Apartment financial management intelligent auxiliary method and system based on artificial intelligence

Abstract

The embodiment of the invention relates to the technical field of artificial intelligence, and provides an apartment financial management intelligent auxiliary method and system based on artificial intelligence, and the method comprises the steps: obtaining apartment management data, hotel management data and transaction data disclosed by a market, and obtaining multi-source data; storing the multi-source data to a cloud financial management center and encrypting the multi-source data; outputting a prediction result of the multi-source data through a financial analysis prediction network; the financial analysis and prediction network comprises an income prediction model and a component analysis optimization model which run in parallel; constructing a visual report form and an analysis report based on the prediction result, and dynamically displaying the report form and the analysis report to the user; and inputting the multi-source data into the financial management model, performing analysis through an auditing rule in the financial management model to generate a financial auditing report, calling the financial risk model to monitor the financial auditing report, and synchronously pushing the financial auditing report to a manual auditing terminal for verification and approval processing. According to the method, the data utilization rate is increased, and the cost optimization accuracy, the financial management efficiency and the safety are improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an intelligent auxiliary method and system for apartment financial management based on artificial intelligence. Background Technology

[0002] With the rapid development of IoT technology, intelligent management of apartments has become an inevitable trend. As the scale of apartment operations continues to expand and the business models become increasingly complex, the need for refined and intelligent financial management, as a core operational link, is becoming increasingly urgent. However, current apartment financial management faces many pressing technical problems that severely restrict operational efficiency.

[0003] In existing solutions, apartment financial management involves diverse and fragmented data sources, including internal management data such as rent collection, water and electricity consumption, and maintenance expenses; reference data such as pricing and occupancy rates of nearby competing hotels; and publicly available transaction data such as market supply and demand and rent trends. Currently, this data is often stored in different systems or platforms, forming information silos and lacking a unified collection and integration mechanism, resulting in low data utilization. Furthermore, financial data contains a large amount of sensitive information, but traditional storage methods lack robust encryption protection schemes, posing security risks of data leakage and tampering. Synchronization delays and data inconsistencies are also prone to occur during cross-system data transfer. In addition, regarding cost optimization, existing solutions often focus on localized adjustments to single cost items (such as energy consumption and maintenance), lacking a holistic consideration of multi-dimensional costs such as energy consumption, maintenance, and labor costs. This makes it difficult to achieve dynamic optimization of overall costs, resulting in resource waste and revenue loss.

[0004] Therefore, it is urgent to design a technical solution to solve at least one of the above-mentioned technical problems. Summary of the Invention

[0005] The main objective of this application is to provide an intelligent auxiliary method and system for apartment financial management based on artificial intelligence, aiming to solve the technical problems of low data utilization, low accuracy of cost optimization, and lack of a global perspective in related technologies.

[0006] In a first aspect, embodiments of this application provide an intelligent auxiliary method for apartment financial management based on artificial intelligence, comprising: Acquire apartment management data, hotel management data, and publicly available market transaction data to obtain multi-source data; store the multi-source data in a cloud-based financial management center and encrypt it; The multi-source data is processed through a financial analysis and prediction network to output prediction results. This network includes a parallel-running revenue prediction model and a component analysis optimization model. The prediction results include revenue changes and dynamic optimization results for global costs. The revenue prediction model uses an attention mechanism to fuse static features, time-series features, and market sentiment semantic vectors learned from the multi-source data, and combines the fusion results to predict revenue changes. The component analysis optimization model is constructed using a multi-agent reinforcement learning model, and obtains dynamic optimization results for global costs through agent collaborative game theory. Visualized reports and analysis reports are constructed based on the prediction results and dynamically displayed to users; The static and time-series features learned from the multi-source data are input into the financial management model. The model performs preliminary analysis using the pre-set audit rules, generates a financial audit report, and uses the financial risk model to monitor the financial audit report. The financial audit report and monitoring results are then pushed to the manual audit end for verification and approval. The verification and approval results are synchronously stored in the cloud financial management center and encrypted. The verification and approval results are also used to dynamically optimize the audit rules.

[0007] Secondly, embodiments of this application provide an intelligent auxiliary system for apartment financial management based on artificial intelligence, comprising: The data acquisition module is used to obtain apartment management data, hotel management data, and publicly available market transaction data to obtain multi-source data; The prediction module is used to output prediction results from the multi-source data through a financial analysis and prediction network. The financial analysis and prediction network includes a parallel-running revenue prediction model and a component analysis optimization model. The prediction results include revenue changes and dynamic optimization results for global costs. The revenue prediction model uses an attention mechanism to fuse static features, time-series features, and market sentiment semantic vectors learned from the multi-source data, and combines the fusion results to predict revenue changes. The component analysis optimization model is constructed using a multi-agent reinforcement learning model, and obtains dynamic optimization results for global costs through agent collaborative game theory. The display module is used to construct visual reports and analysis reports based on the prediction results and dynamically display them to users; The prediction module is also used to input the static features and time-series features learned from the multi-source data into the financial management model, perform preliminary analysis through the audit rules set in the financial management model, generate a financial audit report, call the financial risk model to monitor the financial audit report, and push the financial audit report and monitoring results to the manual audit end for verification and approval. The storage module is used to store the multi-source data to the cloud financial management center and encrypt it; and to synchronously store the verification and approval processing results to the cloud financial management center and encrypt them. The verification and approval processing results are also used to dynamically optimize the audit rules.

[0008] Thirdly, embodiments of this application also provide a terminal device, the terminal device including a processor and a memory for storing computer programs; the processor is used to execute the computer programs and, when executing the computer programs, implement the intelligent auxiliary method for apartment financial management based on artificial intelligence as described in the first aspect or any embodiment of this application.

[0009] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer software program, which, when executed by a processor, implements the AI-based intelligent auxiliary method for apartment financial management as described in the first aspect or any embodiment of this application.

[0010] This application provides an intelligent auxiliary method and system for apartment financial management based on artificial intelligence. The method acquires apartment management data, hotel management data, and publicly available market transaction data to obtain multi-source data; this multi-source data is then stored in a cloud-based financial management center and encrypted. Next, the multi-source data is processed through a financial analysis and prediction network to output prediction results. This network includes a parallel-running revenue prediction model and a component analysis optimization model. The prediction results include revenue changes and dynamic optimization results for global costs. The revenue prediction model uses an attention mechanism to fuse static features, time-series features, and market sentiment semantic vectors learned from the multi-source data, and combines the fusion results to predict revenue changes. The component analysis optimization model is constructed using a multi-agent reinforcement learning model, and obtains dynamic optimization results for global costs through agent collaborative game theory. Finally, based on the prediction results, visual reports and analysis reports are constructed and dynamically displayed to the user. Next, the static and time-series features learned from the multi-source data are input into the financial management model. Preliminary analysis is performed using pre-set audit rules within the model to generate a financial audit report. The financial risk model is then invoked to monitor the audit report, and the report and monitoring results are simultaneously pushed to a human auditor for verification and approval. Finally, the verification and approval results are synchronously stored and encrypted in a cloud-based financial management center. These results are also used to dynamically optimize audit rules. This embodiment achieves integrated management of multi-source financial data, improves data utilization, enhances cost optimization accuracy from a global perspective, assists in intelligent closed-loop control of financial processes, and improves the efficiency and security of financial management. Furthermore, this embodiment enables accurate prediction and dynamic optimization of financial conditions, further strengthening the accuracy of financial data-driven operational decisions and providing a more refined and intelligent financial solution for apartment operations. Attached Figure Description

[0011] Figure 1 A flowchart illustrating an AI-based intelligent auxiliary method for apartment financial management provided in this application; Figure 2 A schematic diagram of the structure of an intelligent auxiliary system for apartment financial management based on artificial intelligence provided in this application; Figure 3 This application provides a schematic block diagram of the structure of a terminal device. Detailed Implementation

[0012] This application provides an AI-based intelligent auxiliary method and system for apartment financial management. The AI-based intelligent auxiliary method for apartment financial management can be applied to a terminal device, which can be a mobile terminal, such as a mobile phone, virtual reality device, tablet computer, laptop computer, desktop computer, wearable device, etc. The terminal device can be a server connected to a cloud service system or a server cluster. The connection can be implemented through hardware circuitry or a communication module.

[0013] In this embodiment, the deep integration of artificial intelligence technology with apartment financial management achieves integrated management of multi-source financial data, accurate prediction and dynamic optimization of financial status, and intelligent closed-loop control of financial processes. This not only significantly improves the efficiency and security of financial management but also strengthens the support of financial data for operational decisions, providing a refined and intelligent financial solution for apartment operations. Addressing the issue of low data utilization in related technologies, this embodiment first constructs a multi-source data integration mechanism to comprehensively acquire apartment management data, hotel management data, and publicly available market transaction data, breaking the information silo dilemma of traditional financial data. Simultaneously, the integrated multi-source data is stored in a cloud-based financial management center and encrypted, achieving centralized data control while ensuring data security. Furthermore, the revenue prediction model deeply integrates static features, time-series features, and market sentiment semantic vectors from multi-source data through an attention mechanism, fully exploring the value of different types of data. This avoids the problem of one-sided data utilization in traditional methods, transforming data from a scattered and useless state into a core resource supporting financial analysis, significantly improving data utilization. To address the issues of low accuracy and lack of a holistic perspective in cost optimization, this embodiment designs a financial analysis and prediction network comprising a revenue forecasting model and a cost analysis and optimization model. The cost analysis and optimization model employs a multi-agent reinforcement learning model, transforming different cost subsystems such as energy consumption, maintenance, and labor costs into independent agents. Through collaborative game theory among these agents, dynamic optimization of global costs is achieved, overcoming the limitations of traditional methods that only adjust single cost items locally. Simultaneously, the revenue forecasting model, incorporating multi-dimensional features, outputs accurate revenue change data, providing a reliable predictive basis for cost optimization. The model's dynamic learning capability adjusts the results based on market changes and operational realities, further enhancing the accuracy of cost optimization and ensuring that it aligns with actual needs while possessing a holistic perspective. Furthermore, the dynamic display of visual reports makes financial results more intuitive and understandable. The intelligent processing of financial audits and the dynamic optimization of audit rules form a complete closed loop from data input to decision output and process iteration, further amplifying the application value of the technical solution and driving the transformation of apartment financial management from a traditional manual model to an intelligent and efficient one.

[0014] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other. Please refer to... Figure 1 , Figure 1 This is a flowchart illustrating an intelligent auxiliary method for apartment financial management based on artificial intelligence, provided as an embodiment of this application. Figure 1 As shown, this AI-based intelligent assistance method for apartment financial management includes the following steps: Step S101: Obtain apartment management data, hotel management data, and publicly available market transaction data to obtain multi-source data; store the multi-source data in the cloud financial management center and encrypt it.

[0015] For example, the cloud-based financial management center is a cloud-based integrated financial data management and service hub that supports the intelligent financial management of apartments. Its core is a secure data storage and processing platform built on cloud computing technology.

[0016] For example, apartment management data, hotel management data, and publicly available market transaction data can be a collection of structured and unstructured data supporting the entire process of apartment financial analysis. Apartment management data, as core internal data, covers complete records of rent collection (including tenant name, room number, rent amount, payment method, payment status, and overdue status), detailed daily operating expenses (such as water and electricity costs in public areas, elevator and fire safety maintenance costs, and outsourced cleaning and security services), basic property information (unit type, building area, decoration standards, and appliance availability), and tenant performance records (historical payment records, lease term, and lease termination settlement information). Hotel management data focuses on operational reference data from competing properties in the same market segment, including real-time pricing and fluctuation rules for similar room types, monthly and quarterly occupancy rate curves, membership system discount strategies, and pricing standards for additional services (such as breakfast and shuttle service), providing a reference for apartment rent pricing and value-added service design. The publicly available transaction data covers both macro and micro levels. It includes data released by official or authoritative institutions such as regional apartment rental indices, vacancy rate statistics reports, and average prices of second-hand housing transactions. It also covers the latest policy documents from the Ministry of Housing and Urban-Rural Development and tax authorities, industry transaction data released by intermediary associations, and public opinion information on regional rental demand and housing reviews on social media and rental platforms. This multi-dimensional data constitutes a reliable data source for subsequent financial analysis and forecasting. After integration, it is stored in a cloud-based financial management center and its data security is ensured by symmetric encryption algorithms, laying the foundation for subsequent processing.

[0017] This system integrates and stores multi-source financial and related data from apartment management, hotel management, and market transactions, ensuring data security through encryption technology. It provides stable data access services for subsequent financial analysis and forecasting networks and financial management models, supporting model access to data for feature extraction and computation. It synchronously stores data from the entire process, including financial audit reports, verification results, and early warning records, forming a complete financial data chain. It also possesses data synchronization and sharing capabilities, supporting visualization and multi-terminal access, serving as the core data foundation for realizing intelligent financial management of apartments.

[0018] As an optional embodiment, in step S101, after storing the multi-source data in the cloud financial management center and encrypting it, the method further includes: extracting tenant data and publicly available data from the multi-source data using a tenant risk prediction model. The tenant data includes tenant performance records, payment behavior, and job stability information, while the publicly available data from the intermediary association includes a blacklist and historical arrears records. A tenant relationship graph is constructed based on a graph neural network GraphSAGE, with co-tenancy and recommendation relationships as edges and tenant credit features as nodes, to extract the correlation between the tenant data and the publicly available data from the intermediary association. The correlation extraction results are processed by a tenant risk prediction model with an adversarial training module, calculating the tenant arrears probability and the expected arrears amount, and determining the arrears risk level based on the tenant arrears probability and the expected arrears amount, and dynamically displaying the arrears risk level to the user.

[0019] Specifically, following step S101, further risk assessment and monitoring of tenant arrears can be conducted. This is achieved by using a pre-trained tenant risk prediction model to accurately extract tenant data and publicly available data from encrypted multi-source data sources. Tenant data includes various performance records, daily payment behavior characteristics, and information reflecting job stability such as occupation type and years of service during the lease term. Publicly available data from the intermediary association primarily includes publicly disclosed lists of defaulting tenants and historical arrears records, among other risk-related data. Furthermore, a tenant relationship network graph is constructed based on the GraphSAGE algorithm. Co-tenancy and mutual referral relationships among tenants are used as connecting edges in the graph, while credit characteristics such as credit scores, historical performance ratings, and timely payment rates are used as nodes. This relationship graph is used to deeply mine the potential correlations between tenant data and publicly available data from the intermediary association, enabling the extraction of risk factors based on their correlation. The integrated data obtained from the correlation extraction is input into the tenant risk prediction model of the adversarial training module for processing. This adversarial training module can improve the model's robustness to missing data and outliers, ensuring the accuracy of the calculation results. Next, the adversarial training module calculates the probability of default and the expected amount of default for each tenant by learning the characteristics of historical risk cases. Then, it combines the default probability and the expected amount of default with preset risk assessment standards to perform a weighted analysis, and finally determines the corresponding default risk level. Subsequently, the default risk level is dynamically displayed to users through the system's visual interface and real-time message push, so that users can keep abreast of tenants' credit status and prepare for risk response in advance.

[0020] For example, in the above steps, an apartment project with more than 500 units, after completing the encrypted storage of multi-source data, initiates a tenant arrears risk prediction process, constructing a targeted tenant relationship network graph based on the GraphSAGE algorithm. Zhang San and Li Si, roommates in apartment 1502 of building 3, form a co-renting relationship edge. Wang Wu, roommate in apartment 801 of building 12, recommended his friend Zhao Liu to move into apartment 303 of building 7 in the same complex, forming a referral relationship edge between Wang Wu and Zhao Liu. The nodes in the graph carry specific credit characteristics. Zhang San has a credit score of 780, a historical performance rating of A, and a payment on time rate of 95% in the past year; Li Si has a credit score of 650, a historical performance rating of B, and a payment on time rate of 82%; Wang Wu has a credit score of 810, a performance rating of A, and a payment on time rate of 98%; and Zhao Liu has a credit score of 590, a performance rating of C, and a payment on time rate of 70%. Simultaneously, it relates to publicly available data from the intermediary association showing Zhao Liu's 1500 RMB historical arrears in another apartment in 2023. In-depth analysis of this relationship graph reveals that Li Si and Zhao Liu previously shared a rental contact, and both have payment on time rates below the apartment average (88%). There is a potential correlation between Zhao Liu's historical arrears and Li Si's two recent instances of delayed payments. Therefore, the core risk factor combination of co-renting / referral association + low payment on time rate + historical arrears is extracted. After inputting this integrated data into the tenant risk prediction model in the adversarial training module, the model maintained stable operation even when faced with situations where Li Si's professional information was incomplete and Zhao Liu's three small payment records were not fully entered, thanks to the optimization of the model's generalization ability using simulated noise data generated during adversarial training. By learning the characteristics of 120 tenant arrears cases from the past three years (such as payment delay frequency, credit score range, and associated risk status), the model accurately calculated that Zhang San's arrears probability was 3% with an estimated arrears amount of 0 yuan, Li Si's arrears probability was 28% with an estimated arrears amount of 1800 yuan (corresponding to one month's rent), and Zhao Liu's arrears probability was 75% with an estimated arrears amount of 3200 yuan (corresponding to one month's rent plus the difference after deducting the deposit). Based on the apartment's pre-set risk assessment standards (probability of arrears <10% is low risk, 10%-50% is medium risk, and >50% is high risk), and using a weighted formula (risk level score = probability of arrears × 0.6 + estimated arrears amount / monthly rent × 0.4), Zhang San was determined to be low risk, Li Si to be medium risk, and Zhao Liu to be high risk. Subsequently, the apartment management system's visual interface in the tenant management section marked Zhang San with green, Li Si with yellow, and Zhao Liu with red, while simultaneously pushing real-time messages to the operations manager, clearly indicating Zhao Liu's high-risk status and historical arrears background, reminding him to communicate about rent payment in advance. For Li Si, it was suggested to shorten the payment cycle to half a month, helping the operations team accurately grasp the tenants' credit status and develop response strategies in advance.

[0021] Optionally, when using this relationship graph to predict tenant default risk, the core of this prediction is to mine potential risk factors in the individual credit characteristics and relationships of tenants through the graph, and combine it with an intelligent model to achieve accurate prediction. This can be divided into the following steps: First, a tenant relationship graph is constructed based on the GraphSAGE algorithm of the graph neural network. The core credit characteristics of tenants, such as credit scores, historical performance ratings, and timely payment rates, are used as graph nodes, and social connections such as co-tenancy relationships and referral relationships between tenants are used as graph connecting edges. At the same time, external risk data such as the blacklist of defaulters published by the intermediary association and historical default records are associated with the corresponding tenant nodes to form a three-dimensional data network of individual characteristics, relationships, and external risks. Next, leveraging the interconnected topology of this graph, we delve deeper into hidden risk relationships beyond individual tenants. For example, graph link analysis reveals hidden connections such as common contacts and indirect co-tenancy among tenants. Simultaneously, by comparing the credit characteristics of related tenants, we extract combined risk factors such as historical arrears, multiple tenants having below-average timely payment rates, and the presence of members with low performance ratings within the co-tenancy group. This breaks through the limitations of traditional methods relying solely on individual data. Subsequently, the individual credit characteristics, associated risk factors, and publicly available external data extracted from the graph are integrated into a complete input dataset and fed into a tenant risk prediction model incorporating an adversarial training module. The model learns the correspondence between features and risks in historical arrears cases and enhances its robustness to missing data and outliers through adversarial training, accurately calculating the probability of arrears and the estimated amount for each tenant. By combining the preset risk assessment standards with the weighted formula to integrate the probability of arrears with the expected amount of arrears, three risk levels—low, medium, and high—are determined. Finally, the risk level is displayed through a visual interface, and risk alerts and targeted response suggestions are pushed to operations personnel, realizing a closed loop of the entire process from related data mining to accurate risk prediction and implementation.

[0022] As an optional embodiment, in step S101, after storing the multi-source data in the cloud financial management center and encrypting it, the method further includes: obtaining regulatory and policy information issued by the Ministry of Housing and Urban-Rural Development and the tax authorities in real time; constructing a regulatory and policy knowledge graph containing policy information, compliance requirements, and financial operations; marking rent control constraints and tax compliance constraints; adapting the apartment's financial operations to the regulatory and policy knowledge graph; wherein, by learning from existing compliance judgment logic through small samples, determining whether the apartment's financial operations are compatible with the updated regulatory and policy information; determining whether there are compliance risks in the apartment's financial operations based on the adaptation results; generating a compliance risk monitoring report; and dynamically displaying it to the user.

[0023] Specifically, following step S101, web crawling technology combined with official channel data interfaces is used to capture policy documents such as rent control and housing rental registration requirements issued by the Ministry of Housing and Urban-Rural Development, as well as relevant laws and regulations on tax collection and invoice management issued by the tax authorities, in real time, ensuring that the policy information obtained is timely, accurate, and complete. Based on natural language processing technology, this legal and policy information is structured and analyzed to extract key content such as core policy clauses, effective dates, and scope of application. A legal and policy knowledge graph is constructed, encompassing three core dimensions: policy information, compliance requirements, and financial operations. The graph clearly marks rent control constraints such as rent collection caps and deposit ratio limits, as well as tax compliance constraints such as tax declaration deadlines and tax base standards, forming a visualized network of policy compliance operations. The apartment's daily financial operation data, including rent collection standards, deposit amounts, tax declaration details, and invoice issuance records, are matched dimension-by-dimensionally with a constructed legal and policy knowledge graph. Addressing the issue of limited labeled samples for newly released policies, a small-sample learning algorithm is used to transfer existing compliance judgment logic, quickly establishing adaptation rules between new policies and financial operations to determine whether the apartment's current financial operations comply with the latest legal and policy requirements. Based on the matching results, if conflicts or deviations are found between financial operations and constraints in the knowledge graph, a compliance risk is identified. The risk type, relevant policy clauses, and scope of impact are clarified, and a compliance risk monitoring report is generated, including a risk description, violation details, and rectification suggestions. This report is dynamically displayed to financial managers, apartment operations managers, and other users through system dashboards and real-time message alerts, ensuring users are promptly aware of the compliance risk status and can take targeted measures.

[0024] Step S102: Output the prediction results from the multi-source data through the financial analysis and prediction network.

[0025] In this embodiment, the financial analysis and forecasting network includes a revenue forecasting model and a component analysis optimization model running in parallel. Further optionally, the forecast results include revenue changes and global cost dynamic optimization results.

[0026] Optionally, the revenue prediction model uses an attention mechanism to fuse static features, temporal features, and market sentiment semantic vectors learned from the multi-source data, and combines the fusion results to predict revenue changes. The component analysis optimization model is constructed using a multi-agent reinforcement learning model, and obtains a dynamic optimization result for global cost through agent collaborative game.

[0027] As an optional embodiment, in step S102 above, the encrypted multi-source data is first preprocessed to remove redundant noise and complete data standardization, and then input into a financial analysis and prediction network composed of a revenue prediction model and a component analysis optimization model in parallel, so that the revenue change and the global cost dynamic optimization results are output synchronously through the collaborative operation of the two models.

[0028] Specifically, in step S102, firstly, the multi-source data is preprocessed. Multi-dimensional feature data is learned from the multi-source data through an embedding layer, and feature encoding is performed using a hierarchical feature encoding method to obtain static features, temporal features, and market sentiment semantic vectors. The static features, temporal features, and market sentiment semantic vectors are then weighted and fused using an attention mechanism. Next, the weighted fused features are input into an income prediction model, which is an improved Temporal Fusion Transformer model. This model learns from historical transaction data to output income change curves at multiple quantiles, and performs multi-scenario predictions (including conservative, neutral, and optimistic scenarios) on the income change curves, fine-tuning them to obtain the income change situation. Furthermore, cost-related data is extracted from the multi-source data and input into a cost analysis and optimization model. This model employs an improved PPO algorithm and a multi-objective reward function, outputting a global cost dynamic optimization result through agent collaborative game. The cost analysis and optimization model is a multi-agent reinforcement learning (MARL) model, which includes independent agents corresponding to the energy consumption system, maintenance system, and labor cost system.

[0029] For example, the income prediction model is constructed using an improved Temporal Fusion Transformer (TFT). This model first extracts static features, time-series features, and market sentiment semantic vectors from multi-source data. Static features cover fixed attributes such as housing type, area, and supporting facilities. Time-series features include time-related data such as historical rental income, occupancy rate fluctuations, and quarterly cycle changes. The market sentiment semantic vector is transformed from textual information such as rental demand and policy evaluations mined from publicly available market information using natural language processing technology. Then, the three types of features are weighted and fused through an attention mechanism to highlight key factors affecting income. At the same time, the model can accurately distinguish between known dynamic features such as holidays and quarterly cycles and unknown dynamic features such as sudden vacancy. It outputs prediction intervals for the 10th, 50th, and 90th quantiles, supports conservative, neutral, and optimistic multi-scenario predictions, and greatly improves the model's prediction accuracy. Finally, it combines the feature fusion results and multi-scenario predictions to output accurate income change information.

[0030] In practical applications, during the feature processing stage, static features not only cover basic attributes such as property type, building area, decoration standards, and appliances (e.g., whether it is furnished or has a private bathroom), but also incorporate key fixed factors affecting rental pricing, such as the property's floor level, orientation, surrounding transportation hubs (e.g., distance to subway stations), and amenities (e.g., distribution of supermarkets and schools). These features are transformed into standardized vectors through an embedding layer, forming stable basic feature support. Temporal features, on the other hand, use daily, weekly, and monthly granularities to extract historical rent collection data from the past 1-3 years, monthly occupancy rate fluctuation curves, quarterly rental cycle changes, and peak passenger flow during holidays. Simultaneously, time coding technology preserves the periodic patterns of different time periods (e.g., peak rental season after the Spring Festival, peak demand during graduation season). The extraction of market sentiment semantic vectors employs a pre-trained BERT model. This model deeply analyzes unstructured texts such as rental platform reviews, social media rental topics, government-issued rental policy documents, and regional planning announcements to uncover positive semantics like expectations of rising rents, housing shortages, and policy encouragement for renting, or negative semantics like inadequate infrastructure, inconvenient transportation, and policy restrictions on rents. This transforms the textual information into a computable high-dimensional vector. Based on this, the model dynamically weights and fuses the three types of features through a multi-head attention mechanism. Feature weights are automatically adjusted for different scenarios. For example, during peak rental seasons like graduation season, the weights of occupancy rate time-series features and rental demand sentiment vectors are increased to 0.35 and 0.3, respectively. During policy adjustment periods (such as the introduction of rent control policies), the weight of the policy sentiment semantic vector is increased to 0.4, ensuring that core influencing factors are given priority. Meanwhile, the improved TFT, through a newly added dynamic feature differentiation module, accurately distinguishes between known dynamic features such as holidays and quarterly cycles, and unknown dynamic features such as sudden vacancy (e.g., reduced customer traffic due to nearby construction sites) and a resurgence of the epidemic. For known features, historical patterns are used for fitting; for unknown features, the model's gating mechanism dynamically adapts, effectively reducing prediction bias caused by unforeseen factors. The final output is revenue change curves at the 10th, 50th, and 90th quantiles. The 10th quantile corresponds to a conservative scenario (e.g., the lowest expected revenue under the combined effects of off-season and unforeseen risks), the 50th quantile to a neutral scenario (the most likely revenue based on historical patterns), and the 90th quantile to an optimistic scenario (e.g., the highest expected revenue under peak season and favorable policies). Compared to the traditional LSTM model, it reduces prediction errors for revenue fluctuations, especially when dealing with unconventional situations such as sudden vacancy and policy adjustments, where the improvement in prediction accuracy is more significant. Finally, combining the fusion results with multi-scenario predictions, the output provides accurate revenue changes covering monthly, quarterly, and annual periods, including peak revenue periods, troughs, year-on-year and month-on-month growth rates, and analysis of key influencing factors.

[0031] For example, the component analysis optimization model is built on multi-agent reinforcement learning (MARL). Its core is to transform the global cost optimization of the apartment into a sequential decision problem. First, it breaks down the cost into subsystems such as energy consumption, maintenance, and labor costs. Each subsystem acts as an independent agent and is responsible for the cost optimization decision of its corresponding dimension. The model adopts an improved PPO (proximal policy optimization) algorithm and designs a multi-objective reward function that includes energy consumption reduction rate, equipment life extension, and service satisfaction. Through the collaborative game between the agents, the relationship between the cost reduction of a single subsystem and the overall operational efficiency is balanced, avoiding global cost loss caused by local optimization. Finally, it outputs a dynamic global cost optimization result that takes into account the needs of each subsystem and the overall benefits.

[0032] Understandably, during the model building phase, the overall cost of the apartment is first broken down into three core subsystems: energy consumption, maintenance, and labor costs. Each subsystem corresponds to an independent intelligent agent, with clearly defined optimization objectives and decision boundaries. The energy consumption agent is responsible for water and electricity consumption control, with decisions including the duration of public area lighting, air conditioning temperature thresholds, and smart meter monitoring frequency. The maintenance agent is responsible for facility and equipment maintenance management, with decisions covering maintenance cycles, repair priorities, and spare parts inventory planning for equipment such as elevators, fire protection systems, and home appliances. The labor cost agent is responsible for optimizing the operations team's scheduling, with decisions including the number of cleaning, security, and customer service personnel on duty, shift patterns, and the allocation of part-time staff. The model employs an improved PPO (Proximity-Based Proximal Policy) optimization algorithm, adding a constraint mechanism to the traditional PPO algorithm to prevent agents from sacrificing overall interests for a single objective. For example, an energy agent excessively reducing air conditioning operating time might trigger tenant complaints. A multi-objective reward function is also designed, balancing different optimization directions through weight allocation, covering core dimensions such as energy consumption reduction rate, equipment lifespan extension, and service satisfaction. The agent's decision-making reward is directly linked to the score of this function. In the collaborative game process, each agent dynamically adjusts its decisions through environmental interaction and information sharing. For example, the maintenance agent plans to extend the elevator maintenance cycle to reduce labor and spare parts costs. However, if this decision may lead to an increase in the frequency of elevator malfunctions, thereby increasing energy consumption (i.e., high energy consumption of backup equipment during malfunctions) and tenant complaints (i.e., decreased service satisfaction), the energy consumption agent and the labor cost agent will raise objections through a feedback mechanism. The three parties adjust their decisions through multiple rounds of game theory, ultimately determining a maintenance plan that balances cost and safety. For example, optimizing the maintenance process to reduce the cost per service, finding a balance between local cost reduction and overall efficiency improvement. Through continuous iterative learning, the model can dynamically adapt to real-time changes in apartment operations. For example, during peak seasons when there are more tenants, the labor agent increases shifts, the energy agent appropriately improves the energy consumption guarantee experience in public areas, and the maintenance agent conducts facility inspections in advance. The final output of the global cost dynamic optimization results not only includes the specific cost reduction targets of each subsystem, but also comes with detailed implementation plans, such as setting reasonable ranges for air conditioning temperature in summer, arranging elevator maintenance during off-peak hours of tenant travel at night, and adopting a peak-hour increase and off-peak-hour shift mode for cleaning staff, to ensure that the optimization results can be directly implemented, while avoiding the chain of negative impacts caused by local optimization.

[0033] Optionally, the income forecasting model further includes a market mutation detection module. Based on this, in step S102, after the financial forecasting network outputs the forecast results, the market mutation detection module can be used to calculate the differences in attention to various features and automatically identify major event nodes. These major event nodes include income mutation event nodes caused by policy adjustments or regional planning. Based on these major event nodes, potential mutation trends in income changes are analyzed, and income risk warning information is generated by combining the scope and degree of the mutation's impact. The income risk warning information is pushed to relevant personnel via system messages and / or SMS, and the income risk warning information is stored in the cloud-based financial management center and encrypted.

[0034] It can be explained that the market mutation detection module included in the income prediction model works in synergy with the main model. Based on this, after the financial prediction network outputs the income change and global cost dynamic optimization results in step S102, this detection module automatically starts its workflow. First, it retrieves the attention weight data of each feature during the income prediction model's calculation process. By calculating the difference in feature attention within different time windows, it locates the key features of weight mutations. For example, when the attention weight of the policy-related public opinion semantic vector spikes from 0.15 to 0.5 within 24 hours, or when the weight of the regional planning text feature suddenly exceeds three times the historical average, the major event node identification mechanism is triggered. Further verification is performed by combining the data sources associated with the features, automatically distinguishing different types of mutation triggers such as policy adjustments, regional planning, and major public events. Policy adjustments may include new rent control regulations issued by the Ministry of Housing and Urban-Rural Development and rental tax adjustment policies issued by the tax authorities. Regional planning includes events that directly affect apartment rent levels, such as the implementation of new subway line plans and announcements of surrounding business district construction. Finally, it accurately marks the specific event nodes that may cause income mutations. The module then correlates major event nodes with the output revenue change curves, combining the event's effective date and scope of application to predict potential abrupt changes in revenue. For example, if a planning announcement for a new industrial park in the vicinity is detected, by comparing historical data from similar areas, it is determined that the event will drive an increase in apartment occupancy rates within 6 months, thereby boosting revenue growth by 15% to 20%. If a rent cap policy is detected, it predicts that future quarterly revenue growth will narrow from 8% to less than 3%. Simultaneously, the module quantitatively analyzes the scope of the impact of the abrupt change, clarifying whether it affects only a specific building or covers the entire apartment project, and whether the impact is a short-term fluctuation or a long-term trend. Based on this, structured revenue risk warning information is generated, including the event name, occurrence time, impact characteristics, predicted magnitude of the change, and risk level. After the warning information is generated, the system will push it through two channels. A red or yellow warning pop-up will appear on the apartment management system's operations dashboard, and a text message containing the core warning content will be sent to the operations manager, finance manager, and other relevant personnel to ensure rapid information delivery. Finally, the income risk warning information will be uploaded to the cloud-based financial management center and stored using the same encryption algorithm as the multi-source data. This ensures the integrity of the warning record and provides sample data containing the impact of sudden events for subsequent model iteration and optimization.

[0035] As an optional embodiment, after step S102 outputs the prediction results from the multi-source data through the financial analysis and prediction network, new transaction data and market dynamic feedback data generated in real time can also be collected to form a supplementary data set with the multi-source data; the FTRL algorithm is used to fine-tune the revenue prediction model and cost analysis optimization model in the financial analysis and prediction network online, and the fine-tuned model parameters are synchronously updated to the financial analysis and prediction network; a model performance monitoring threshold is set for the financial analysis and prediction network. When the prediction error of revenue changes or global cost dynamic optimization results exceeds the preset threshold three times consecutively, incremental training of the model is triggered to further enhance the model's adaptability to the supplementary data set and effectively suppress model drift.

[0036] Specifically, the implementation of model optimization and iteration in step S102 is as follows: First, a real-time data acquisition mechanism is constructed. New transaction data, such as newly generated rent collection records, water and electricity consumption payment data, and maintenance service settlement vouchers, are captured in real time through the transaction interface of the apartment management system. Simultaneously, relying on web crawlers and official data push channels, market feedback data, including the regional apartment rent index, the latest statistics on rental market vacancy rates, policy updates from housing and construction and tax departments, and social media rental sentiment dynamics, are collected concurrently. This real-time data is integrated with the initial multi-source data to form a supplementary data set containing historical accumulation and real-time dynamics, ensuring the timeliness and comprehensiveness of data input. Next, the FTRL (Follow The Regularized Leader) algorithm is used to fine-tune the financial analysis and prediction network online. For the income prediction model, based on the newly added rent fluctuation characteristics and market sentiment changes in the supplementary data set, the algorithm dynamically adjusts the attention weights and gating unit parameters of each layer of the improved TemporalFusion Transformer model, especially strengthening the ability to capture features of recent sudden market factors. For the cost analysis optimization model, the algorithm incorporates actual expenditure deviations in energy consumption, maintenance, and labor costs from new transaction data to optimize the reward function weights of the multi-agent reinforcement learning model and the policy update step size of the improved PPO algorithm. This makes the decisions of each cost subsystem agent more aligned with current operational realities. After fine-tuning, the system instantly pushes the updated model parameters to the runtime of the financial analysis and prediction network via a parameter synchronization channel, achieving seamless model performance upgrades. Simultaneously, the system presets model performance monitoring thresholds for the financial analysis and prediction network, setting the prediction error threshold for revenue changes at 5% and the prediction error threshold for global cost dynamic optimization results at 8%. By comparing the model's output prediction results with the actual data in the supplementary dataset in real time, the system calculates and records the error value for each prediction. When the prediction error of revenue changes exceeds 5% for three consecutive times, or the prediction error of the global cost dynamic optimization result exceeds 8% for three consecutive times, the system automatically triggers the incremental training mechanism of the model. The system uses the supplementary data set as the core training sample and combines historical multi-source data to construct an incremental training dataset. It adopts a training strategy of freezing the base layer and fine-tuning the upper-layer network. While retaining the original learning results of the model, it focuses on strengthening the model's ability to adapt to the characteristics of new data distribution. By increasing the number of training iterations and adjusting the learning rate, it further corrects the deviation of model parameters, effectively suppresses the model drift problem caused by changes in data distribution, and ensures that the financial analysis and prediction network maintains stable and accurate prediction performance in the long term.

[0037] Step S103: Based on the prediction results, construct a visualization report and analysis report, and dynamically display them to the user.

[0038] As an optional embodiment, in step S103, firstly, a dynamic dashboard is constructed using ECharts. This dashboard generates a visual report by overlaying predicted curves with actual data, displaying revenue changes and the results of dynamic global cost optimization. A drill-down function is then configured for this visual report. Further, the drill-down function allows for layer-by-layer viewing of revenue breakdown details and cost breakdowns across multiple projects, single buildings, and individual apartments, meeting the need for refined data viewing. Next, based on the customized needs of the owner or financial personnel, a target time period and analysis dimension are selected, and predicted data for the corresponding dimension is extracted from the revenue changes and the results of dynamic global cost optimization. Then, the predicted data and visual report for the corresponding dimension are parsed using a large language model, generating a structured analysis and prediction report that includes data trend charts, a summary of key financial indicators, and an explanation of the causes of discrepancies. Finally, the visual report and structured analysis and prediction report are dynamically displayed to the user, supporting online viewing and export. The visual report and structured analysis and prediction report are also synchronously stored in the cloud-based financial management center and encrypted to ensure data security.

[0039] Specifically, the ECharts data visualization tool is first used to build a dynamic dashboard adapted to the apartment financial management scenario. This dashboard visualizes the revenue changes output by the financial analysis and forecasting network, as well as the dynamic optimization results of global costs. Intuitive visual reports are generated by overlaying revenue forecast curves with historical actual revenue data and cost optimization target curves with current actual cost data, clearly showing the difference between forecasts and actual results. A flexible drill-down function is also configured for this visualization report. This drill-down function allows users to click on summary data in the visualization report and drill down layer by layer, from the overall financial overview of multiple apartment projects to the revenue and cost data of individual buildings, and further down to the detailed revenue composition of individual apartments, such as rental income and value-added service income, as well as cost breakdown data such as water and electricity consumption, maintenance parts, and personnel services. This comprehensively meets the needs of refined data verification and analysis during operation. The system then receives customized requests from property owners or financial personnel, selecting target time periods (monthly, quarterly, or annually) and specific analytical dimensions (such as revenue growth, cost control, or return on investment) based on their core concerns. It accurately extracts corresponding predictive data and historical comparative data from revenue changes and overall cost dynamic optimization results. Next, it uses a large language model to deeply analyze the extracted predictive data and core information in the visualized reports. Combining this with professional logic for apartment financial operations, it automatically generates a structured analysis and prediction report. This report includes customized data trend charts tailored to specific needs, summaries of key financial indicators such as rental income growth rate, cost reduction rate, and net profit expectations, as well as explanations of discrepancies between predictions and historical data, such as changes in the market environment, the impact of policy adjustments, and the effectiveness of operational strategy optimization. Finally, the system dynamically displays the generated visualized reports and structured analysis and prediction reports to users through multiple channels, including the apartment management system's web and mobile applications. Users can zoom in and out to view detailed data and switch between different display dimensions online, and the system also offers export options in PDF and Excel formats. At the same time, the visualized reports and structured analysis and forecast reports will be uploaded to the cloud-based financial management center and encrypted using the same encryption algorithm as the multi-source data. This ensures the data security of the financial analysis results and provides a reliable basis for subsequent data traceability and review.

[0040] Step S104: Input the static features and time-series features learned from the multi-source data into the financial management model, perform preliminary analysis through the audit rules preset in the financial management model, generate a financial audit report, call the financial risk model to monitor the financial audit report, and push the financial audit report and monitoring results to the manual audit end for verification and approval.

[0041] In practical applications, real-time verification is further conducted on various types of financial data that have been identified and classified. The focus is on checking the completeness, logical consistency, and numerical reasonableness of the data to ensure the accuracy of every financial record. After verification, the results are immediately fed back to the data entry personnel, clearly indicating the possible error locations, specific types, and targeted correction suggestions to facilitate rapid data correction. Furthermore, different time periods, including monthly, quarterly, and annual periods, and different analytical dimensions, such as revenue composition, cost control, and return on investment, can be flexibly selected according to the personalized needs of owners or financial personnel. This allows for the precise extraction of corresponding data from revenue changes and overall cost dynamic optimization results, combined with visual charts and in-depth analysis to generate customized analysis and prediction reports, supporting online viewing and export in multiple formats. In the automated expense reimbursement review process, the rule engine and intelligent model work together to automatically mark questionable reimbursement applications and submit them directly to the human review end. Human reviewers, based on the marked points of doubt, complete the review and then feed the final approval result back to the system. This feedback data is used to dynamically optimize the review rules and model parameters, continuously improving the accuracy of automated review.

[0042] As an optional embodiment, in step S104, the static features and time-series features learned from the multi-source data are input into the financial management model, and a preliminary analysis is performed using the pre-set audit rules in the financial management model to generate a financial audit report, including: According to the apartment's financial regulations, expense reimbursement review rules are configured in the financial management model. These rules include at least the following dimensions: reimbursement amount, voucher requirements, and approval process. The financial management model employs a hybrid algorithm architecture combining a rule engine and a Neural Gradient Boosting Tree (NGBoost) model. Static and time-series features learned from the multi-source data are input into the financial management model. Combined with the expense reimbursement review rules, the model undergoes joint analysis using both the rule engine and the NGBoost model to obtain automatic review results. Financial statements are generated and preliminarily analyzed according to accounting standards and apartment management requirements. These statements indicate at least one core financial indicator among revenue, cost, and profit. The preliminary analysis includes calculating year-on-year and month-on-month changes in these indicators, identifying abnormal fluctuations, and preliminarily determining their causes. The automatic review results are integrated with the preliminary analysis conclusions of the financial statements to generate a financial audit report containing audit opinions, indicator details, anomaly explanations, and feature contribution explanations. The questionable data, compliance probability distribution data, and feature contribution data generated during the automatic review process are synchronized to the cloud-based financial management center.

[0043] Specifically, in step S104, the expense reimbursement review rules are first precisely configured in the financial management model based on the apartment's internal financial management system and industry standards. These rules comprehensively cover core dimensions such as reimbursement amount, voucher requirements, and approval process. The reimbursement amount specifies the single reimbursement limit and annual cumulative reimbursement limit for different positions and different expense types. The voucher requirements define the types, issuance time, and information completeness standards of legal and valid reimbursement vouchers. The approval process specifies the approval level and transfer nodes corresponding to different reimbursement amounts. Furthermore, static features learned from multi-source data, such as basic property attributes, departmental functional attributes, and inherent characteristics of expense types, as well as time-series features, such as historical reimbursement frequency, monthly reimbursement amount fluctuations, and expense patterns in different time periods, are input into the financial management model. The financial management model, combined with pre-defined expense reimbursement review rules, uses a rule engine to quickly verify whether reimbursement applications meet limits, vouchers are compliant, and approval processes are complete. Simultaneously, the NGBoost model calculates the feature similarity between reimbursement data and historical compliant cases, analyzing the rationality and probability of anomalies in reimbursement behavior. Both jointly output automatic review results, including three categories: approved, rejected, and questionable. Subsequently, the NGBoost model automatically generates financial statements covering core financial indicators such as revenue, cost, and profit, according to national accounting standards and the apartment's internal management needs. It also performs preliminary analysis, calculating year-on-year and month-on-month changes for each core indicator. By setting reasonable thresholds, it identifies abnormal fluctuations in indicators and, combined with related feature data, preliminarily determines the causes of these fluctuations, such as whether abnormal revenue growth stems from increased occupancy rates or whether abnormal cost increases are due to increased maintenance frequency. The system integrates the results of automated audits with preliminary analysis of financial statements to generate a comprehensive financial audit report. This report includes clear audit opinions, detailed breakdowns of core indicators, explanations of abnormal fluctuations, and explanations of the contribution of each feature to the audit results and indicator changes, ensuring the audit basis is traceable. Finally, questionable data marked during the automated audit process, probability distribution data reflecting the compliance of reimbursement applications, and contribution data of various features to the audit results are simultaneously uploaded to the cloud-based financial management center, providing support for subsequent risk monitoring, model optimization, and data traceability.

[0044] Optionally, in the above steps, the static and time-series features learned from the multi-source data are input into the financial management model, and combined with the expense reimbursement review rules, a rule engine and an NGBoost model are used for joint analysis to obtain automatic review results. This includes: using the rule engine to perform rigid review and screening of non-compliant applications based on the expense reimbursement review rules using the static and time-series features learned from the multi-source data; using the NGBoost model to perform fine-grained judgment on the remaining applications other than non-compliant applications, and learning the implicit features of questionable applications among the remaining applications, outputting the compliance probability distribution and feature contribution; automatically marking questionable applications whose compliance probability distribution lower limit is lower than a preset threshold, marking the questionable points according to the feature contribution, and obtaining automatic review results.

[0045] The specific implementation method for obtaining automatic review results through joint analysis of the rule engine and NGBoost model in the above steps is as follows: First, the static features and time-series features learned from multi-source data are completely input into the rule engine module of the financial management model. The rule engine performs rigid review on each reimbursement application according to the preset expense reimbursement review rules, and verifies one by one whether the application meets the reimbursement limit, whether the voucher meets the requirements of information completeness and legality, and whether the approval process flows according to the prescribed levels. It directly filters out applications that exceed the limit, applications with missing vouchers or incorrect information, applications that have not completed the corresponding approval process, and marks the reasons for the violation. For the remaining applications, excluding those approved by the rules engine for violations, the NGBoost model performs fine-grained compliance judgments. This model learns the characteristic patterns of a large number of compliant and questionable applications in historical reimbursement data to deeply mine hidden abnormal features in the remaining applications. These include hidden issues such as the same applicant submitting similar expense reimbursements at a significantly higher frequency than the historical average within a short period, reimbursement amounts approaching the limit without reasonable business support, and mismatches between expense payment time and business operation time. The model also outputs the compliance probability distribution for each application and the contribution of each feature to the compliance judgment. For example, if the compliance probability distribution of a certain application is between 30% and 55%, the contribution of the abnormal reimbursement frequency feature reaches 0.6, becoming the core factor affecting the compliance judgment. The system presets a threshold for determining the lower limit of the compliance probability distribution. When the lower limit of the compliance probability distribution of an application is lower than this threshold, it is automatically marked as a questionable application. Based on the feature contribution output by the NGBoost model, the specific questionable points that lead to the application being questionable are accurately marked, such as "the monthly reimbursement amount for similar expenses increased by 40% month-on-month, without corresponding business vouchers to support it" or "the applicant's historical average compliance probability is 85%, but the compliance probability this time is only 40%, which is abnormal". Finally, the system integrates the results of non-compliant applications screened by the rule engine with the results of questionable and compliant applications marked by the NGBoost model to form an automatic review result that includes three categories: approved, non-compliant, and questionable, along with corresponding explanations.

[0046] Alternatively, when learning the implicit characteristics of questionable applications among the remaining applications, the implicit characteristics of questionable applications may include, for example, high-frequency small-amount split reimbursement, abnormal cross-department reimbursement association, and ambiguity of key information on vouchers.

[0047] In the above steps, the financial management model adopts a hybrid algorithm architecture that combines a rule engine and an NGBoost model. The rule engine is responsible for executing clear and rigid audit standards, while the NGBoost model relies on its powerful nonlinear fitting capabilities to handle flexible judgment requirements in complex scenarios.

[0048] The financial management model employs a hybrid algorithm architecture combining a rule engine and Neural Gradient Boosting Tree (NGBoost). This two-layer collaborative review mechanism achieves the dual goals of rigid standard implementation and flexible scenario adaptation. The rule engine, acting as the first layer of review, transforms the clearly quantifiable and well-defined rigid review standards of the apartment's financial regulations into machine-executable logical rules. These rules are solidified in the system in the form of decision tables and if-else logic chains, covering core scenarios such as hierarchical restrictions on reimbursement amounts (e.g., the maximum daily expense reimbursement limit for ordinary employees, the reimbursement amount threshold within the department head's approval authority, and tiered rules requiring general manager approval for amounts exceeding a certain limit), stringent requirements for voucher compliance (e.g., invoices must include a unified social credit code, the invoice date must be within the reimbursement period, the consumption details must be consistent with the business purpose, and electronic invoices must be verified), and standardized approval process nodes (e.g., expense reimbursements must be confirmed by the immediate supervisor, reviewed by the department head, and verified by the financial specialist in a fixed sequence; the absence of any node constitutes a process violation). The rule engine quickly verifies whether the application data conforms to preset rules by batch traversing the static and temporal features corresponding to each expense reimbursement application. For example, when the amount of an expense reimbursement application exceeds the single reimbursement limit for the applicant's position, or the voucher lacks the necessary verification mark, or the approval process has not completed the direct supervisor confirmation step, the rule engine will immediately mark it as an illegal application and record the specific violation, achieving efficient preliminary screening, preventing obviously non-compliant applications from entering the subsequent process, and greatly improving the review efficiency.

[0049] As the core of the second-layer fine-grained review, the NGBoost model, relying on its gradient boosting framework and probabilistic prediction capabilities, as well as its powerful nonlinear fitting ability, is specifically designed to handle complex and flexible scenarios that rule engines cannot cover. These scenarios lack clear quantitative standards and require comprehensive judgment based on historical data and related features. The model first learns from a large amount of data on compliant reimbursement cases, questionable cases, and hidden violations from the past few years to uncover nonlinear relationships between different features. Examples include the matching relationship between the frequency of reimbursements and business volume of the same applicant at different times, the normal fluctuation range of similar expenses in different seasons and business cycles, and the structural differences in cross-departmental expense expenditures, thus constructing a multi-dimensional feature mapping model. When a compliant application selected by the rule engine enters the NGBoost model, the model inputs the static features (such as the applicant's position, expense type, and department), time-series features (such as the reimbursement frequency in the past 6 months, the average monthly reimbursement amount, and historical fluctuations of similar expenses), and related features (such as the progress stage of the corresponding business project and the department's business volume data during the same period). Through multiple rounds of gradient boosting calculations, the model outputs the compliance probability distribution of the application and the contribution of each feature to the compliance judgment. For example, an employee's expense reimbursement amount did not exceed the limit set by the rules engine, and the vouchers and approval process were fully compliant. However, the NGBoost model's analysis revealed that an employee's frequency of similar expense reimbursements over the past three months had increased threefold compared to the historical average, and the reimbursement times were concentrated in non-peak business periods. Furthermore, there were no new expenditure needs for the corresponding business projects. Therefore, the model judged the employee's compliance probability to be low and marked it as questionable. The characteristics of "abnormally high reimbursement frequency" and "mismatch between expense expenditures and business cycles" contributed over 70% of this, providing clear evidence for the review.

[0050] Thus, the two models work together to form a complete review process that combines rigid screening with flexible verification. The rule engine is responsible for quickly filtering obviously non-compliant applications, ensuring the seriousness and efficiency of the review. The NGBoost model focuses on identifying hidden risks in complex scenarios, compensating for the limitations of rigid rules and ensuring the comprehensiveness and accuracy of the review. At the same time, the compliance probability distribution and feature contribution output by the model can provide clear judgment criteria for subsequent manual review, making the review process both based on clear standards and flexible in responding to various complex business scenarios, effectively balancing review efficiency and review quality.

[0051] As an optional embodiment, the financial risk model adopts a fusion mechanism of rule verification and self-supervised learning anomaly detection. Rule verification is based on preset data verification specifications, while the self-supervised learning model learns the distribution characteristics of normal financial data in an unsupervised manner, adapting to scenarios where there are few abnormal samples and many normal samples in financial data. Based on this, in step S104, when the financial risk model is called to monitor the financial audit report, multi-dimensional real-time verification is performed on various types of financial data that have been identified and classified in the report. Data integrity verification focuses on checking whether necessary fields such as reimbursement vouchers, transaction records, and account correspondence are missing, ensuring that the core information of each financial record is complete and traceable. Logical consistency verification focuses on comparing whether the reimbursement amount matches the corresponding account amount in the financial statements, whether the time and amount of the payment record match the reimbursement reason, and whether the reconciliation relationship between cross-report data is established, avoiding data logic contradictions. Numerical reasonableness verification leverages the anomaly detection capabilities of the Contrastive Learning model to identify outliers that deviate from the normal data distribution. Examples include single expense claims exceeding the cluster center of similar scenarios by three standard deviations, or unfounded abrupt changes in monthly transactions for the same financial item. After verification, the system provides all results to data entry personnel in a structured form. The form clearly indicates the specific location and type of error, including missing data, logical contradictions, and numerical anomalies, along with the confidence level and targeted correction suggestions, such as supplementing property fee payment serial numbers for a specific time period or verifying the consistency between rental income and the amount stipulated in the lease agreement. For financial audit reports that pass verification, the model generates risk monitoring conclusions based on the verification results, classifying them as risk-free, low-risk, or high-risk. High-risk reports, such as those with significant data inconsistencies, suspected fraudulent expense claims, or abrupt changes in item amounts, will be additionally marked with specific risk levels and core risk points and simultaneously pushed to higher-level audit nodes for focused review. Subsequently, the system pushes the verification results, risk monitoring conclusions, and financial audit report to the human review end, allowing reviewers to conduct efficient verification and approval based on the machine verification results. Simultaneously, the verification data, risk monitoring data, and subsequent human review feedback data generated during this verification process are uploaded to the cloud-based financial management center and stored using a unified encryption method to ensure data security and traceability. The verification and approval results from the human review end are then fed back to the financial risk model to dynamically update the normal financial data distribution characteristics learned by the Contrastive Learning model. This also optimizes relevant parameters and standards in the rule-based verification specifications, continuously improving the model's adaptability to real-world financial scenarios and the accuracy of data verification.

[0052] Furthermore, the financial risk level is determined by comprehensively calculating the probability of occurrence and the degree of impact of the risk. Differentiated early warning methods are set for different risk levels, including system message pop-ups, SMS reminders, email notifications, and corresponding processing procedures. Low-risk matters are handled according to the regular process, while medium- and high-risk matters are handled through enhanced processes such as multi-level review and higher-level approval to ensure that all types of risks are dealt with in a timely and effective manner.

[0053] Step S105: The verification and approval processing results are synchronously stored in the cloud financial management center and encrypted. The verification and approval processing results are also used to dynamically optimize the audit rules.

[0054] As an optional embodiment, in step S105, after the verification and approval process of the financial audit report and risk monitoring results is completed at the manual review end, the system will automatically collect and integrate the complete verification and approval process results. These results cover core content such as the final confirmation information of approved applications, detailed reasons for rejection and the basis for determining violations in rejected applications, review conclusions and risk level adjustment opinions for questionable applications, manually corrected data errors, and supplementary explanations. Subsequently, these processing results are synchronously uploaded to the cloud-based financial management center through a secure data transmission channel. They are encrypted and stored using the same encryption algorithm as multi-source data and financial audit reports, ensuring both the integrity and security of the verification and approval records and ensuring the traceability of the data link in the entire financial processing flow. Simultaneously, the verification and approval process results will be used as key feedback data in the dynamic optimization mechanism of audit rules. The system will automatically extract the manual review decision-making logic and adjustment opinions, such as analyzing the oversights of the original audit rules for cases of machine misjudgment corrected manually, identifying rule blind spots based on the common characteristics of frequently rejected applications, and identifying unreasonable aspects of the risk judgment standards in the rules based on the risk level adjustment after review. These extracted optimization requirements are comprehensively analyzed along with historical verification and approval data and model calculation data. This allows for dynamic updates to the expense reimbursement review rules in the financial management model, the logical judgment conditions of the rule engine, the feature weight parameters of the NGBoost model, and the verification specifications of the financial risk model. For example, this includes adjusting reimbursement thresholds for different positions, supplementing verification standards for new compliant vouchers, optimizing the judgment parameters for identifying abnormal fluctuations, and improving the collaborative judgment logic between the rule engine and the model. Through continuous iteration, the review rules are constantly adapted to the actual changes in the apartment's financial operations, gradually improving the accuracy and efficiency of automated review, reducing unnecessary manual intervention, and ensuring that the review rules always comply with the apartment's financial regulations and industry standards.

[0055] In this embodiment, not only is integrated management of multi-source financial data realized, assisting in the intelligent closed-loop control of financial processes and improving the efficiency and security of financial management, but it also enables accurate prediction and dynamic optimization of financial conditions, further enhancing the accuracy of financial data-driven operational decisions and providing a more refined and intelligent financial solution for apartment operations.

[0056] Please see Figure 2 , Figure 2 An AI-based intelligent auxiliary system for apartment financial management is provided for embodiments of this application. The AI-based intelligent auxiliary system for apartment financial management includes: a data acquisition module, used to acquire apartment management data, hotel management data, and publicly available market transaction data to obtain multi-source data; The prediction module is used to output prediction results from the multi-source data through a financial analysis and prediction network. The financial analysis and prediction network includes a parallel-running revenue prediction model and a component analysis optimization model. The prediction results include revenue changes and dynamic optimization results for global costs. The revenue prediction model uses an attention mechanism to fuse static features, time-series features, and market sentiment semantic vectors learned from the multi-source data, and combines the fusion results to predict revenue changes. The component analysis optimization model is constructed using a multi-agent reinforcement learning model, and obtains dynamic optimization results for global costs through agent collaborative game theory. The display module is used to construct visual reports and analysis reports based on the prediction results and dynamically display them to users; The prediction module is also used to input the static features and time-series features learned from the multi-source data into the financial management model, perform preliminary analysis through the audit rules set in the financial management model, generate a financial audit report, call the financial risk model to monitor the financial audit report, and push the financial audit report and monitoring results to the manual audit end for verification and approval. The storage module is used to store the multi-source data to the cloud financial management center and encrypt it; and to synchronously store the verification and approval processing results to the cloud financial management center and encrypt them. The verification and approval processing results are also used to dynamically optimize the audit rules.

[0057] In some implementations, the AI-based intelligent auxiliary system for apartment financial management can be applied to terminal devices. It should be noted that, for the sake of convenience and brevity, the specific working process of the AI-based intelligent auxiliary system for apartment financial management described above can be referred to the corresponding process in the aforementioned embodiments of the AI-based intelligent auxiliary method for apartment financial management, and will not be repeated here.

[0058] Please see Figure 3 , Figure 3 This is a schematic block diagram of the structure of a terminal device provided in an embodiment of this application.

[0059] like Figure 3 As shown, the terminal device 300 includes a processor 301 and a memory 302, which are connected via a bus 303, such as an I2C bus. Specifically, the processor 301 provides computing and control capabilities to support the operation of the entire terminal device. The processor 301 can be a central processing unit, or it can be 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. Those skilled in the art will understand that... Figure 3 The structures shown are merely block diagrams of some structures related to the embodiments of this application and do not constitute a limitation on the terminal devices on which the embodiments of this application are applied. Specific servers may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. The processor is used to run a computer program stored in the memory, and when executing the computer program, implements any of the AI-based intelligent auxiliary methods for apartment financial management provided in the embodiments of this application. It should be noted that those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the terminal device described above can be referred to the foregoing embodiments of the AI-based intelligent auxiliary method for apartment financial management, and will not be repeated here.

[0060] This application also provides a storage medium for computer-readable storage, wherein the storage medium stores one or more programs that can be executed by one or more processors to implement the steps of any of the AI-based intelligent auxiliary methods for apartment financial management provided in the specification of this application.

Claims

1. An intelligent auxiliary method for apartment financial management based on artificial intelligence, characterized in that, The method includes: Acquire apartment management data, hotel management data, and publicly available market transaction data to obtain multi-source data; store the multi-source data in a cloud-based financial management center and encrypt it; The multi-source data is processed through a financial analysis and prediction network to output prediction results. This network includes a parallel-running revenue prediction model and a component analysis optimization model. The prediction results include revenue changes and dynamic optimization results for global costs. The revenue prediction model uses an attention mechanism to fuse static features, time-series features, and market sentiment semantic vectors learned from the multi-source data, and combines the fusion results to predict revenue changes. The component analysis optimization model is constructed using a multi-agent reinforcement learning model, and obtains dynamic optimization results for global costs through agent collaborative game theory. Visualized reports and analysis reports are constructed based on the prediction results and dynamically displayed to users; The static and time-series features learned from the multi-source data are input into the financial management model. The model performs preliminary analysis using the pre-set audit rules to generate a financial audit report. The financial risk model is then used to monitor the financial audit report. The financial audit report and monitoring results are then simultaneously pushed to the manual audit end for verification and approval. The verification and approval results are synchronously stored in the cloud financial management center and encrypted. The verification and approval results are also used to dynamically optimize the audit rules.

2. The method according to claim 1, characterized in that, After storing the multi-source data in the cloud-based financial management center and encrypting it, the process further includes: The tenant risk prediction model extracts tenant data and publicly available data from the multi-source data and intermediary association data. The tenant data includes tenant performance records, payment behavior, and job stability information. The publicly available data from the intermediary association includes a blacklist and historical arrears records. A tenant relationship graph is constructed based on GraphSAGE, with co-renting relationships and recommendation relationships as edges and tenant credit features as nodes, to extract correlations between tenant data and publicly available data from real estate agencies. The tenant risk prediction model, which incorporates an adversarial training module, processes the correlation extraction results, calculates the tenant's default probability and expected default amount, and determines the default risk level by combining the default probability and expected default amount, dynamically displaying the default risk level to the user.

3. The method according to claim 1, characterized in that, After storing the multi-source data in the cloud-based financial management center and encrypting it, the process further includes: Real-time access to regulations and policies issued by the Ministry of Housing and Urban-Rural Development and tax authorities; construction of a regulations and policies knowledge graph including policy information, compliance requirements, and financial operations; and marking rent control constraints and tax compliance constraints. The apartment's financial operations are adapted to the aforementioned legal and policy knowledge graph; specifically, existing compliance judgment logic is transferred through small-sample learning to determine whether the apartment's financial operations are compatible with updated legal and policy information. Based on the adaptation results, it is determined whether there are compliance risks in the apartment's financial operations, and a compliance risk monitoring report is generated and dynamically displayed to users.

4. The method according to claim 1, characterized in that, The step of outputting prediction results from the multi-source data through a financial analysis and prediction network includes: The multi-source data is preprocessed; multi-dimensional feature data is learned from the multi-source data through an embedding layer, and feature encoding is performed using a feature hierarchical encoding method to obtain static features, time-series features, and market sentiment semantic vectors; the static features, time-series features, and market sentiment semantic vectors are weighted and fused through an attention mechanism. The weighted fusion features are input into the income prediction model, which is an improved Temporal FusionTransformer model. By learning historical transaction data, it outputs income change curves of multiple quantiles and makes multiple scenario predictions including conservative, neutral and optimistic predictions on the income change curves, and fine-tunes them to obtain the income change situation. Cost-related data is extracted from the multi-source data and input into the cost analysis and optimization model. The cost analysis and optimization model adopts an improved PPO algorithm and a multi-objective reward function, and outputs the global cost dynamic optimization result through agent collaborative game. The cost analysis and optimization model is a multi-agent reinforcement learning (MARL) model, which includes independent agents corresponding to the energy consumption system, maintenance system, and labor cost system.

5. The method according to claim 4, characterized in that, After outputting the prediction results from the multi-source data through the financial analysis and prediction network, the method further includes: Collect real-time new transaction data and market dynamic feedback data to form a supplementary data set with the multi-source data; The FTRL algorithm is used to fine-tune the revenue prediction model and cost analysis optimization model in the financial analysis and prediction network online, and the fine-tuned model parameters are synchronously updated to the financial analysis and prediction network. Set a model performance monitoring threshold for the financial analysis and prediction network. When the prediction error of revenue changes or global cost dynamic optimization results exceeds the preset threshold three times in a row, incremental training of the model is triggered.

6. The method according to claim 1, characterized in that, The revenue forecasting model also includes a market mutation detection module; after outputting the forecast results through the financial forecasting network, it further includes: The market mutation detection module calculates the differences in attention to various features and automatically identifies major event nodes; the major event nodes include income mutation event nodes caused by policy adjustments or regional planning. Based on the analysis of the major event nodes, potential abrupt changes in revenue are identified, and revenue risk warning information is generated by combining the scope and degree of the impact of these abrupt changes. Income risk warning information will be pushed to relevant personnel via system messages and / or SMS, and the income risk warning information will be stored in the cloud financial management center and encrypted.

7. The method according to claim 1, characterized in that, The construction of visualization reports and analysis reports based on the prediction results includes: ECharts is used to build dynamic dashboards that generate visual reports by overlaying predicted curves and actual data on revenue changes and global cost optimization results. Data drill-down functionality is also configured for the visual reports. Based on the customized needs of the owner or financial personnel, select the target time period and analysis dimensions, and extract the corresponding predictive data from the revenue changes and the overall cost dynamic optimization results. By parsing the prediction data and visualization reports of the corresponding dimensions through a large language model, a structured analysis and prediction report is generated, which includes data trend charts, a summary of key financial indicators, and an explanation of the causes of differences. The system dynamically displays visual reports and structured analysis and forecast reports to users; and synchronously stores and encrypts these reports in the cloud-based financial management center.

8. The method according to claim 1, characterized in that, The step involves inputting the static and time-series features learned from the multi-source data into the financial management model, performing preliminary analysis using pre-set audit rules within the financial management model, and generating a financial audit report, including: According to the apartment's financial system, expense reimbursement review rules are configured in the financial management model. These expense reimbursement review rules include at least the following dimensions: reimbursement amount, voucher requirements, and approval process. The financial management model adopts a hybrid algorithm architecture of rule engine and gradient boosting tree neural gradient boosting tree (NGBoost). The static features and time-series features learned from the multi-source data are input into the financial management model, and the expense reimbursement review rules are combined with the rule engine and NGBoost model for joint analysis to obtain automatic review results. In accordance with accounting standards and apartment management requirements, financial statements are generated and preliminary analysis is performed. The financial statements are used to indicate at least one core financial indicator among revenue, cost, and profit. The preliminary analysis includes calculation of year-on-year changes in indicators, calculation of month-on-month changes in indicators, identification of abnormal fluctuations, and preliminary determination of their causes. Integrate the results of automatic audits with the preliminary analysis conclusions of financial statements to generate a financial audit report that includes audit opinions, detailed indicators, explanations of anomalies, and explanations of characteristic contributions. Data on questionable entries, compliance probability distribution, and feature contribution during the automatic review process will be synchronized to the cloud-based financial management center.

9. The method according to claim 8, characterized in that, The process involves inputting the static and time-series features learned from the multi-source data into the financial management model, and combining them with the expense reimbursement review rules through a rule engine and the NGBoost model for joint analysis to obtain automatic review results, including: The static and time-series features learned from the multi-source data are used by a rule engine to perform rigid review and screening of non-compliant applications based on the expense reimbursement review rules. The NGBoost model is used to make fine-grained judgments on the remaining applications other than those that violate regulations, and to learn the implicit features of questionable applications among the remaining applications, outputting the compliance probability distribution and feature contribution. Applications with questionable compliance probability distributions that fall below a preset threshold are automatically marked, and questionable points are identified based on feature contribution, resulting in automatic review results.

10. An intelligent auxiliary system for apartment financial management based on artificial intelligence, characterized in that: The system includes: The data acquisition module is used to obtain apartment management data, hotel management data, and publicly available market transaction data to obtain multi-source data; The prediction module is used to output prediction results from the multi-source data through a financial analysis and prediction network. The financial analysis and prediction network includes a parallel-running revenue prediction model and a component analysis optimization model. The prediction results include revenue changes and dynamic optimization results for global costs. The revenue prediction model uses an attention mechanism to fuse static features, time-series features, and market sentiment semantic vectors learned from the multi-source data, and combines the fusion results to predict revenue changes. The component analysis optimization model is constructed using a multi-agent reinforcement learning model, and obtains dynamic optimization results for global costs through agent collaborative game theory. The display module is used to construct visual reports and analysis reports based on the prediction results and dynamically display them to users; The prediction module is also used to input the static features and time-series features learned from the multi-source data into the financial management model, perform preliminary analysis through the audit rules set in the financial management model, generate a financial audit report, call the financial risk model to monitor the financial audit report, and push the financial audit report and monitoring results to the manual audit end for verification and approval. The storage module is used to store the multi-source data to the cloud financial management center and encrypt it; and to synchronously store the verification and approval processing results to the cloud financial management center and encrypt them. The verification and approval processing results are also used to dynamically optimize the audit rules.

Images