions;

[0050] A data processing module is used to extract structured data, semi-structured data and unstructured data from multiple data sources within the EPC enterprise, and perform cleaning and normalization pre-processing to obtain pre-processed data;

[0051] a model training module for training a machine learning model based on the applicable regulations, applicable tax rates, applicable tax exemptions and reductions, and pre-processed data;

[0052] A model evaluation module, used to perform multi-dimensional evaluation of the machine learning model;

[0053] The reference value acquisition module is used to obtain the tax cost reference value through the evaluated machine learning model when monitoring new project data, and generate a tax planning plan.

[0054] An AI device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0055] Obtain the applicable regulations, applicable tax rates and applicable exemption conditions for EPC companies based on current tax policies and regulations and the company's circumstances;

[0056] Extracting structured data, semi-structured data, and unstructured data from multiple data sources within the EPC enterprise, and performing cleaning and normalization preprocessing to obtain preprocessed data;

[0057] Train machine learning models based on the applicable regulations, applicable tax rates, applicable tax exemptions and reductions, and pre-processed data;

[0058] Performing a multi-dimensional evaluation of the machine learning model;

[0059] When new project data is monitored, the tax cost reference value is obtained through the evaluated machine learning model, and a tax planning plan is generated.

[0060] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the following steps:

[0061] Obtain the applicable regulations, applicable tax rates and applicable exemption conditions for EPC companies based on current tax policies and regulations and the company's circumstances;

[0062] Extracting structured data, semi-structured data, and unstructured data from multiple data sources within the EPC enterprise, and performing cleaning and normalization preprocessing to obtain preprocessed data;

[0063] Train machine learning models based on the applicable regulations, applicable tax rates, applicable tax exemptions and reductions, and pre-processed data;

[0064] Performing a multi-dimensional evaluation of the machine learning model;

[0065] When new project data is monitored, the tax cost reference value is obtained through the evaluated machine learning model, and a tax planning plan is generated.

[0066] The above-mentioned AI-based EPC enterprise invoice tax management method, device, equipment and storage medium obtains the applicable regulations, applicable tax rates and applicable exemption conditions of the EPC enterprise according to the current tax policies and regulations and the enterprise situation; extracts structured data, semi-structured and unstructured data from the multi-source data sources within the EPC enterprise, and performs cleaning and normalization pre-processing to obtain pre-processed data; trains the machine learning model according to the applicable regulations, applicable tax rates, applicable exemption conditions and pre-processed data; performs multi-dimensional evaluation on the machine learning model; when new project data is obtained, the tax payable is obtained through the evaluated machine learning model, and a tax planning scheme is generated. Through the deep integration of machine learning, the intelligent advancement of tax management of the entire life cycle of EPC projects is realized, and key tax management problems such as large deviations in tax cost predictions, difficulties in real-time tax calculation, and incomplete analysis of economic activities are effectively overcome. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 This is a flowchart of an AI-based EPC enterprise invoice tax management method in one embodiment;

[0068] Figure 2 This is a schematic diagram of the AI-based EPC enterprise invoice tax system architecture in one embodiment;

[0069] Figure 3 A schematic diagram of a tax cost prediction process in one embodiment;

[0070] Figure 4 This is a schematic diagram of the refined operation process of AI-based EPC enterprise financial invoice and tax management in one embodiment;

[0071] Figure 5 This is a schematic diagram of the refined operation process of AI-based EPC enterprise financial invoice and tax management in one embodiment;

[0072] Figure 6This is a structural block diagram of an AI-based EPC enterprise financial invoice and tax management device in one embodiment;

[0073] Figure 7 FIG. 4 is a diagram showing the internal structure of an AI device in one embodiment. DETAILED DESCRIPTION

[0074] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0075] The AI-based EPC enterprise invoice tax management method provided in this application can be applied to AI devices. Among them, the AI device can be implemented with an independent server or a server cluster composed of multiple servers. The AI device is used to obtain the applicable regulations, applicable tax rates and applicable exemption conditions of the EPC enterprise based on the current tax policies and regulations and the company's situation; extract structured data, semi-structured and unstructured data from the multi-source data sources within the EPC enterprise, and perform cleaning and normalization pre-processing to obtain pre-processed data; train the machine learning model based on the applicable regulations, applicable tax rates, applicable exemption conditions and pre-processed data; perform multi-dimensional evaluation of the machine learning model; when obtaining new project data, obtain the taxable amount through the evaluated machine learning model, and generate a tax planning plan. Through the deep integration of machine learning, the intelligent advancement of tax management of the entire life cycle of the EPC project is achieved, effectively overcoming key tax management problems such as large deviations in tax cost forecasts, difficulties in real-time tax calculations, and incomplete analysis of economic activities.

[0076] In one embodiment, Figure 1 As shown in the figure, an AI-based EPC enterprise invoice tax management method is provided. The method is applied to AI equipment as an example to illustrate, including the following steps:

[0077] Step S101: Obtain applicable regulations, applicable tax rates, and applicable exemption conditions of the EPC enterprise based on current tax policies and regulations and the enterprise's situation.

[0078] In this example, current tax policies and regulations refer to the country's current and effective tax policies and regulations, including 12 categories of texts such as laws, administrative regulations, and departmental regulations, totaling over 500,000. Enterprise conditions include factors that affect tax application, such as enterprise type, whether it is a high-tech enterprise, and the region where it is located. The following steps are used to obtain the applicable regulations, applicable tax rates, and applicable exemption conditions for EPC enterprises based on current tax policies and regulations and enterprise conditions:

[0079] The pre-trained ALBERT-xxlarge model was used to analyze current tax policies and regulations. Combined with the BiLSTM-CRF model, the model was used to extract policy terms, tax rates, geographical conditions, exemption conditions, and applicable entities. These included policy terms such as "Article 16 of the Provisional Regulations on Value-Added Tax," tax rate elements such as "9%" and "immediate tax refund," applicable entities such as "high-tech enterprises" and "small and micro enterprises," and geographical conditions such as "Western Development Region" and "Free Trade Pilot Zone."

[0080] A relational model based on the attention mechanism (BERT + Pointer Network) identifies relationships between nodes and constructs a tax policy knowledge graph based on Neo4j. This tax policy knowledge graph has a multi-layered architecture, with Neo4j being a high-performance NOSQL graph database that stores structured data on the network rather than in tables. The top layer represents the tax legal system, from the Tax Collection and Administration Law to various tax regulations and implementation details. The middle layer is indexed based on the three dimensions of "industry-region-taxpayer type." The bottom layer links metadata such as the effective date, document number, and revision history of specific policy clauses. Node attributes innovatively incorporate dynamically calculated fields, such as "policy popularity index" = recent query frequency × number of associated enterprises. Relationship weights are quantified using the TF-IDF algorithm to quantify the strength of the clause association. For example, the association weight between "additional deduction for R&D expenses" and "high-tech enterprises" is set to 0.92.

[0081] According to the enterprise situation, the applicable regulations, applicable tax rates and applicable exemption conditions of the EPC enterprise are obtained from the tax policy knowledge graph.

[0082] Preferably, the method further comprises:

[0083] Receive new tax policies and regulations in real time through Apache Kafka and update the tax policy knowledge graph.

[0084] Specifically, the Graph Algorithms library is used to automatically detect policy clusters, with a modularity threshold set to 0.65. Complex logical reasoning is performed using SPARQL, such as "find all preferential policies applicable to cross-border e-commerce with a tax rate below 13%." Cypher is used for path analysis, such as "tracing the path from the current policy to the superior law." A custom tax rule DSL is also implemented.

[0085] Step S102 extracts structured data, semi-structured data, and unstructured data from multiple sources within the EPC enterprise, and performs cleaning and normalization preprocessing to obtain preprocessed data. The multi-source data includes historical project data, historical financial data, historical invoice data, and historical tax costs in the ERP system or financial management system.

[0086] In this embodiment, based on the data characteristics of financial and tax scenarios under the EPC model, it includes three categories: structured, semi-structured and unstructured. Among them, structured data is stored in a standardized two-dimensional table format with the help of a relational database to facilitate efficient management. Structured data includes taxpayer declaration data, tax registration information, and historical project data, historical financial data, historical invoice data and historical tax costs in the ERP system or financial management system; semi-structured and unstructured data are stored in the form of documents such as HTML, JSON, and XML. For structured data, Apache Atlas is used to build a dynamic metadata network, encompassing over 200 business entities and over 5,000 field relationships. Field semantic constraints are automatically inferred, such as requiring "payment amount" to be positive, and the Apriori algorithm is applied to discover field association rules, such as "Contract Type = Government Procurement" → "Tax Rate = 9%." Data conversion templates are generated using a combination of XSLT 3.0 and Schematron validation, and template matching is optimized using a genetic algorithm. The fitness function incorporates field coverage and conversion accuracy. For unstructured data, a multi-model integration architecture is employed, with the TableNet neural network and DocBank annotation system incorporated into document analysis. Entity recognition incorporates domain-adaptive training based on RoBERTa-wwm-ext, using a BILSTM-CRF decoding architecture. The label set includes 27 categories of fiscal and tax entities. Relationship extraction utilizes an improved version of the CasRel model, incorporating a multi-head attention mechanism, resulting in an F1-score of 0.94. Time series feature engineering supports Spark distributed computing, with feature selection sorted by SHAP values and then validated using XGBoost feature importance.

[0087] Specifically, the cleaning and normalization preprocessing to obtain preprocessed data includes:

[0088] The structured data is automatically processed using the Python pandas library. Specifically, the mean and standard deviation of numerical features are calculated, the range of outliers is determined according to the 3σ principle, and explainable outliers are corrected by mean filling or linear interpolation. Unexplained outliers are directly eliminated.

[0089] For structured data that needs to retain its original distribution characteristics, MinMaxScaler is used to linearly map it to the interval [0, 1]. For structured data that needs to eliminate dimensionality, StandardScaler is used to convert it to a standard normal distribution with a mean of 0 and a standard deviation of 1. This series of preprocessing operations ensures the standardization and stability of the data and effectively reduces the interference of abnormal data on model training.

[0090] The key information of the semi-structured and unstructured data is extracted through natural language processing and OCR technology; the key information is relevant information extracted from project contracts, invoices, cross-regional prepayment data, etc., including but not limited to contract amount, executed amount, execution progress; invoice amount, tax rate, invoice date, recipient information; prepayment amount, prepayment location, and project to which it belongs.

[0091] Dynamic matching of tax rules is achieved based on the tax policy knowledge graph and the rule engine, wherein the rule engine is a tax policy rule library built based on the Drools tool, which extracts, verifies, deduplicates and converts the format of unstructured data to realize the automated processing of business rules.

[0092] Step S103: training a machine learning model based on the applicable laws and regulations, applicable tax rates, applicable tax exemption conditions, and preprocessed data.

[0093] In this embodiment, a deep learning model is built based on the TensorFlow or PyTorch framework. By setting the mean square error (MSE) as the loss function, the stochastic gradient descent (SGD) optimizer is used for model training, and the mean_squared_error and mean_absolute_error functions in the Scikit-learn library are used to calculate the root mean square error (RMSE) and mean absolute error (MAE) to ensure the accuracy of tax forecasting. The specific implementation method is to generate a four-dimensional tensor structure through the feature engineering pipeline through the input data: the first dimension is the structure The first dimension is the business features, including multiple numerical fields such as order amount (logarithmically transformed), supplier default history, industry risk factor, etc., which are normalized by Z-Score and then spliced into a [Batch_size×32] matrix; the second dimension is the time series features, which are the rolling window statistics extracted from the transaction flow by TSFRESH, including 58 features such as the moving average of the payment cycle in the past 6 months, Hurst exponent and Fourier transform frequency, forming a [Batch_size×58×6] three-dimensional tensor; the third dimension is the text embedding feature, which is the 1024-dimensional semantic vector generated by the contract terms, which is consistent with the email communication The recorded Sentence-BERT encoding [Batch_size×768] is weightedly fused with attention; the fourth dimension is the graph structure feature, and the supplier relationship graph is embedded in the multi-dimensional node generated by GraphSAGE and aggregated into a [Batch_size×256] matrix according to the weighted transaction relationship; the model architecture adopts a multimodal cross-attention mechanism, in which the TabNet branch processes structured features, the number of decision steps = 6, the feature selection ratio is 0.7, and it supports processing time series data. The graph features are gated and fused with the text features after the GATv2 convolution layer, and the sigmoid threshold is = 0.6; the loss The loss function uses an improved Huber-MSE hybrid loss, switched when δ=1.5, and the weight of high-tax samples is increased by 2.5 times. The optimizer uses Lamb with linear warmup and cosine annealing scheduling. Distributed training is performed for each GPU processing 4096 samples / batch, with a gradient synchronization cycle of 4 steps, and random weights are integrated to improve model robustness. In addition to conventional indicators, the verification phase adds business-customized evaluation: stratified MAE is calculated for different tax types, and feature contribution is analyzed using Shapley values, such as eliminating noise features with a contribution of less than 1%. TensorRT optimization is used for model deployment, achieving a single prediction time of less than 8ms on an NVIDIA T4 GPU, a throughput of 1200 requests / second, and a capture rate of 98.7% for abnormal tax amounts (such as deviations >10%). Through the online learning module, continuous optimization of model indicators is achieved, with an average weekly improvement of 0.4%. According to the applicable regulations, applicable tax rates, and preprocessed data, training of the machine learning model includes:

[0094] Based on the scikit-learn library: for data with significant linear relationships, use Ridge regression to build a linear regression model; or for data with significant nonlinear characteristics, use the ExtraTreesRegressor class to build a decision tree model; and

[0095] Ensemble learning, including XGBoost and LightGBM:

[0096] XGBoost method: End-to-end training is performed by defining a tree structure, setting a loss function, and optimizing the optimizer. By defining precise tree structure parameters and a business-adaptive loss function, an adaptive optimization strategy is employed. The feature engineering phase innovatively introduces third-order interaction features and time-series dynamic features, and employs hierarchical feature importance screening. The training process utilizes distributed incremental learning, with early stopping and adversarial validation ensuring model robustness. The model interpreter outputs real-time feature contribution heat maps, such as "Input tax volatility contribution 35% in the last three months," enabling tax administration personnel to dynamically adjust audit strategies, increasing the efficiency of identifying high-risk enterprises by 3.4 times.

[0097] The LightGBM approach allows for flexible training by defining model parameters, setting loss functions, and optimizing the optimizer. During model training, stratified cross-validation is used to evaluate model performance, with Bayesian optimization used to search for optimal hyperparameters and business-oriented loss function design, complemented by an adaptive optimization strategy. The feature processing phase innovatively implements dynamic feature encoding and introduces a temporal attention mechanism for weighted processing of historical tax data. The training process utilizes GPU-accelerated histogram construction and distributed data parallelism, with elastic weight solidification technology enabling incremental model updates. Embedded feature importance analysis automatically generates economic activity analysis reports, guiding grassroots tax authorities in achieving precise tax collection and management, thereby ensuring optimal predictive performance. This application utilizes a multi-level modeling strategy combining traditional models with ensemble learning.

[0098] Step S104: perform a multi-dimensional evaluation on the machine learning model.

[0099] In this embodiment, the multi-dimensional evaluation of the machine learning model includes:

[0100] Regression task evaluation, using the XGBoost framework to calculate the mean squared error (MSE) and compare the deviation between the predicted value and the true value to evaluate the prediction accuracy of the machine learning model; or / and

[0101] For classification task evaluation, the AUC-ROC curve is used as the evaluation metric. By balancing the true positive rate and false positive rate, the ROC curve is first constructed based on the risk probability output of the test set samples. The optimal decision threshold is selected by maximizing the Youden index. The model is deployed when the AUC value is > 0.92. SHAP value analysis is then used to identify high-contribution feature combinations and convert them into Drools monitoring rules and derived features. FGSM is used to generate adversarial samples for model reinforcement training. SMOTE oversampling is used for missed samples to balance the data distribution. Finally, the uncertain samples of the model are pushed for manual annotation to form a closed-loop feedback loop; or / and

[0102] Named Entity Recognition evaluation: Building an evaluation system based on the BERT pre-trained model to evaluate entity recognition performance through F1 score and entity boundary accuracy; or / and

[0103] Text classification evaluation uses a multi-dimensional feature fusion engine to construct structured input text features. These features are first processed through a multi-source data collection layer, including PDF / OCR parsing of unstructured text such as tax documents, local regulations, and corporate contracts. All text is then domain-adaptively segmented to generate character-level features. At the word embedding level, the tax professional version of Word2Vec is combined with dynamic context embeddings generated by RoBERTa-wwm-ext to achieve adaptive fusion of static and dynamic word embeddings. A tax clause structure-aware encoder is introduced into the sentence-level representation to identify the three-part logical structure of "policy subject-applicable conditions-implementation standards" in the text, embedding the clause type label as an additional position. Document-level features are cross-modally fused with metadata specific to tax policies. Training is performed in stages based on text complexity using difficult samples generated through adversarial training. The training process begins with training the underlying feature extractor using basic policy text, using a class-weighted cross-entropy loss function. Difficult samples such as complex local regulations and cross-border tax agreements are then introduced, using focal loss (γ=2) to strengthen edge case learning. Adversarial training is also initiated to improve robustness. Mixed-precision training is then used to optimize the model structure.

[0104] Data processing efficiency evaluation, by calculating the throughput of data preprocessing, feature engineering, and model inference to evaluate the processing efficiency of the distributed computing framework; or / and

[0105] This data consistency and integrity assessment establishes a data quality assessment system based on the data governance framework. By implementing data traceability and data validation rules, we ensure data reliability and consistency at every stage. This multi-layered assessment approach comprehensively improves model performance, significantly enhancing the intelligence and decision-making efficiency of tax management, and providing enterprises with efficient and accurate tax management support.

[0106] Preferably, after step S104, the method further includes:

[0107] Based on the scoring results of the machine learning model, the customer base is divided into three tiers: high-value, medium-value, and low-value. Each tier is further segmented based on business rules, and differentiated tiered management strategies are implemented for customers with different scores. A proactive service model is adopted for high-value customers, with personalized tax optimization solutions regularly delivered through relevant systems. A guided service model is adopted for medium-value customers, with standardized tax consulting services provided through an intelligent customer service system. A self-service model is adopted for low-value customers, with basic tax query functions provided through a mobile application. Based on the customer stratification results, a reinforcement learning algorithm is used to identify the optimal management strategy. The effectiveness of the strategy is verified through A / B testing, and the strategy combination is continuously optimized. A return-on-input (ROI) evaluation model is also constructed, taking into account customer lifetime value and service costs, to dynamically adjust resource allocation. Finally, a closed-loop feedback mechanism is established to continuously optimize cross-domain application effectiveness through real-time monitoring of key indicators (such as customer satisfaction and service response time). Ultimately, the goal is to maximize resource investment and returns, significantly improving the accuracy and efficiency of tax management.

[0108] Step S105: When new project data is monitored, the tax cost reference value is obtained through the evaluated machine learning model, and a tax planning plan is generated.

[0109] In this embodiment, the preprocessed data is stored in Kafka, an enterprise-level message queue, using the Hadoop and Spark frameworks to achieve efficient data storage and parallel computing. The new project data includes contract execution data, invoicing and collection data, and cross-regional prepayment data, which are packaged in JSON format and sent to the Kafka queue. The new project features are formed into a feature matrix according to the preprocessing rules, and the trained model is called to obtain the prediction results. Multi-dimensional error analysis is performed based on the evaluation function of the scikit-learn library; the median absolute error is calculated and the explained variance score is calculated. The evaluation results are compared with the preset threshold: if the threshold range is exceeded, the model is retrained by adjusting the model parameters or supplementing the training data; if it is within the threshold range, the prediction results are used as the reference value for the tax cost of the project budget.

[0110] Preferably, after step S105, the method further includes:

[0111] The machine learning model is retrained at every preset period or when the tax policy knowledge graph is updated or the model performance indicators do not meet the standards. The model is retrained regularly based on the latest data to ensure the timeliness and accuracy of the prediction results.

[0112] Preferably, after step S105, the method further includes:

[0113] Based on real-time monitoring of the tax cost reference value, the model monitors indicator fluctuations to track model performance. The degradation handling mechanism adopts a graded response strategy: mild degradation (indicator fluctuation <20%) automatically initiates incremental training; moderate degradation (20-50%) triggers Bayesian optimization hyperparameter adjustment (multiple rounds of iterations); and severe degradation (>50%) triggers a rollback to the previous stable version with an email notification. Furthermore, the underlying system collects metrics through Prometheus (with a 15-second sampling frequency). Key model snapshots and evaluation results are persisted to MongoDB, and integrated with the CI / CD pipeline for automatic model retraining. In actual operation, model anomalies are detected an average of 7 days earlier, a 5x improvement compared to traditional methods. The automatic recovery time for degraded models is reduced to 12 minutes, ensuring that fluctuations in prediction accuracy are consistently within a ±3% range. All model changes are recorded on the blockchain, allowing real-time tracking of model performance indicators to promptly identify and address model degradation issues.

[0114] Preferably, after step S105, the method further includes:

[0115] Generate economic activity analysis reports. Specifically, Figure 5As shown in the figure, the data integration layer adopts the strategy of "multi-source aggregation + ETL processing". In the project planning, construction and completion stages, it relies on professional ETL tools to carefully configure the connection parameters for multiple data sources such as financial databases, project management systems, invoice management systems, etc. to ensure stable and reliable data transmission. According to the data structure and business logic of different data sources, the extraction rules are customized. For example, the contract revenue and cost data are accurately extracted from the contract table of the financial database, and key information such as the invoice amount and tax rate are obtained from the invoice management system. In the cleaning and conversion stage, the data cleaning algorithm is used to unify the format of the contract revenue and cost data, such as unifying the amount unit in different contracts to 10,000 yuan; the expense expenditure data is standardized and classified according to the cost category, such as dividing the expenses into labor cost, material cost, equipment cost, etc.Based on the data volume and query requirements, data storage solutions can be flexibly selected. If the data volume is small and the real-time query response requirements are high, a relational database can be selected and its mature transaction processing mechanism can be used to ensure data consistency and integrity. If the data volume is large and complex distributed analysis is required, the Hive distributed column storage database can be used to give full play to its advantages in massive data storage and parallel computing, achieve efficient aggregation and preliminary processing of financial data, break the data silo phenomenon under the traditional model, and lay a solid foundation for subsequent in-depth analysis. In the in-depth analysis layer, the "data mining + intelligent modeling" method is used to implement the Apriori algorithm through the mlxtend library to conduct a deep scan of the integrated financial data, and explore the potential association rules between financial and economic indicators such as contract revenue and cost, expense expenditure and tax, and capital flow and project progress. With the help of the KMeans class of the scikit-learn library, cluster analysis of project cost data is performed, projects with similar cost structures are classified into one category, and project groups with different cost characteristics are identified to help enterprises accurately locate the focus of cost control. The library builds a VAR (vector autoregression) model, comprehensively considering the mutual influence of multiple financial variables, inputting time series data such as historical project revenue, cost, and cash flow, predicting the future economic trend of the project, and providing forward-looking information for corporate strategic decision-making. Compared with traditional analysis methods, it has achieved a leap from single indicator analysis to multi-dimensional and in-depth data analysis, and deeply excavates the laws and trends behind financial data; the report generation layer builds a "template-driven + visual presentation" mechanism, using Python's Jinja2 template engine to design professional project economic activity analysis report templates, cleverly embedding Python code blocks and placeholders in the templates to dynamically fill in project economic activity analysis data and conclusions, reading key financial data such as total project revenue, total cost, profit, etc. from the data warehouse or analysis result storage medium, and passing it to the Jinja2 template engine for rendering. At the same time, using Matplotlib and Seaborn The library's powerful data visualization capabilities generate a rich variety of visual charts. For example, bar charts intuitively display the cost structure of different project stages, clearly showing the contribution of each cost item; line charts accurately show the changing trend of project revenue over time, helping companies understand the reasons for revenue growth or fluctuations; and scatter plots provide in-depth analysis of the relationship between costs and benefits, assisting companies in evaluating project profitability. These visual charts are organically integrated with textual analysis content and embedded into report documents to generate project economic activity analysis reports in formats such as PDF and Word, tailored to company needs. This intuitive and easy-to-understand approach provides comprehensive, in-depth, and decision-making data support for company management to review project financial work and plan subsequent project finances, greatly improving the readability and practicality of reports.

[0116] In the above-mentioned AI-based EPC enterprise invoice tax management method, the applicable regulations, applicable tax rates and applicable exemption conditions of the EPC enterprise are obtained according to the current tax policies and regulations and the enterprise's situation; structured data, semi-structured and unstructured data are extracted from the multi-source data sources within the EPC enterprise, and cleaned and normalized pre-processed to obtain pre-processed data, and the multi-source data includes historical project data, historical financial data, historical invoice data and historical tax costs in the ERP system or financial management system; a machine learning model is trained according to the applicable regulations, applicable tax rates, applicable exemption conditions and pre-processed data; the machine learning model is evaluated in multiple dimensions; when new project data is obtained, the tax payable is obtained through the evaluated machine learning model, and a tax planning scheme is generated. Through the deep integration of machine learning, the intelligent advancement of tax management of the entire life cycle of EPC projects is achieved, effectively overcoming key tax management problems such as large deviations in tax cost predictions, difficulties in real-time tax calculation, and incomplete analysis of economic activities.

[0117] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. In addition, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0118] In one embodiment, Figure 6 As shown, an AI-based EPC enterprise invoice tax management device is provided, comprising: a regulation acquisition module 61, a data processing module 62, a model training module 63, a model evaluation module 64 and a reference value acquisition module 65, wherein:

[0119] The regulation acquisition module 61 is used to obtain the applicable regulations, applicable tax rates and applicable exemption conditions of the EPC enterprise based on the current tax policies and regulations and the enterprise's situation;

[0120] The data processing module 62 is used to extract structured data, semi-structured data and unstructured data from multiple data sources within the EPC enterprise, and perform cleaning and normalization pre-processing to obtain pre-processed data;

[0121] a model training module 63 for training a machine learning model based on the applicable regulations, applicable tax rates, applicable tax exemptions and reductions, and preprocessed data;

[0122] A model evaluation module 64 is used to perform a multi-dimensional evaluation of the machine learning model;

[0123] The reference value acquisition module 65 is used to obtain the tax cost reference value through the evaluated machine learning model when monitoring new project data, and generate a tax planning plan.

[0124] Furthermore, the regulation acquisition module 61 is specifically used to parse current tax policies and regulations through the pre-trained ALBERT-xxlarge model, extracting policy terms, tax rates, geographical conditions, exemption conditions, and applicable entities;

[0125] A relational model based on the attention mechanism identifies relationships between nodes and constructs a Neo4j-based tax policy knowledge graph. This tax policy knowledge graph has a multi-layered architecture: the top layer represents the tax legal system, the middle layer is indexed by the three dimensions of "industry, region, and taxpayer type," and the bottom layer contains metadata related to specific policy clauses, including effective date, document number, and revision history. Node attributes innovatively incorporate dynamically calculated fields, and relationship weights are quantified using the TF-IDF algorithm to quantify the strength of clause associations.

[0126] According to the enterprise situation, the applicable regulations, applicable tax rates and applicable exemption conditions of the EPC enterprise are obtained from the tax policy knowledge graph.

[0127] Furthermore, the device further comprises:

[0128] The regulations update module is used to receive new tax policies and regulations in real time through Apache Kafka and update the tax policy knowledge graph.

[0129] Furthermore, the data processing module 62 is specifically used to automatically process the structured data based on the Python pandas library. Specifically, it calculates the mean and standard deviation of numerical features, determines the range of outliers according to the 3σ principle, corrects explainable outliers by using mean filling or linear interpolation, and directly eliminates unexplained outliers.

[0130] For the structured data that needs to retain the original distribution characteristics, MinMaxScaler is used to linearly map it to the interval [0,1]. For the structured data that needs to eliminate the dimension effect, StandardScaler is used to convert it into a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0131] Extract key information from the semi-structured and unstructured data through natural language processing and OCR technology;

[0132] Dynamic matching of tax rules is achieved based on the tax policy knowledge graph and rule engine.

[0133] Furthermore, the model training module 63 is specifically used to build a linear regression model using Ridge regression for data with significant linear relationships based on the scikit-learn library; or to build a decision tree model using the ExtraTreesRegressor class for data with significant nonlinear characteristics; and

[0134] Perform integrated learning, including XGBoost and LightGBM:

[0135] XGBoost method: End-to-end training is performed by defining the tree structure, setting the loss function, and optimizing the model. Precise tree structure parameters and a business-adaptive loss function are defined, along with an adaptive optimization strategy. The feature engineering phase innovatively introduces third-order interaction features and temporal dynamic features, and employs hierarchical feature importance screening. The training process utilizes distributed incremental learning, ensuring model robustness through early stopping and adversarial validation. A model interpreter outputs feature contribution heatmaps in real time.

[0136] LightGBM method: Flexible training is performed by defining model parameters, setting loss functions and optimizers; during model training, a stratified cross-validation method is used to evaluate model performance, and Bayesian optimization is used to search for optimal hyperparameters and business-oriented loss function design, combined with an adaptive optimization strategy; in the feature processing stage, dynamic feature encoding is innovatively implemented, and a temporal attention mechanism is introduced to weightedly process historical tax data; the training process uses GPU-accelerated histogram construction and distributed data parallelism, and incremental model updates are achieved through elastic weight solidification technology.

[0137] Furthermore, the pre-processed data is stored in the enterprise-level message queue Kafka by adopting Hadoop and Spark framework to achieve efficient data storage and parallel computing; the new project data includes contract execution data, invoicing and collection data, and cross-region prepayment data, and is packaged in JSON format and sent to the Kafka queue.

[0138] Furthermore, the model evaluation module 64 is specifically used for regression task evaluation, and calculates the mean square error by using the reg:squarederror loss function based on the XGBoost framework, and evaluates the prediction accuracy of the machine learning model by comparing the deviation between the predicted value and the true value; or / and

[0139] For classification task evaluation, the AUC-ROC curve is used as the evaluation metric. By balancing the true positive rate and false positive rate, the ROC curve is first constructed based on the risk probability output of the test set samples. The optimal decision threshold is selected by maximizing the Youden index. The model is deployed when the AUC value is > 0.92. SHAP value analysis is then used to identify high-contribution feature combinations and convert them into Drools monitoring rules and derived features. FGSM is used to generate adversarial samples for model reinforcement training. SMOTE oversampling is used for missed samples to balance the data distribution. Finally, the uncertain samples of the model are pushed for manual annotation to form a closed-loop feedback loop; or / and

[0140] Named Entity Recognition evaluation: Building an evaluation system based on the BERT pre-trained model to evaluate entity recognition performance through F1 score and entity boundary accuracy; or / and

[0141] Text classification evaluation involves constructing a multi-dimensional feature fusion engine to structure input text features. These features are first processed through a multi-source data collection layer, including PDF / OCR parsing of unstructured text such as tax documents, local regulations, and corporate contracts. All text is then domain-adaptively segmented to generate character-level features. At the word embedding level, the tax professional version of Word2Vec is combined with dynamic context embeddings generated by RoBERTa-wwm-ext to achieve adaptive fusion of static and dynamic word embeddings. A sentence-level representation incorporates a tax clause structure-aware encoder to identify the three-part logical structure of "policy subject-applicable conditions-implementation standards" in the text, embedding the clause type label as an additional position. Document-level features are cross-modally fused with metadata specific to tax policies. Training is performed in stages based on text complexity using difficult samples generated through adversarial training. The training process begins with training the underlying feature extractor using basic policy text, using a class-weighted cross-entropy loss function. Difficult samples such as complex local regulations and cross-border tax agreements are then introduced, using focal loss to enhance edge case learning. Adversarial training is also initiated to enhance robustness. Mixed-precision training is then used to optimize the model structure.

[0142] Data processing efficiency evaluation, by calculating the throughput of data preprocessing, feature engineering, and model inference to evaluate the processing efficiency of the distributed computing framework; or / and

[0143] The data consistency and integrity assessment establishes a data quality assessment system based on the data governance framework, and ensures the reliability and consistency of data in all links through data traceability and data verification rules.

[0144] Furthermore, the device further comprises:

[0145] The customer management module is used to divide the customer base into three levels: high value, medium value and low value according to the scoring results of the machine learning model through the K-Means clustering algorithm, and to subdivide each level in combination with business rules to implement differentiated hierarchical management strategies for customers with different scores.

[0146] Furthermore, the device further comprises:

[0147] The model tracking module is used to monitor the tax cost reference value in real time through a monitoring model to track model performance.

[0148] Furthermore, the device further comprises:

[0149] The model update module is used to retrain the machine learning model at preset intervals or when the tax policy knowledge graph is updated or the model performance indicators do not meet the standards.

[0150] Furthermore, the device further comprises:

[0151] Report generation module, used to generate project economic activity analysis report.

[0152] For the specific definition of the AI-based EPC enterprise invoice tax management device, please refer to the definition of the AI-based EPC enterprise invoice tax management method above, which will not be repeated here. The various modules in the above-mentioned AI-based EPC enterprise invoice tax management device can be implemented in whole or in part through software, hardware, and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0153] In one embodiment, an AI device is provided. The AI device may be a server, and its internal structure diagram may be as follows: Figure 7 As shown. The AI device includes a processor, memory, network interface and database connected via a system bus. The processor of the AI device is used to provide computing and control capabilities. The memory of the AI device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the AI device is used to store EPC enterprise invoice tax management data. The network interface of the AI device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, an AI-based EPC enterprise invoice tax management method is implemented.

[0154] Those skilled in the art will understand that Figure 7The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0155] In one embodiment, an AI device is provided, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0156] Obtain the applicable regulations, applicable tax rates and applicable exemption conditions for EPC companies based on current tax policies and regulations and the company's circumstances;

[0157] Extracting structured data, semi-structured data, and unstructured data from multiple data sources within the EPC enterprise, and performing cleaning and normalization preprocessing to obtain preprocessed data;

[0158] Train machine learning models based on the applicable regulations, applicable tax rates, applicable tax exemptions and reductions, and pre-processed data;

[0159] Performing a multi-dimensional evaluation of the machine learning model;

[0160] When new project data is monitored, the tax cost reference value is obtained through the evaluated machine learning model, and a tax planning plan is generated.

[0161] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0162] Obtain the applicable regulations, applicable tax rates and applicable exemption conditions for EPC companies based on current tax policies and regulations and the company's circumstances;

[0163] Extracting structured data, semi-structured data, and unstructured data from multiple data sources within the EPC enterprise, and performing cleaning and normalization preprocessing to obtain preprocessed data;

[0164] Train machine learning models based on the applicable regulations, applicable tax rates, applicable tax exemptions and reductions, and pre-processed data;

[0165] Performing a multi-dimensional evaluation of the machine learning model;

[0166] When new project data is monitored, the tax cost reference value is obtained through the evaluated machine learning model, and a tax planning plan is generated.

[0167] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0168] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0169] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. An AI-based EPC enterprise invoice tax management method, characterized by: The method comprises: Obtain the applicable regulations, applicable tax rates and applicable exemption conditions for EPC companies based on current tax policies and regulations and the company's circumstances; Extracting structured data, semi-structured data, and unstructured data from multiple data sources within the EPC enterprise, and performing cleaning and normalization preprocessing to obtain preprocessed data; Train machine learning models based on the applicable regulations, applicable tax rates, applicable tax exemptions and reductions, and pre-processed data; Performing a multi-dimensional evaluation of the machine learning model; When new project data is monitored, the tax cost reference value is obtained through the evaluated machine learning model, and a tax planning plan is generated.

2. The method according to claim 1, characterized in that According to the current tax policies and regulations and the company's situation, the applicable regulations, applicable tax rates and applicable exemption conditions for EPC companies include: Use the pre-trained ALBERT-xxlarge model to analyze current tax policies and regulations, extracting policy terms, tax rates, geographical conditions, exemption conditions, and applicable entities; A relational model based on the attention mechanism identifies relationships between nodes and constructs a Neo4j-based tax policy knowledge graph. This tax policy knowledge graph has a multi-layered architecture: the top layer represents the tax legal system, the middle layer is indexed by the three dimensions of "industry, region, and taxpayer type," and the bottom layer contains metadata related to specific policy clauses, including effective date, document number, and revision history. Node attributes innovatively incorporate dynamically calculated fields, and relationship weights are quantified using the TF-IDF algorithm to quantify the strength of clause associations. According to the enterprise situation, the applicable regulations, applicable tax rates and applicable exemption conditions of the EPC enterprise are obtained from the tax policy knowledge graph.

3. The method according to claim 1, characterized in that The cleaning and normalization preprocessing to obtain preprocessed data includes: The structured data is automatically processed using the Python pandas library. Specifically, the mean and standard deviation of numerical features are calculated, the range of outliers is determined according to the 3σ principle, and explainable outliers are corrected by mean filling or linear interpolation. Unexplained outliers are directly eliminated. For the structured data that needs to retain the original distribution characteristics, MinMaxScaler is used to linearly map it to the interval [0,1]. For the structured data that needs to eliminate the dimension effect, StandardScaler is used to convert it into a standard normal distribution with a mean of 0 and a standard deviation of 1. Extract key information from the semi-structured and unstructured data through natural language processing and OCR technology; Dynamic matching of tax rules is achieved based on the tax policy knowledge graph and rule engine.

4. The method according to claim 1, wherein The training of the machine learning model according to the applicable regulations, applicable tax rates and pre-processed data includes: Based on the scikit-learn library: for data with significant linear relationships, use Ridge regression to build a linear regression model; or for data with significant nonlinear characteristics, use the ExtraTreesRegressor class to build a decision tree model; and Perform integrated learning, including XGBoost and LightGBM: XGBoost method: End-to-end training is performed by defining the tree structure, setting the loss function, and optimizing the model. Precise tree structure parameters and a business-adaptive loss function are defined, along with an adaptive optimization strategy. The feature engineering phase innovatively introduces third-order interaction features and temporal dynamic features, and employs hierarchical feature importance screening. The training process utilizes distributed incremental learning, ensuring model robustness through early stopping and adversarial validation. A model interpreter outputs feature contribution heatmaps in real time. LightGBM method: Flexible training is performed by defining model parameters, setting loss functions and optimizers; during model training, a stratified cross-validation method is used to evaluate model performance, and Bayesian optimization is used to search for optimal hyperparameters and business-oriented loss function design, combined with an adaptive optimization strategy; in the feature processing stage, dynamic feature encoding is innovatively implemented, and a temporal attention mechanism is introduced to weightedly process historical tax data; the training process uses GPU-accelerated histogram construction and distributed data parallelism, and incremental model updates are achieved through elastic weight solidification technology.

5. The method according to claim 1, wherein The pre-processed data is stored in Kafka, an enterprise-level message queue, using Hadoop and Spark frameworks to achieve efficient data storage and parallel computing. The new project data includes contract execution data, invoicing and collection data, and cross-regional prepayment data, which are packaged in JSON format and sent to the Kafka queue.

6. The method according to claim 1, characterized in that The multi-dimensional evaluation of the machine learning model includes: Regression task evaluation, using the XGBoost framework to calculate the mean squared error (MSE) and compare the deviation between the predicted value and the true value to evaluate the prediction accuracy of the machine learning model; or / and For classification task evaluation, the AUC-ROC curve is used as the evaluation metric. By balancing the true positive rate and false positive rate, the ROC curve is first constructed based on the risk probability output of the test set samples. The optimal decision threshold is selected by maximizing the Youden index. The model is deployed when the AUC value is > 0.

92. SHAP value analysis is then used to identify high-contribution feature combinations and convert them into Drools monitoring rules and derived features. FGSM is used to generate adversarial samples for model reinforcement training. SMOTE oversampling is used for missed samples to balance the data distribution. Finally, the uncertain samples of the model are pushed for manual annotation to form a closed-loop feedback loop; or / and Named Entity Recognition evaluation: Building an evaluation system based on the BERT pre-trained model to evaluate entity recognition performance through F1 score and entity boundary accuracy; or / and Text classification evaluation involves constructing a multi-dimensional feature fusion engine to structure input text features. These features are first processed through a multi-source data collection layer, including PDF / OCR parsing of unstructured text such as tax documents, local regulations, and corporate contracts. All text is then domain-adaptively segmented to generate character-level features. At the word embedding level, the tax professional version of Word2Vec is combined with dynamic context embeddings generated by RoBERTa-wwm-ext to achieve adaptive fusion of static and dynamic word embeddings. A sentence-level representation incorporates a tax clause structure-aware encoder to identify the three-part logical structure of "policy subject-applicable conditions-implementation standards" in the text, embedding the clause type label as an additional position. Document-level features are cross-modally fused with metadata specific to tax policies. Training is performed in stages based on text complexity using difficult samples generated through adversarial training. The training process begins with training the underlying feature extractor using basic policy text, using a class-weighted cross-entropy loss function. Difficult samples such as complex local regulations and cross-border tax agreements are then introduced, using focal loss to enhance edge case learning. Adversarial training is also initiated to enhance robustness. Mixed-precision training is then used to optimize the model structure. Data processing efficiency evaluation, by calculating the throughput of data preprocessing, feature engineering, and model inference to evaluate the processing efficiency of the distributed computing framework; or / and The data consistency and integrity assessment establishes a data quality assessment system based on the data governance framework, and ensures the reliability and consistency of data in all links through data traceability and data verification rules.

7. The method according to claim 1, characterized in that The method further comprises: Generate project economic activity analysis report.

8. An AI-based EPC enterprise invoice tax management device, characterized by: The device comprises: The regulations acquisition module is used to obtain the applicable regulations, applicable tax rates and applicable exemption conditions of EPC companies based on current tax policies and regulations and corporate conditions; A data processing module is used to extract structured data, semi-structured data and unstructured data from multiple data sources within the EPC enterprise, and perform cleaning and normalization pre-processing to obtain pre-processed data; a model training module for training a machine learning model based on the applicable regulations, applicable tax rates, applicable tax exemptions and reductions, and pre-processed data; A model evaluation module, used to perform multi-dimensional evaluation of the machine learning model; The reference value acquisition module is used to obtain the tax cost reference value through the evaluated machine learning model when monitoring new project data, and generate a tax planning plan.

9. An AI device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images