A government investment project fund puzzle intelligent generation system and method

Abstract

The application discloses a government investment project fund puzzle intelligent generation system and method, and relates to the technical field of data processing.The system takes a fund channel rule graph database and a multi-constraint rule engine as cores, and comprises project parameter analysis, multi-dimensional criterion extraction, support ratio and leverage constraint calculation, compliance constraint conversion, fund puzzle closed constraint solving, parameter library version management, batch processing scheduling and combined scheme generation modules.The method extracts and matches a candidate fund channel from structured project parameters through multi-dimensional criterion extraction and rules, integrates compliance constraints and total amount closure, channel exclusion, negative list, income feasibility and capital upper limit constraints into a same constraint satisfaction problem for joint solving, iteratively adjusts in a directional manner according to conflict reasons when there is no feasible solution, guarantees batch calculation caliber consistency through parameter library version locking, outputs a fund puzzle scheme which is closed, compliant and has rule traceability information, and realizes computable generation of ten-thousand-level project batch fund puzzle.

Description

Technical Field

[0001] This invention relates to the fields of data processing and intelligent decision-making technology, specifically to the fields of matching funding channels and compiling funding portfolios for government investment projects, and particularly to an intelligent generation system and method for government investment project funding portfolios. Background Technology

[0002] Funding for government investment projects involves multiple channels, including central government budgetary investment, ultra-long-term special treasury bonds, local government special bonds (hereinafter referred to as special bonds), policy-based financial instruments, and bank loans. Each funding channel is subject to national and provincial policy guidelines regarding its scope of support, support ratio, exclusivity, negative lists, capital requirements, and return requirements. These policy guidelines are updated annually, and multiple guidelines coexist. Funding consolidation refers to a funding arrangement plan that combines multiple funding channels for a specific project, ensuring that the sum of the funds from each channel precisely covers the total investment of the project and meets all compliance constraints. Its preparation is a prerequisite for government investment project application and investment decision-making.

[0003] The technical solutions in the prior art that are closest to this invention include the following categories: General-purpose enterprise financing solution generation / matching systems include: financial financing planning systems based on large language models, automatic inquiry and comparison financing solution generation systems, financing matching methods (CN111861757A), matching of enterprise financing needs with financial services (CN119539961B), and matching systems for science and technology innovation projects with venture capital funds. These systems target enterprise financing scenarios and typically perform semantic or rule-based matching based on the characteristics of the funding demander (industry, size, term, collateral) and the product characteristics of financial institutions (interest rate, term, amount), outputting financing portfolio suggestions. Some systems also incorporate large language models to generate financing solution text.

[0004] Intelligent response and intelligent agent systems based on large-scale models in the government sector: such as intelligent question-and-answer systems for government affairs based on self-developed large-scale government models with hundreds of billions of parameters, intelligent agents for policy consultation, and tools for assisting in the writing of application materials. These systems integrate policy documents, service guides, and other corpora into large language models to provide text generation services such as policy consultation, application Q&A, and material writing assistance. They solve the "consultation and response" problem, but lack the ability to perform structured matching, constraint calculation, and closed-loop verification of funding channels, and fail to solve the computable generation problem of compliant funding from multiple channels.

[0005] Intelligent judgment rule engine and artificial intelligence fusion systems: such as the category adaptive judgment system (CN121902957A) which integrates large language models and deterministic rule engines, and the government policy recommendation and application diagnosis platform based on dynamic profiles (CN121352630A). These systems use a combination of technologies—feature extraction from large language models + deterministic rule engine matching + priority arbitration + versioned hot updates + basis chain tracing—to solve certain types of government judgment / recommendation problems. The feature matching and arbitration mechanisms are similar to some aspects of this invention, but their output is a single-category judgment or policy recommendation text, not involving multi-channel fund allocation, nor possessing the closed-loop calculation verification capability for fund aggregation. CN121902957A uses a JSON rule base + inverted index to store matching rules, which differs from the fund channel rule graph database of this invention at the data structure level (graph structure storage + hit path records, with the hit path directly serving subsequent conflict attribution and tracing).

[0006] Responsible AI deployment systems with decision-gating: such as responsible, risk-controlled, and hazard-prevention AI deployment systems in financial decision-making systems (US20260154749A1). These systems perform compliance verification, risk assessment, and authorization / restriction / delay / blocking gating on individual AI decision-making actions before execution, generating immutable audit logs. Their compliance verification and decision-gating mechanisms are similar to some aspects of this invention, but they only handle the execution and release of individual decision-making actions. They do not generate multi-channel funding portfolios, do not allocate funding amounts or perform closure-based solutions, and do not handle the determination of whether the portfolio is closure-based and feasible. More importantly, their compliance verification occurs in an independent gating stage after the solution is generated; compliance conditions do not participate in the construction and solution of the funding plan.

[0007] Constraint-satisfaction / multi-objective optimization systems oriented towards general resource allocation models include: Intelligent road maintenance closed-loop decision-making system based on budget constraints and funding matching (CN121599380A, mixed-integer programming solution with budget ceiling + funding matching engine); multi-objective optimization method for joint scheduling of water conservancy projects (CN121599390A, structured constraint set + hierarchical solution of feasible region + parameter correction and recalculation); and ESG integrated performance management system with multi-dimensional business value quantification (US20260170441A1, multi-objective optimization and constraint satisfaction algorithm for cross-theme resource allocation). These systems possess general constraint-satisfaction / multi-objective optimization capabilities, but their constraints are general constraints such as budget ceilings, water conservancy feasible regions, and ESG resource allocation. They do not address the closed semantics of the five types of funding constraints in government investment projects ("total amount closure + channel exclusivity + negative list + capital ceiling + revenue feasibility"), nor do they treat compliance requirements as part of the constraint-satisfaction process.

[0008] Systems for fund routing and portfolio optimization include: fund routing methods (CN114240619B), fund routing methods (CN112053143B), power grid portfolio optimization (CN114708105B), intelligent fund routing matching for multiple channels and multiple funders (CN110348985A), and candidate recommendation schemes for multiple channel financing amount ranges (CN116151985A). These systems have proven that "multi-channel fund routing priority matching" and "portfolio constraint optimization" are existing technical approaches. However, their implementation involves single-layer matching of fund demand characteristics and fund channel characteristics, plus priority ranking / scoring and filtering, or general portfolio constraint optimization. Neither of these methods includes the closed-loop semantics of the five types of constraints mentioned above for government investment projects, nor does it incorporate compliance constraints into the solution. Furthermore, it lacks versioned management of policy parameter libraries and consistency guarantees for tens of thousands of batches of policy statements. Its output is an optimal channel or portfolio recommendation, rather than a closed-loop fund allocation scheme and feasibility proof that satisfies all government constraints.

[0009] System optimization based on project portfolio and funding mapping constraints: For example, IBM's Stochastic Investment Planning System (US20140330748A1) discloses a complete process of imposing constraints on multiple projects, combining projects according to compatibility, and iteratively solving using an optimization model until the parameters fall within an acceptable range. Its constraints include total amount constraints, categorized funding constraints, and funding mapping constraints. This system is designed for general scenarios of enterprise or project portfolio investment planning. Its constraints are based on general resource allocation semantics and do not include government funding exclusivity (e.g., ultra-long-term special treasury bonds and central budget investments should not overlap in principle), negative lists, revenue coverage review lines, positive capital lists, or other government-mandated constraints. Its iterative adjustments are numerical optimization searches, rather than a cyclical process of "attributing to conflict causes → targeted adjustments → re-solving" where constraints are met. Furthermore, there are no compliance constraints incorporated into the solution, and no policy version locks or batch consistency guarantees.

[0010] The practice of fund aggregation in the field of government investment: Currently, provincial and municipal governments generally rely on manual processes in compiling fund aggregation data for government investment projects. This involves manually compiling policy guidelines, manually matching channels, and manually verifying closure. Business practices (such as guidance documents for government investment projects at all levels, and training materials on project planning and investment-loan linkage) have resulted in a systematic set of channel rules. These rules include the correspondence between national economic sectors and channels such as central government budgetary investment, ultra-long-term special treasury bonds, and special bonds; differentiated support ratios for eastern, central, and western regions; negative lists; positive lists for capital; leverage constraints in investment-loan linkage; exclusive / misaligned rules between channels; and revenue coverage review lines. However, these rules are scattered across policy texts and the experience of business personnel, and there is currently no structured rule engine or calculable verification method specifically for this field.

[0011] General wealth management and credit decision-making rule engine systems, such as Monetaire Wealth Management Platform (US2004054610A1), Internet Lending System and Methodology (CN108389122A), and Counterparty Credit Risk System (US2004024692A1), etc. These systems use rule engines or decision engines to achieve fund matching and risk control for personal / corporate wealth planning or consumer credit. Their constraints are general financial or credit risk constraints, or independent gating by a risk control module after the solution is generated. They do not involve government investment project mappings such as "industry category ↔ funding channel," differentiated support ratios in the eastern, central, and western regions, negative lists, positive capital lists, or revenue coverage review lines. They do not solve problems related to the closure of the five types of constraints: total amount closure + channel exclusivity + negative list + revenue feasibility + capital ceiling. They also lack compliance constraints incorporated into the feasible domain construction and policy version lock batch consistency mechanisms. The output is a channel recommendation or risk control judgment, rather than a closed-loop funding scheme and feasibility proof that satisfies all government constraints.

[0012] In summary, the existing technology has the following drawbacks: Lack of understanding of the rules and regulations in the field of government investment: The aforementioned general financing system does not have decision-making rules for government investment project funding channels (correspondence between national economic industry categories and various funding channels, differentiated support ratios in the eastern, central and western regions, negative lists, positive lists of capital, investment and loan linkage leverage constraints, etc.), and the generated plans cannot guarantee compliance with national and provincial application standards, often resulting in funding plans that are "unsuitable for application".

[0013] Lack of exclusive / complementary constraints among funding channels: The failure to consider the misalignment rules such as the principle that ultra-long-term special treasury bonds and central budgetary investments should not overlap, loan replacement of special bonds, and policy-based financial instruments serving as "bridges" for special bonds may easily lead to funding gaps or duplicate applications.

[0014] Lack of computable verification of the closure of the funding pool: Neither the rule engine fusion system nor the resource allocation solution system mentioned above can model the five types of government funding constraints—"total amount closure, channel exclusivity, negative list, revenue feasibility, and capital ceiling"—as a solvable constraint satisfaction problem and provide a feasible / infeasible judgment. The feasibility of the solution relies on manual verification and cannot provide a verifiable feasibility guarantee.

[0015] The disconnect between compliance verification and scheme generation: Decision-gated systems treat compliance verification as a pre-execution release step (authorization / restriction / delay / blocking) independent of the scheme generation process. Compliance conditions are not involved in the construction and solution of funding schemes. In the scenario of government funding pooling, compliance conditions (such as the return coverage ratio reaching the review threshold or not meeting the bond issuance conditions, being excluded if it hits the negative list, and strictly prohibiting duplicate applications between channels) are the constructive conditions that determine whether a scheme can be established. If a "generate scheme first, then gate and intercept" approach is adopted, a large number of generated schemes will be intercepted at the gate control stage and need to be redone in the entire round (e.g., special bond schemes with insufficient coverage ratio to the review threshold, combinations that hit the negative list). It is impossible to internalize compliance conditions into constraints at the scheme generation stage and avoid generating non-compliant schemes from the source.

[0016] Lack of large-scale batch processing and consistency of standards: Policy standards are updated annually and multiple standards coexist—the scope of special-purpose bonds was adjusted from 9 categories in 2023 to 11 categories in 2024, and then changed to negative list management; the scope of special-purpose bonds used as capital expanded from 13 categories to 15 / 17 / 22 categories; the upper limit of capital ratio was adjusted from 25% to 30%; the return coverage ratio has multiple standards, with a national minimum of 1.2, a review threshold of 1.3, and 1.5 in some regions. The same project receives different amounts of support under different versions of standards. The existing system has not established a parameter library version management and batch parallel processing mechanism. When dealing with provincial-level projects of tens of thousands, it cannot guarantee the consistency of standards and processing efficiency of batch calculation results; while manually compiling a single project takes several days, which is difficult to fit into the short application window. Summary of the Invention

[0017] To address the shortcomings of existing technologies, the technical problem this invention aims to solve is: under the conditions of provincial-level project data scale of tens of thousands, annual policy iterations, and the coexistence of multiple policy definitions, how to achieve efficient calculation, consistency of definitions, and verifiable results for matching funding channels and verifying the closure of funding combinations for government investment projects—that is, to automatically map structured project parameters into multi-channel funding combination schemes that conform to current policy definitions through a rule engine, and to ensure the funding closure, compliance, and profitability of the schemes through overall constraint solving of five types of constraints: "total amount closure + channel exclusivity + negative list + revenue feasibility + capital ceiling" (compliance requirements are used as constraints in the construction of the feasible domain, rather than being independent of the solution gate). Versioned parameter library management ensures the consistency of batch calculation results under annual policy iterations, and hit path backtracking ensures that each calculation result is verifiable and traceable, achieving fully automatic and computable generation of "project parameter input—compliant and closed-loop funding combination scheme output".

[0018] To address the aforementioned technical problems, in a first aspect, embodiments of the present invention provide an intelligent generation system for government investment project funding pooling, comprising: The project parameter parsing module is used to obtain the structured project parameters of the target project and generate a project parameter vector, which includes at least industry category and sub-category fields; The multidimensional criterion extraction module is used to extract the criterion values ​​of each criterion dimension from the project parameter vector according to multiple preset criterion dimensions, and the value spaces of each criterion dimension are independent of each other. The funding channel rule graph database stores funding channel matching rules in a graph data structure. The nodes of the graph data structure include channel identifier, support scope, support ratio identifier, exclusivity identifier, and negative list identifier. Edges represent judgment conditions and pointing relationships. The matching process records the matching path, which is used for conflict cause attribution and result tracing. The multi-constraint rule engine is used to match the criteria values ​​of each criterion dimension with the funding channel rule graph database to obtain a set of candidate channels. When the hit results of different criteria dimensions conflict, the engine arbitrates according to the preset priority rules and outputs the hit path of each channel. The module supports the calculation of support ratios and leverage constraints, which reads the support ratios and leverage constraints of each candidate channel from a preset parameter library and calculates the support amount for each channel. The compliance constraint conversion module is used to convert negative list constraints, channel exclusivity constraints, and compliance requirements into a set of compliance constraint conditions. The fund pooling closure constraint solving module is used to model and solve a constraint satisfaction problem by taking the fund amount of each channel as variables and the fund pooling verification conditions as constraints. The constraints include total amount closure constraint, channel exclusivity constraint, negative list constraint, compliance constraint, revenue feasibility constraint, and capital ceiling constraint. The total amount closure constraint is the constraint that the sum of the fund amounts of each channel equals the total investment of the project. The channel exclusivity constraint is the constraint determined by the exclusivity identifier in the fund channel rule graph database. The compliance constraint condition set is incorporated into the constraint satisfaction problem by the compliance constraint transformation module and participates in the construction of the feasible region in the form of constraint equations along with other constraints, rather than being intercepted as an independent gate after the solution is output. When a feasible solution exists, a closed fund pooling scheme is output. When no feasible solution exists, the conflict reason is output. The candidate channels are adjusted according to the conflict reason type by matching the preset adjustment strategy, and the multi-constraint rule engine is re-triggered. When the preset iteration limit is reached or there is no improvement for two consecutive rounds, the iteration is terminated and the latest feasible solution and conflict report are output. The parameter library version management module is used to independently store the parameter library according to the policy caliber version number, perform incremental synchronization in response to policy caliber updates, perform new and old version mapping verification and freeze or roll back in-transit calculation tasks during the update, and lock the same policy caliber version for the same batch of batch calculations. The batch processing scheduling module is used to establish rule graph indexes and parameter caches for multiple projects in the project library, solve batch tasks in parallel after splitting them, and summarize them. When a single split fails to solve, it automatically retryes and outputs the reason for failure and a list of unfinished items. All tasks in the same batch use the same policy version. The combination scheme generation module is used to generate a combination scheme output containing rule tracing information based on the closed fund pooling scheme and the hit path.

[0019] Furthermore, the combined scheme generation module includes a text presentation submodule, which is used to convert the combined scheme into an application text, and the funding channels, amounts, proportions and policy basis fields in the output are constrained by the rule traceability information.

[0020] Furthermore, the text presentation submodule adopts a generation plus structured verification mechanism: the input is limited to structured scheme data containing fields such as channel, amount, ratio, policy basis, and corresponding traceability fields; a text draft is generated according to a preset application text template; the funding channel name, amount, ratio, and policy basis fields in the text draft are verified field by field, the funding channel name must be in a preset whitelist, and the value must be exactly consistent with the traceability field. Content that is out of range or inconsistent is rejected and triggers regeneration. After reaching the preset maximum number of retries, it falls back to a deterministic template splicing output.

[0021] To address the aforementioned technical problems, in a second aspect, embodiments of the present invention provide a method for intelligently generating government investment project funding pools, comprising the following steps: S1. Obtain the structured project parameters of the target project and generate a project parameter vector, wherein the project parameter vector includes at least industry category and sub-category fields; S2. Extract the criterion values ​​of each criterion dimension from the project parameter vector according to multiple preset criterion dimensions, and the value spaces of each criterion dimension are independent of each other. S3. The criteria values ​​of each criterion dimension are matched with the funding channel rule graph database to obtain a set of candidate channels. When different criteria dimensions hit different channels and conflicts occur, arbitration is carried out according to the preset priority rules, and the hit path of each channel is recorded. The hit path is used for conflict cause attribution and result tracing. The funding channel rule graph database stores the funding channel hit rules in a graph data structure. The nodes of the graph data structure include channel identifier, support scope, support ratio identifier, exclusivity identifier and negative list identifier, and the edges represent the criteria conditions and pointing relationships. S4. Read the support ratio and leverage constraints of each candidate channel according to the preset parameter library, and calculate the support amount of each channel; S5. Transform the negative list constraints, channel exclusivity constraints, and compliance requirements into a set of compliance constraint conditions; S6. Model the funding amount of each channel as a variable and the funding pooling verification conditions as constraints into a constraint satisfaction problem and solve it. The constraints include total amount closure constraint, channel exclusivity constraint, negative list constraint, compliance constraint, revenue feasibility constraint, and capital ceiling constraint. The total amount closure constraint is the constraint that the sum of the funding amounts of each channel equals the total investment of the project. The channel exclusivity constraint is the constraint determined by the exclusivity identifier in the funding channel rule graph database. The compliance constraint set is incorporated into the constraint satisfaction problem and participates in the construction of the feasible region in the form of constraint equations along with the other constraints, rather than being used as an independent gate to intercept the solution after the output. When a feasible solution exists, output a closed funding pooling scheme. When there is no feasible solution, output the conflict cause, adjust the candidate channels according to the pre-set adjustment strategy according to the conflict cause type, and return to S3 for re-matching and calculation. When the pre-set iteration limit is reached or there is no improvement for two consecutive rounds, terminate the iteration and output the latest feasible solution and conflict report. S7. A combination scheme based on the closed-loop fund pooling scheme and the hit path output of each channel, which includes rule tracing information; S8. Perform batch processing on multiple projects in the project library and output the funding pooling plan for each project. All tasks in the same batch use parameter library data of the same policy version.

[0022] Furthermore, prior to S1, a batch sub-project bundling step is included: for multiple sub-projects with an investment scale lower than the single project application threshold, bundling optimization is performed according to a pre-set grouping rule based on the same application entity, similar industry categories, and complementary revenue attributes. The optimization objective is to maximize the overall revenue coverage multiple after bundling, and a negative list is removed. The bundled project units are then output as inputs to the structured project parameters. In S6, the bundled project units pass a closure check using the overall parameters after bundling.

[0023] Furthermore, the combined scheme output by S7 also includes investment-loan linkage leverage arrangement and application timing; the investment-loan linkage leverage arrangement is triggered based on the target project's entry status in the external project database and the hit status of policy-based funding channels, matching the target project's industry tags with a pre-set bank credit product tag database, and calculating the combined amount of policy-based funding and bank credit; or, based on the target project having a planned but not yet issued special bond quota identifier, a loan replacement arrangement is triggered, replacing the planned but not yet issued special bond quota with bank loans.

[0024] Furthermore, the constraint satisfaction problem is solved using a mixed integer programming model; the revenue feasibility constraint is converted into a cash flow constraint based on the project revenue calculation data, and the revenue feasibility constraint is defined as the ratio of the total net revenue of the project's entire life cycle to the total principal and interest of debt financing reaching the revenue coverage review line, where the revenue coverage multiple is the ratio of the total net revenue of the project's entire life cycle to the total principal and interest of debt financing; the preset adjustment strategies are preset according to the conflict cause type, including adjustment actions corresponding to funding gaps, channel duplication, negative list hits, non-compliance, unmet revenue feasibility, and capital over-limit.

[0025] Furthermore, the parameter library is stored independently according to the policy version number. In response to policy updates, incremental synchronization is performed. During the update, the mapping between the old and new versions is verified and the in-transit calculation tasks are frozen or rolled back. The same policy version is locked for the same batch of calculations.

[0026] Furthermore, the batch processing in S8 includes: establishing a rule graph index and parameter cache according to industry categories, solving the batch tasks in parallel after splitting them into pieces and summarizing the results; automatically retrying and recording the retry log when a single piece fails to solve; and outputting the failure reason and incomplete item list for that piece when the retry still fails.

[0027] Furthermore, the pre-set adjustment strategy based on the type of conflict cause includes: when the conflict cause is a funding gap, the best alternative channel is selected according to the gap amount; when the conflict cause is channel duplication, the highest priority channel is retained and duplicates are eliminated according to the priority arbitration result; when the conflict cause is a negative list, the channel is eliminated and alternative channels are selected; when the conflict cause is non-compliance, the corresponding non-compliant combination is eliminated and alternative channels are selected; when the conflict cause is unsatisfactory profitability, the channel is downgraded to a combination channel or an additional capital-based channel is added; when the conflict cause is excessive capital, the amount of special bonds used as capital is reduced and converted to ordinary special bonds plus bank matching funds.

[0028] Furthermore, the set of compliance constraints is solved jointly with the total amount closure constraint, channel exclusivity constraint, negative list constraint, revenue feasibility constraint, and capital ceiling constraint within the same constraint set. Channel combinations that hit the exclusivity flag and channels that hit the negative list are excluded during the feasible domain construction stage, so as to eliminate non-compliant funding combinations from the source during the scheme construction stage.

[0029] Furthermore, the multiple criteria dimensions include at least revenue attributes, project type, national strategic attributes, industry sub-items, and regional attributes. Each criterion dimension takes an independent value, and the value of any criterion dimension is not determined by the values ​​of other criterion dimensions.

[0030] Furthermore, the preset priority rule is a multi-level priority order configured according to the criteria dimension; when channels hit by different criteria dimensions conflict, the channel hit by the highest priority criteria dimension is retained as the main channel, and the other hit channels are recorded as alternative channels; after locking the main channel according to the highest priority criteria dimension, the alternative channels with an exclusive relationship with the main channel are closed according to the exclusive identifier in the funding channel rule graph database; the candidate channel set output by arbitration is used as the candidate channel variable domain of the constraint satisfaction problem.

[0031] Furthermore, S4 includes: based on the regional attributes of the target project's location, reading the upper limit of the support ratio for each candidate channel from a differentiated support ratio parameter library organized by regional attributes, and calculating the support amount for each channel based on the product of the total project investment and the upper limit of the support ratio; when using special bonds as project capital, calling the capital positive list library to verify the target project's category eligibility, and verifying that the scale of special bonds used as capital does not exceed the preset upper limit of the ratio.

[0032] Furthermore, the entry status of the target project in the external project library is obtained through a standard data interface or batch import method. Both methods share the same entry status snapshot. The batch import method generates an entry status snapshot and adds a version mark at the time of import. Calculations within the same batch are performed based on the same snapshot. If data retrieval fails, it will automatically retry and record the retry log.

[0033] To address the aforementioned technical problems, in a third aspect, embodiments of the present invention provide an electronic device, comprising: a memory for storing a computer program; and a processor for executing the computer program; wherein, when the processor executes the computer program, it implements the intelligent generation method for government investment project funding pooling as described in any of the second aspects above.

[0034] To address the aforementioned technical problems, in a fourth aspect, embodiments of the present invention provide a computer-readable storage medium storing a computer program, which, when executed, implements the intelligent generation method for government investment project funding pooling as described in any of the second aspects above.

[0035] To address the aforementioned technical problems, in a fifth aspect, embodiments of the present invention provide a computer program product, including computer instructions, which, when executed by a processor, implement the intelligent generation method for government investment project funding pooling as described in any of the second aspects above.

[0036] Compared with the prior art, the present invention has the following beneficial effects: The closure of the fund pooling system is verified by overall modeling and constraint solver of five types of constraints. The output scheme must satisfy the constraints of total amount closure, channel exclusivity, negative list, revenue feasibility and capital ceiling, which is different from the existing system's method of "generating text and then manually checking it". Compliance requirements are incorporated into the closed-loop solution in the form of constraint equations and participate in the construction of the feasible region along with other constraints. They are internalized as constraints during the solution construction stage, eliminating non-compliant combinations from the source. This is different from the independent release stage after solution generation in decision-gated systems. Gated implementations require interception and complete redoing after solution generation. However, in the context of government consolidation, compliance is a constructive condition (insufficient revenue coverage means no bond issuance conditions, negative list violation means removal, and duplicate applications between channels are strictly prohibited). "Incorporating into the solution" is an effective path to solve this real technical problem. The multidimensional criteria orthogonal extraction and priority arbitration mechanism resolves the logical conflicts of orthogonal criteria such as benefit attributes, project types, and national strategic attributes in a single tree structure. The arbitration result serves as the input for constraint solving (determining the variable domain), ensuring that each project receives a unique and determined channel combination. Rule graph indexing, parameter caching, and parallel scheduling support batch processing of tens of thousands of provincial-level projects, reducing the generation of a single project plan from several days of manual work to minutes, which is in line with the short application window period; The parameter library version management and hot update consistency verification mechanism ensure that batch calculation results use the same version of the caliber under annual policy iterations (such as adjustments to investment field classification, adjustments to the upper limit of capital ratio, and the coexistence of multiple calibers for return coverage multiples), avoiding outdated calibers and incomparability of cross-batch results caused by hard-coded rules. The negative list and the positive list of capital are structured and verified. Projects containing prohibited content are removed during the generation stage to prevent "application with defects". Automatic calculation of investment-loan linkage leverage and determination of special bonds for loan replacement maximize the available funds for projects. The output of the hit path backtracking and rule source field embedding provides the rule basis, proportion source version and calculation formula for each matching step; the output of the large language model presentation layer is constrained by the source field (generation plus structured verification mechanism) to prevent the generation of unfounded content, which facilitates the preparation of application materials, leadership decision-making and subsequent review and traceability; the hit path also serves to attribute the cause of conflict when there is no feasible solution, forming a closed loop of "matching - solving - attribution - adjustment"; The batch sub-project bundling step supports the combined application of small and scattered projects, which expands the system's coverage of small-scale sub-projects, and the bundling unit is also subject to complete closure and compliance verification. Unlike the general solution of "project portfolio + constraint optimization" (US20140330748A1), this invention outputs a funding portfolio solution that satisfies the five types of closed constraints in government affairs and provides feasibility proof in the form of a problem where constraints can be determined. It also has features that the general solution does not have, such as incorporating compliance constraints into the solution, conflict attribution iteration, and batch consistency of parameter library version lock. Unlike existing paths of "funding routing / priority matching" (CN114240619B, CN112053143B), this invention does not perform "optimal channel selection" for a single project, but rather performs a closed-loop solution for "a patchwork of government affairs with five types of constraints coupled together", and the output includes feasibility proof and traceability, solving the problem of consistency between scale and scope. Attached Figure Description

[0037] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the following description of the embodiments are briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 The system overall architecture diagram provided for embodiments of the present invention; Figure 2 This is a schematic diagram of the data structure of the funding channel rule graph database provided in an embodiment of the present invention; Figure 3 A diagram showing the relationship between the parameter library and the rule engine in an embodiment of the present invention; Figure 4 A flowchart illustrating the method implementation provided in this embodiment of the invention; Figure 5 This is a batch parallel scheduling diagram provided for an embodiment of the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0040] It should be noted that in the embodiments described in this section, the project names, project types, geographical attributes, and various amounts, proportions, multiples, and other data are all illustrative constructed data and do not refer to any real projects, regions, or entities. They are only used to illustrate the specific implementation process of the technical solution of the present invention.

[0041] The core idea of ​​this invention is to use a structured government funding rule engine (multi-dimensional criterion orthogonal extraction, constraint rule matching and priority arbitration, support for ratio and leverage calculation, compliance constraint transformation, five types of constraint closure solution, parameter library version management, batch parallel processing) as the main technology. The large language model only undertakes the solution presentation and interpretation layer, and its output is constrained by the structured results and traceability fields of the rule engine. In this way, the intellectual activity rules of "generating consultation text" are transformed into a processing flow with definite technical meaning for data in specific application fields.

[0042] (I) System Overall Architecture See Figure 1 The system includes: a project parameter parsing module 11, a multi-dimensional criterion extraction module 12, a multi-constraint rule engine 13 (including a funding channel rule graph database 131 and a rule priority arbitration unit 132), a module supporting ratio and leverage constraint calculation 14, a compliance constraint conversion module 15, a funding consolidation closure constraint solution module 16 (including a compliance constraint input interface 161, which merges the compliance constraint condition set output by the compliance constraint conversion module into the constraint set), a combined scheme generation module 17 (including a large language model presentation sub-module 171), a parameter library version management module 18, and a batch processing scheduling module 19. These modules form a pipeline of "parameter parsing → criterion extraction → rule matching and arbitration → constraint calculation → compliance constraint conversion → overall solution of five types of constraints → scheme output," and support batch parallel processing of provincial project databases. Among them, the compliance constraint transformation module 15 and the closure constraint solution module 16 are coupled through the compliance constraint input interface 161: the compliance conditions are not used as a gate control link independent of the solution, but are transformed into constraint equations and incorporated into the solution. They are solved together with the total amount closure, channel exclusivity, negative list, revenue feasibility and capital ceiling constraints in the same constraint set and jointly construct the feasible region.

[0043] (II) Core Data Structure (1) Funding Channel Rule Graph Database: This database stores funding channel matching rules using a graph data structure. Nodes include at least the channel identifier, support scope, support ratio identifier, exclusivity identifier, negative list identifier, and applicable criterion dimension. Edges represent criterion conditions and pointing relationships. The database is based on business specifications and reporting standards. Business experts manually or semi-manually input current policy standards using pre-set templates. After consistency verification, the data is stored and maintained synchronously according to a versioning mechanism when policy standards are updated. During the matching process, the matching path (criterion dimension, rule identifier, node sequence) is recorded for result backtracking and tracing, and serves as the basis for attributing conflict causes when S6 has no feasible solution (the matching path identifies which criterion dimension and which channel rule the conflict occurred in). For node and edge definitions, matching path records, and matching examples in the funding channel rule graph database, please refer to [link to database documentation]. Figure 2 .

[0044] (2) Parameter library system: This includes a differentiated support ratio parameter library (organized according to the regional attributes of the eastern, central and western regions), a negative list library (including the negative list of special bonds and the negative list of ultra-long-term special treasury bonds), a capital positive list library, a revenue coverage review line parameter library, and a bank credit product tag library (e.g., 8 primary tags, 42 secondary tags, and 87 tertiary tags). Each parameter library is stored independently according to the policy caliber version number. Incremental synchronization is performed when the policy caliber is updated, and consistency verification is performed during the update period (new and old version mapping verification, freezing or rolling back of in-transit calculation tasks) to ensure that the same version caliber is used for batch calculations in the same batch. For the calling relationship between the parameter library and the rule engine and the versioning mechanism, please refer to [link to relevant documentation]. Figure 3 .

[0045] The parameter library content is stored in a structured field format. The field schema examples are as follows (each field is configurable data in the parameter library and does not constitute a limitation on the scope of protection of the claims): Channel Identifier, Support Scope, Support Ratio Upper Limit, East-Central-West Difference, Exclusivity Identifier, Negative List Identifier, Applicable Criterion Dimension, Threshold Condition, Version Number.

[0046] Example of parameter library rules (the values ​​are for illustrative purposes only): ① Urban sewage and garbage treatment (sub-item of central budget investment): Support ratio 30% in the east / 45% in the central region / 60% in the west, with a total investment threshold of no less than 100 million yuan; ② Revenue coverage assessment line: 1.3 for special bonds (national minimum 1.2, 1.5 for some regions); ③ Negative list of special bonds: government buildings, image projects, real estate development, theme parks, office buildings of party and government organs, etc., will be excluded if they are included; ④ Ultra-long-term special treasury bonds and central budget investment: Exclusivity is marked as "choose one of the three, strictly prohibit duplicate applications, and misaligned arrangements"; ⑤ Investment-loan linkage leverage: government funds leverage bank credit by about 4-5 times. The above rules indicate that the rule engine processes structured data with specific technical meaning (channel identifier, support ratio, exclusivity mark, negative list mark, and criterion dimension), rather than abstract "business rules".

[0047] (3) External data interface: The entry status of the project in the National Major Construction Project Database is obtained through the standard data interface or batch import method, which serves as the trigger condition for investment-loan linkage judgment. The two acquisition methods share the same entry status snapshot (batch import generates a snapshot and adds a version mark at the time of import, and the standard interface ensures that the data retrieved within the same batch is consistent according to the interface call convention. If the data retrieval fails, it will be automatically retried and the retry log will be recorded), so as to avoid inconsistencies in the entry status in the same batch calculation.

[0048] (III) Method Implementation Steps See the method implementation process. Figure 4 It includes steps S0 to S8, where S0 is an optional preliminary step.

[0049] S0 Batch Sub-project Bundling (Optional Pre-step): For sub-project sets with investment scales lower than the single-project application threshold, bundling optimization is performed according to the pre-defined grouping rules of "same applicant entity + similar industry category + complementary revenue attributes"—the decision variable is whether each sub-project is packaged into the same application unit, and the constraints are consistency of the applicant entity, consistency of industry category, and exclusion from the negative list for the packaged unit. The objective is to maximize the overall revenue coverage multiple after bundling; the output is the bundled project unit, which serves as the input for S1. Bundling is an optional step; if the bundling conditions are not met, the project directly enters S1 as a single project. The bundled unit must pass the overall closure check after packaging in S6.

[0050] S1 Project Parameter Analysis and Industry Classification: This involves obtaining structured parameters such as project name, construction content, total investment, construction nature (new construction / renovation / expansion / equipment upgrade), national economic industry category and subcategories, project location (province / city / county and regional attributes of eastern, central, and western China), and revenue calculation data (annual revenue, total life cycle cost, and operating period) to form a project parameter vector. The project is mapped to the corresponding industry category (20 categories A-T, with international organization category T generally not involving fixed asset investment declarations; actual mapping is mainly based on A-S) and subcategories according to the "National Economic Industry Classification" (GB / T 4754-2017), serving as the index key for channel matching. If the input is a bundled unit output from S0, the project parameter vector is generated from the packaged overall parameters.

[0051] S2 Multidimensional Criterion Extraction: This extracts multidimensional criteria from the project parameter vector—revenue attributes (no revenue / revenue with a revenue coverage ratio reaching the review threshold / revenue but not reaching the review threshold / revenue uncertain), project type (new infrastructure / renovation / equipment upgrade), national strategic attributes (whether it belongs to the "two-fold" investment direction, whether it belongs to the "two-new" equipment upgrade field), industry sub-items, and regional attributes. Each criterion is independently valued and not nested, avoiding conflicts between parent and child criteria in a single tree structure. Here, "orthogonality" means that the value spaces of each criterion dimension are independent (revenue attributes, project type, national strategic attributes, industry sub-items, and regional attributes are valued separately and do not serve as judgment conditions for each other); "priority arbitration" means that when different criterion dimensions conflict with different channels, a decision is made according to the pre-set priority rules of S3. The arbitration result serves as input for solving the closed-loop constraint problem (determining the candidate channel variable domain for satisfying the constraint). Orthogonality determines how the criteria are valued, and arbitration determines how to converge and constrain when the results conflict. These two are mechanisms at different levels and do not conflict with each other.

[0052] S3 Multi-Constraint Rule Matching and Priority Arbitration: Each criterion is matched against the funding channel rule graph database. Matching channels are the union of those satisfying the corresponding constraints. When conflicts arise due to different criteria matching different channels, they are resolved according to pre-set priority arbitration rules: First priority is for criteria specific to national strategic attributes or equipment upgrade attributes (e.g., "two-fold / two-new" criteria), prioritizing the locking of corresponding specific channels upon matching (based on the requirement that applications for such channels must include a commitment not to repeatedly apply for other central government funding, thus having exclusive locking effect); Second priority is for revenue criteria: projects with a revenue coverage ratio reaching the review threshold qualify for local government special bonds, while those not reaching the review threshold are downgraded to combined channels; Third priority is for industry sub-item criteria: specific funding sub-items (e.g., the "urban sewage and garbage treatment" sub-item within the central budget, and the corresponding category of special bonds) and alternative channels are retrieved from the rule graph database according to industry category. The above priority order, specific channel names, and review threshold values ​​are all configurable data in the parameter library, and this invention is not limited to these. The candidate channel set and matching path output by arbitration serve as variable domain inputs for constraint solving. During the matching process, the hit path of each channel is recorded. The hit path is used for attribution of conflict causes when there is no feasible solution in S6 and for tracing the results in S7.

[0053] S4 Support Ratio and Leverage Constraint Calculation: Based on the project's geographical location (East, Central, West), the upper limit of the support ratio for each candidate channel is read from the differentiated support ratio parameter library, and the support amount is calculated (support amount = total project investment × upper limit of support ratio). Capital Determination: When special bonds are used as project capital, the capital positive list library is called to verify the project category eligibility and to verify that the scale of special bonds used as capital does not exceed the prescribed upper limit of the province's special bond quota. Investment-loan linkage and loan replacement are two independent parallel judgment scenarios: (a) Investment-loan linkage (leverage superposition): The trigger condition is that the project has been confirmed by the external data interface to be included in the national major construction project database and has hit the policy-based funding channel. The action is to match the project's industry tag with the bank credit product tag library, calculate the combined quota of policy-based funds and bank credit, and realize the leverage of government funds to bank credit; (b) Loan replacement of special bonds: The trigger condition is that the project has a special bond quota identifier of "planned but not yet issued". The action is that the bank intervenes in advance and replaces the planned but not yet issued special bond quota with a loan to alleviate the project's funding pressure during the bond issuance waiting period. The two scenarios are triggered independently based on their respective data premises, and the output enters the initial funding pooling scheme.

[0054] S5 Compliance Constraint Transformation: The negative list constraints, channel exclusivity constraints, and compliance requirements (such as fund management compliance, data security and sharing compliance, audit compliance, etc.) involved in the initial funding scheme are transformed into a set of compliance constraint conditions—the negative list library (special bond negative list: government buildings, image projects, real estate development, theme parks, party and government office buildings, etc.; ultra-long-term special treasury bond negative list). If a channel is matched, it is removed from the list. Channel exclusivity rules (ultra-long-term special treasury bonds and central budget investment should not overlap in principle, and the same project should not repeatedly apply for the same central funding) are transformed into channel deduplication constraints. Compliance requirements are transformed into compliance sub-constraints that can be incorporated into the constraint solution. The aforementioned set of compliance constraint conditions is not used as a gating mechanism independent of the solution process, but is incorporated into the constraint satisfaction problem of S6 through the constraint set interface, participating in the feasible region construction simultaneously with other constraints in the form of constraint equations.

[0055] S6 Funding Pool Closure Constraint Solution: The validation model of the initial funding pooling scheme is modeled as a constraint satisfaction problem—the variables are the funding amounts of each channel; the constraint set includes five types of funding pooling constraints and compliance constraints: total amount closure constraint (the sum of the funding amounts of each channel equals the total project investment), channel exclusivity constraint (the exclusivity between different channels is determined by the exclusivity identifier in the funding channel rule graph database; combinations that hit the exclusivity identifier are excluded during the feasible region construction stage), negative list constraint, compliance constraint, revenue feasibility constraint (the project revenue calculation data is converted into cash flow constraints, the caliber is: revenue coverage multiple = total net revenue of the project throughout its entire life cycle ÷ total principal and interest of debt financing, reaching the revenue coverage review line), and capital ceiling constraint; if necessary, linearization is performed and incorporated into the mixed integer programming model. The compliance constraint set is incorporated into the constraint satisfaction problem in the form of constraint equations, participating in the feasible region construction along with the other constraints, rather than being used as an independent gating interception after the solution result is output—compliance conditions exclude non-compliant combinations from the source during the scheme construction stage. The constraint solver is invoked to solve the problem: if a feasible solution exists, the validated funding pooling scheme is output; if no feasible solution exists, the conflict cause is output, and a preset adjustment strategy is matched according to the conflict cause type. After adjusting the candidate channels, the process returns to S3 for re-matching and calculation. The correspondence between conflict cause types and preset adjustment strategies is shown in the table below (this correspondence supports the dependent claims):

[0056] The solution process sets an upper limit on the number of iterations (preset value, 5 times in the example). The iteration terminates when the upper limit is reached or when there is no improvement for two consecutive rounds, and outputs "the most recent feasible solution + conflict report" to ensure that the method terminates within a finite number of steps and gives a definite result.

[0057] S7 Combined Solution Output: Outputs the main solution and alternative solutions. Each solution includes channel combinations, their respective supported ratios and amounts, investment-loan linkage leverage arrangements, application timelines (e.g., application windows in October-November of the current year and March-April of the following year), key constraints, and rule traceability information (hit rule identifiers, support ratio source version numbers, and support amount calculation formulas). Under the constraints of the rule traceability information, the large language model presentation submodule converts the structured solutions into capital balance solutions, application instructions, and other application material texts. The implementation mechanism of the large language model presentation submodule is "generation + structured validation": the input is limited to structured solution data (JSON, including channel, amount, ratio, policy basis fields, and their traceability fields); the large language model generates a draft according to a pre-set application text template; the output undergoes strong field-by-field validation by a structured validator—the channel name, amount, ratio, and policy basis fields must be consistent with the traceability fields (names must be within the whitelist, and values ​​must match precisely). Content that exceeds the range or is inconsistent is rejected and triggers regeneration (maximum of 3 retries; if it still fails, it falls back to a deterministic template concatenation), ensuring that the presentation layer does not generate any content other than the traceability information.

[0058] S8 Batch Processing: This step generates batch funding pools for multiple projects (in the tens of thousands) in the provincial investment project database. It establishes rule graph indexes and parameter caches by industry category, schedules batch matching tasks (task sharding, parallel solving, and result aggregation) based on a parallel computing framework, and outputs funding pooling schemes and verification reports for each project in batches, ensuring that tasks in the same batch use the same policy version. If a single shard fails to solve, it automatically retryes and records the retry log. If the retry still fails, it outputs the reason for the failure and a list of incomplete items. Abnormal tasks during version updates trigger consistency checks and rollbacks. The "tens of thousands" scale addressed in this step refers to an objectively existing project scale: taking a provincial government investment project database as an example, its regularly reserved projects have reached several thousand, and the batch processing scale reaches tens of thousands. Manually compiling each project individually cannot meet the short application window. Rule graph indexing and parallel sharding are necessary means to solve this scale problem, not optional engineering optimizations. See the batch parallel scheduling process for details. Figure 5 .

[0059] (iv) Examples Example 1 (Urban sewage and garbage treatment project, total investment of 500 million yuan, central region) ① Criterion Extraction (S2) – Revenue Attribute: There is revenue (sewage treatment fees and reclaimed water reuse income). Calculations show that if the full debt financing is used, the coverage ratio is approximately 1.0, which is lower than the 1.3 review threshold, and therefore does not meet the conditions for separate full financing of special bonds; Project Type: New Construction; National Strategic Attribute: No; Industry Sub-item: Water Conservancy, Environment and Public Facilities Management (N categories) "Urban Sewage and Garbage Treatment". ② Rule Matching (S3) – First Priority (Two-fold / Two-new) Not Matched; Second Priority Revenue Criterion: Coverage is insufficient to meet the review threshold, special bonds are not the main channel but can be used as capital (the category is within the positive list of capital); Third Priority Industry Sub-item Criterion: Matches the central budget investment "Urban Sewage and Garbage Treatment" sub-item (the support ratio in the central region is capped at 45%, and the total investment threshold of not less than 100 million yuan is met), the main channel is locked in the central budget. ③ Quota Calculation (S4) – Central budgetary support quota = 500 million yuan × 45% = 225 million yuan; special bonds as capital: 100 million yuan (accounting for 20% of total investment, not exceeding the prescribed upper limit); remaining funding requirement = 500 million - 225 million - 100 million = 175 million yuan. ④ Investment-Loan Linkage (S4) – The project has been confirmed by the data interface to be included in the national major construction project database, the industry tag matches the bank credit product, and the bank provides a matching loan of 175 million yuan. ⑤ Constraint Solution (S6) – Variables x1=2.25 (Central Budget), x2=1 (Special Bonds), x3=1.75 (Bank Loans); Constraint Verification: Σx=500 million yuan = Total Investment (Closed-loop solution passed); Channel Exclusivity (Central Budget and Special Bonds are staggered, passed); Negative List Not Hit (Passed); Compliance Constraints Met (Passed); Capital of 100 million yuan does not exceed the prescribed limit (Passed); Profit Feasibility: The project's annual net income is approximately 22 million yuan, with a total operating period of approximately 660 million yuan over 30 years. The total principal and interest of 275 million yuan from special bonds and bank loans is approximately 490 million yuan, with a coverage ratio of approximately 1.35, reaching the 1.3 review threshold (Passed); A feasible solution exists, output plan: 45% from the Central Budget + 20% from Special Bonds (as capital) + 35% from Bank Loans (investment-loan linkage). ⑥ Source Information (S7): Hit rule identifier, proportion source version, calculation formula (2.25=5×45%, etc.) are output with the plan and embedded in the application text.

[0060] Example 2 (Metropolitan Area Intercity Railway Project, Total Investment of 20 Billion Yuan, Eastern Region) ① Criterion Extraction (S2) – National Strategic Attribute: Yes (included in the intercity railway planning of key metropolitan areas, belonging to the newly added investment direction of "two major"); Revenue Attribute: Revenue exists (ticketing and properties along the line), coverage multiple is about 1.1, lower than the review line; Project Type: New Construction; Industry Sub-item: Transportation Industry (G Category) Railway. ② Rule Matching (S3) – The first priority "two major" exclusive criterion is hit, and the ultra-long-term special treasury bonds (two major) are prioritized as the main channel; the revenue criterion and industry sub-item criterion (the upper limit of the support ratio of the railway sub-item in the central budget shall not exceed 60%) are used as alternative records; because the "two major" application must promise not to apply for other central funds repeatedly, after the treasury bond channel is locked, the central budget alternative channel is closed according to the exclusivity rule. ③ Quota Calculation (S4) – The support quota of ultra-long-term special treasury bonds is determined according to the special standard. ④ Investment and Loan Linkage (S4) – The treasury bond funds are used as project capital and matched with syndicated loans (referring to the syndicated loan model of large water conservancy projects). ⑤ After the constraint solution (S6) passes, the combined solution of "ultra-long-term special treasury bonds + syndicated loan" is output. This example illustrates the process of resolving conflicts between the national strategic attribute criteria and the benefit attribute criteria by arbitration according to priority. It should be noted that the use of ultra-long-term special treasury bond funds as project capital is subject to the issuance and use arrangements of ultra-long-term special treasury bonds in the current year (see the policy interpretation in section (VI) of this manual for details).

[0061] Example 3 (Equipment upgrade project in an industrial park, total investment of 200 million yuan, in the eastern region) ① Criterion Extraction (S2) – Project Type: Equipment Renewal and Upgrading; National Strategic Attribute: Belongs to the "Two New" Equipment Renewal Field (Industrial Direction); Revenue Attribute: Operating Revenue; Industry Sub-item: Manufacturing (Category C) General Equipment Manufacturing. ② Rule Matching (S3) – The first priority "Two New" exclusive criterion hits the ultra-long-term special treasury bond (Two New – Industrial Sector Equipment Renewal); According to the exclusivity rule, once the "Two New" channel is hit, the central budget's industrial sector equipment renewal sub-item will not be applied for repeatedly. ③ Quota Calculation (S4) – Assume that the annual support ratio limit for "Two New" – Industrial Sector Equipment Renewal is r=15% (illustrative value, based on the current year's caliber), and the support quota = 200 million yuan × 15% = 30 million yuan. ④ Constraint Solution (S6) – Variables x1=0.3 (two new types of ultra-long-term special treasury bonds), x2=1.7 (enterprise self-financing and capital); Constraint Verification: Σx=200 million yuan = total investment (closed loop passed); Exclusivity (two new types of bonds and two major types of bonds, central budget funds do not overlap, passed); Negative list not hit (passed); Compliance constraints met (passed); Profitability Feasibility: The project's annual net income is approximately 25 million yuan, covering the debt funding gap (x1 corresponds to principal and interest) without pressure (passed); Capital Constraint: The proportion of self-raised funds for the equipment renewal project meets the requirements (passed); A feasible solution exists, outputting a single-channel solution: 15% from ultra-long-term special treasury bonds (two new types of bonds) + 85% from self-financing. This example illustrates the complete calculation chain of the single-channel solution when the "two new types of bonds" specific criterion is met.

[0062] Example 4 (Bundled renovation project of old urban communities, corresponding to S0 bundling step) ① Bundling (S0) – Includes 52 sub-projects such as water conservancy, roads, parking, and elderly care. The investment of any single sub-project is less than the threshold for special bond application. Bundling optimization is carried out according to the grouping rule of "same applicant + similar industry category + complementary revenue attributes". The 52 sub-items are packaged into 1 application unit: profit-making sub-items such as parking fees, charging piles, advertising space, and public property rental are matched with purely public welfare sub-items, and negative list items such as sculptures and greening are removed. ② Criterion Extraction (S1 / S2) – Project type: urban renewal / renovation; revenue attributes: the revenue of individual sub-projects varies greatly (uneven profitability), and the overall packaged integration has revenue. ③ Rule Matching (S3) – The bundled application unit as a whole hits the urban renewal category of special bond (illustrated data: the total investment of the bundled application unit is about 3 billion yuan, the planned issuance of special bonds is about 1.8 billion yuan, accounting for about 60% of the total investment, and the overall debt repayment reserve ratio is about 1.6). ④ Constraint Solving (S6) – The overall revenue coverage multiple after packaging is approximately 1.6, reaching the review threshold (passed); total amount closure (passed); compliance constraints are met (passed); output solution. ⑤ Supporting Mechanisms (S4) – Before the issuance of special bonds, policy-based financial instruments can act as a "bridge," and after issuance, the bonds can be withdrawn. This example illustrates how the system uses the S0 bundling optimization capability (combining rich and poor projects, bundling small and scattered projects, and removing negative list items) to achieve batch packaging of sub-projects, and explains the overall closure verification of the bundling unit after packaging in S6.

[0063] Example 5 (Loan replacement of special bonds for cross-sea channel projects, corresponding to the S4 loan replacement scenario) A cross-sea channel project has a total investment of 13 billion yuan. The project has been included in the national major construction project database, and a special bond quota of 10 billion yuan has been planned but not yet issued (there is a special bond quota marker of "planned but not yet issued"). ① Criterion Extraction (S2) - Industry Sub-item: Transportation Industry (G Category) Highway / Cross-sea Channel; Revenue Attribute: There is toll revenue; National Strategic Attribute: Yes. ② Loan Replacement Judgment (S4) - The "planned but not yet issued" special bond quota marker is detected, triggering the loan replacement scenario: The bank intervenes in advance, replacing the planned but not yet issued special bond quota with a loan of 5.2 billion yuan, releasing more than 5 billion yuan of bond quota space (of the original 10 billion yuan special bond quota, 5.2 billion yuan is replaced by bank loans, and the remaining quota is re-issued according to procedures, realizing the release of bond space and alleviating the funding pressure during the bond issuance waiting period). ③ Constraint Solving (S6) – Variables: Loan replacement of 5.2 billion yuan + Special bonds (subsequent issuance) of 5 billion yuan + Capital and other funds of 2.8 billion yuan, Σ=13 billion yuan=Total investment (closed loop successful); Exclusivity and compliance constraints satisfied (passed); Profitability: Fee revenue covers principal and interest of debt funds (passed); Output scheme: Loan replacement of special bonds + subsequent issuance of special bonds + Capital. This example illustrates the system's identification of the "planned but not yet issued" special bond quota and the loan replacement trigger determination, as well as the closed loop calculation of funds after releasing bond space.

[0064] Example 6 (No feasible solution, backflow, or termination, corresponding to S6) A municipal road project, with a total investment of 300 million yuan, is located in the central region and has a low-yield nature (coverage multiple of approximately 0.9). ① Initial Solution (S6) – The initial candidate channel is the full amount of 300 million yuan in special-purpose bonds; constraint verification: total amount closure (passed); revenue feasibility fails (coverage multiple 0.9 < 1.3 review line, not up to standard); output conflict reason "revenue feasibility not up to standard". ② Targeted Adjustment (S6) – According to the pre-set adjustment strategy "revenue feasibility not up to standard → downgrade to combined channel": special-purpose bonds are downgraded to a combined channel, and the central government budgetary investment (the upper limit of the central region's support ratio is 45%) is extended to the candidate channel. ③ Secondary Solution (S6) – Variables: 135 million yuan from the central budget + 45 million yuan from special bonds as capital + 120 million yuan from bank loans, Σ=300 million yuan = total investment (closed loop approved); Feasibility of returns: Total principal and interest of debt financing is approximately 230 million yuan, total net income over the entire project lifecycle is approximately 290 million yuan, coverage ratio is approximately 1.26 < 1.3 review line, still not up to standard; Output conflict reason: "Feasibility of returns not up to standard (gap of approximately 8 million yuan)". ④ Tertiary Solution (S6) – Continue to reduce and compress debt financing according to the adjustment strategy: 135 million yuan from the central budget + 45 million yuan from special bonds as capital + 90 million yuan from bank loans + 30 million yuan from enterprise self-raised funds, Σ=300 million yuan (closed loop approved); Total principal and interest of debt financing is approximately 210 million yuan, coverage ratio is approximately 1.38, reaching the 1.3 review line (up to standard); Compliance constraints are met (up to standard); A feasible solution exists, output the plan. This example illustrates the process of targeted adjustment based on the type of conflict cause when no feasible solution is found, iterative backflow until convergence, and the termination guarantee of the iteration limit (5 times) and "termination after two consecutive rounds without improvement" (in this example, convergence occurred after only 2 iterations, far below the limit). In this example, the conflict cause attribution is based on the location of the hit path of each channel recorded by S3 (the conflict occurs on the review line constraint of the benefit criterion dimension), and the synergy between the hit path and the conflict attribution is reflected here.

[0065] Example 7 (batch production with consistent version specifications, corresponding to S8) A batch of funding aggregation was performed on 5,000 projects in the project database, including Project A (equipment upgrade, total investment of 120 million yuan). ① Version Locking (S8) – At the start of the batch calculation, the current policy version in the parameter database is V2026-02. All 5,000 projects in the batch (including Project A) are uniformly locked to V2026-02 for matching and solving. ② Result Output (S7 / S8) – Project A outputs a support amount of 18 million yuan (maximum proportion of 15%) under the V2026-02 policy version, and the output traceability field includes "policy version V2026-02". ③ Cross-batch comparison (S8) – The previous batch used the V2025-01 standard, and Project A's supported quota under this standard was 0.12 billion yuan (the maximum proportion is 10%). The results of the two batches differ due to version differences. After the version update, the system verifies the source of the difference through a mapping check between the old and new versions (in this example, the maximum proportion adjustment), and marks the version number in the batch report to ensure that the cross-batch results are interpretable and comparable. ④ Anomaly handling (S8) – If the policy standard is updated to V2026-03 during batch execution, the shards already started before the update are completed according to the V2026-02 standard and marked with the version number. Shards not started are switched to the V2026-03 standard for execution. The two sets of results are listed separately and the version differences are recorded to avoid incomparable results caused by mixing standards within the same batch. This example illustrates the parameter library version lock's mechanism for ensuring comparability between batches of the same version and across batches. The policy interpretation is indeed frequently iterated: the classification and management of investment areas, the upper limit of capital ratio, the scope of capital investment areas, the minimum return coverage ratio and the value of the review line, etc., all have multiple interpretations coexisting with the policy version (illustration). The version lock is designed to deal with this coexistence of multiple interpretations and annual drift.

[0066] (v) Examples of electronic devices, computer-readable storage media, and computer program products This invention also provides an electronic device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program. When the processor executes the computer program, it implements the intelligent generation method for government investment project funding pooling described in any of the above embodiments. The processor may be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, field-programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.; the memory may include high-speed random access memory, or non-volatile memory, etc.

[0067] This invention also provides a computer-readable storage medium storing a computer program thereon. When executed, the computer program implements the intelligent generation method for government investment project funding pooling described in any of the above embodiments. The computer-readable storage medium may include various media capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory, random access memory, magnetic disk, or optical disk.

[0068] This invention also provides a computer program product, including computer instructions, which, when executed by a processor, implement the intelligent generation method for government investment project funding pooling described in any of the above embodiments.

[0069] (vi) Explanation of Policy Scope The policy parameters involved in the embodiments of this specification, including but not limited to the upper limit of the support ratio for each funding channel and its differentiated values ​​in the eastern, central and western regions, the national minimum and regional values ​​of the revenue coverage review line, the classification of special bond investment areas and their negative lists, the scope and upper limit of special bonds as capital, the leverage ratio of investment-loan linkage, and the application window arrangements, are all exemplary data under the corresponding policy versions, and are subject to the national and provincial policy standards of the current year. All the above parameters are configurable data in the parameter library of the system, and are incrementally synchronized through the parameter library versioning management mechanism when the policy is updated, which does not constitute a limitation on the scope of protection of this invention.

[0070] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A government investment project funding pooling intelligent generation system, characterized in that, include: The project parameter parsing module is used to obtain the structured project parameters of the target project and generate a project parameter vector, which includes at least industry category and sub-category fields; The multidimensional criterion extraction module is used to extract the criterion values ​​of each criterion dimension from the project parameter vector according to multiple preset criterion dimensions, and the value spaces of each criterion dimension are independent of each other. The funding channel rule graph database stores funding channel matching rules in a graph data structure. The nodes of the graph data structure include channel identifier, support scope, support ratio identifier, exclusivity identifier and negative list identifier. Edges represent judgment conditions and pointing relationships. During the matching process, the matching path is recorded for conflict cause attribution and result tracing. The multi-constraint rule engine is used to match the criteria values ​​of each criterion dimension with the funding channel rule graph database to obtain a set of candidate channels. When the hit results of different criteria dimensions conflict, the engine arbitrates according to the preset priority rules and outputs the hit path of each channel. The module supports the calculation of support ratios and leverage constraints, which reads the support ratios and leverage constraints of each candidate channel from a preset parameter library and calculates the support amount for each channel. The module for solving the closure constraint of the fund pooling scheme is used to model the fund amount of each channel as a variable and the fund pooling verification conditions as constraints into a constraint satisfaction problem and solve it. The constraints include total amount closure constraint, channel exclusivity constraint, negative list constraint, compliance constraint, revenue feasibility constraint, and capital ceiling constraint. The compliance constraint is a set of compliance constraint conditions obtained by transforming the negative list constraint, channel exclusivity constraint, and compliance requirements. The set of compliance constraint conditions is incorporated into the constraint satisfaction problem in the form of constraint equations and participates in the construction of the feasible region along with the other constraints, rather than being used as an independent gating interception after the solution is output. When a feasible solution exists, a closed fund pooling scheme is output; when no feasible solution exists, the conflict reason is output and iterative adjustments are made. The parameter library version management module is used to store the parameter library independently according to the policy caliber version number, and lock the same policy caliber version for the same batch of batch calculations. The batch processing scheduling module is used to establish rule graph indexes and parameter caches for multiple projects in the project library, and to solve and summarize batch tasks in parallel after splitting them into chunks. The combination scheme generation module is used to generate a combination scheme output containing rule tracing information based on the closed fund pooling scheme and the hit path.

2. A method for intelligently generating government investment project funding pools, characterized in that, Includes the following steps: S1. Obtain the structured project parameters of the target project and generate a project parameter vector, wherein the project parameter vector includes at least industry category and sub-category fields; S2. Extract the criterion values ​​of each criterion dimension from the project parameter vector according to multiple preset criterion dimensions, and the value spaces of each criterion dimension are independent of each other. S3. The criteria values ​​of each criterion dimension are matched with the funding channel rule graph database to obtain a set of candidate channels. When different criteria dimensions hit different channels and cause conflicts, arbitration is carried out according to the preset priority rules, and the hit path of each channel is recorded. The funding channel rule graph database stores the funding channel hit rules in a graph data structure. Its nodes include channel identifier, support scope, support ratio identifier, exclusivity identifier and negative list identifier, and the edges represent the criteria conditions and pointing relationships. S4. Read the support ratio and leverage constraints of each candidate channel according to the preset parameter library, and calculate the support amount of each channel; S5. Transform the negative list constraints, channel exclusivity constraints, and compliance requirements into a set of compliance constraint conditions; S6. Model the funding amount of each channel as a variable and the fund pooling verification conditions as constraints into a constraint satisfaction problem and solve it. The constraints include total amount closure constraint, channel exclusivity constraint, negative list constraint, compliance constraint, revenue feasibility constraint, and capital ceiling constraint. The compliance constraint set is incorporated into the constraint satisfaction problem in the form of constraint equations and participates in the construction of the feasible region along with the other constraints, rather than being used as an independent gating interception after the solution is output. When a feasible solution exists, output a closed fund pooling scheme; when no feasible solution exists, output the conflict reason and iteratively adjust. S7. A combination scheme based on the closed-loop fund pooling scheme and the hit path output of each channel, which includes rule tracing information; S8. Perform batch processing on multiple projects in the project library and output the funding pooling plan for each project. All tasks in the same batch use parameter library data of the same policy version.

3. The intelligent generation method for government investment project funding pooling as described in claim 2, characterized in that, Before S1, there is also a step of bundling sub-projects in batches: for multiple sub-projects with an investment scale lower than the threshold for single project application, the bundling combination optimization is performed according to the pre-set grouping rules of the same application entity, similar industry category, and complementary income attributes. The optimization objective is to maximize the overall income coverage multiple after bundling, and the negative list is removed. The bundled project units are output as the input of the structured project parameters. The bundled project units pass the closure check in S6 with the overall parameters after bundling.

4. The intelligent generation method for government investment project funding pooling as described in claim 2, characterized in that, The combined scheme output by S7 also includes investment-loan linkage leverage arrangement and application timing; the investment-loan linkage leverage arrangement is triggered based on the target project's entry status in the external project database and the hit status of policy-based funding channels, matching the target project's industry tags with a pre-set bank credit product tag database, and calculating the combined amount of policy-based funding and bank credit; or, based on the target project having a planned but not yet issued special bond quota identifier, a loan replacement arrangement is triggered, replacing the planned but not yet issued special bond quota with bank loans.

5. The intelligent generation method for government investment project funding pooling as described in claim 2, characterized in that, The constraint satisfaction problem is solved using a mixed integer programming model; the revenue feasibility constraint is converted into a cash flow constraint based on the project revenue calculation data, and the revenue coverage multiple reaching the revenue coverage review line is used as the revenue feasibility constraint, whereby the revenue coverage multiple is the ratio of the total net revenue of the project throughout its entire life cycle to the total principal and interest of debt financing; the iterative adjustment when there is no feasible solution matches a pre-set adjustment strategy according to the conflict cause type, whereby the conflict cause types include funding gap, channel duplication, negative list hit, non-compliance, unmet revenue feasibility, and capital over-limit, and each conflict cause type corresponds to a pre-set adjustment action.

6. The intelligent generation method for government investment project funding pooling as described in claim 2, characterized in that, The set of compliance constraints, together with the total amount closure constraint, channel exclusivity constraint, negative list constraint, revenue feasibility constraint, and capital ceiling constraint, are solved in the same constraint set. Channel combinations that hit the exclusivity flag and channels that hit the negative list are excluded during the feasible domain construction stage, so as to eliminate non-compliant funding combinations from the source during the scheme construction stage.

7. The intelligent generation method for government investment project funding pooling as described in claim 2, characterized in that, The multiple criterion dimensions include at least revenue attributes, project type, national strategic attributes, industry sub-items, and regional attributes. Each criterion dimension takes independent values, and the value of any criterion dimension is not used as a judgment condition based on the values ​​of other criterion dimensions. The preset priority rule is a multi-level priority order configured according to the criterion dimensions. When channels hit by different criterion dimensions conflict, the channel hit by the highest priority criterion dimension is retained as the primary channel, and the remaining channels are recorded as alternative channels. The alternative channels with an exclusive relationship with the primary channel are closed according to the exclusive identifier in the funding channel rule graph database. The candidate channel set output by arbitration serves as the candidate channel variable domain for the constraint satisfaction problem.

8. The intelligent generation method for government investment project funding pools as described in claim 2, characterized in that, The parameter library is stored independently according to the policy version number. In response to policy updates, incremental synchronization is performed. During the update, the mapping between the old and new versions is verified, and in-transit calculation tasks are frozen or rolled back. The batch processing in S8 includes: establishing a rule graph index and parameter cache according to industry categories, solving batch tasks in parallel after splitting them into pieces and summarizing the results. When a single piece fails to solve, it is automatically retried and a retry log is recorded. If the retry still fails, the failure reason and incomplete item list of the piece are output. The same policy version is locked for the same batch of batch calculations.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program; When the processor executes the computer program, it implements the intelligent generation method for government investment project funding pools as described in any one of claims 2 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed, implements the intelligent generation method for government investment project funding pooling as described in any one of claims 2 to 8.

Images