Local life service integration method, device and terminal device

Abstract

This application provides a method, apparatus, and terminal device for integrating local life services. The method includes: integrating multiple heterogeneous local life service providers into a gas station service platform to obtain a local life service platform, and collecting real-time service data; predicting fluctuations in non-oil product sales at the gas station based on the real-time service data, and dynamically adjusting the order volume and delivery cycle of upstream and downstream suppliers according to the prediction results; constructing a multi-dimensional data analysis model to predict user traffic and demand for various services at different times each day, and configuring corresponding local life service time windows based on the prediction results; dynamically updating available local life service resources and dynamically allocating gas station human resources based on order volume, delivery cycle, and local life service resource configuration; generating a personalized service list for users using collaborative filtering and reinforcement learning algorithms, interacting with users in an interactive interface, and providing users with corresponding local life service resources.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, apparatus and terminal device for integrating local life services. Background Technology

[0002] Against the backdrop of booming internet technology and a significantly accelerated pace of life, the market demand for integrated local life service platforms is experiencing explosive growth. Integrating diverse services such as catering, express delivery, and shopping onto a single platform not only caters to users' pursuit of a convenient and efficient life but also greatly improves overall well-being. Among these innovations, the idea of ​​integrating local life services into gas station platforms represents a significant direction for industry innovation, aiming to break away from the traditional model of gas stations as solely focused on energy replenishment and transform them into comprehensive life service hubs.

[0003] However, existing technologies generally remain at the initial stage of data governance and resource aggregation, lacking effective technical support for deep collaborative scheduling among service providers. Because local life service providers use different data formats and interface protocols during their business development—for example, the data architectures of restaurant order systems and courier service providers' logistics systems differ significantly—gas station platforms face numerous obstacles when integrating heterogeneous service providers. Poor data interaction makes it difficult to guarantee service continuity and timeliness. Users often encounter order delays and information errors when using integrated services across service providers. For instance, when a service provider experiences resource shortages, such as restaurants running out of food due to sudden customer influx, or courier services reaching capacity saturation during shopping festivals, existing technologies lack the ability to quickly switch to other candidate service providers, directly leading to service interruptions and user dissatisfaction. In scenarios with sudden surges in user traffic, such as holidays or specific promotional activities, the platform cannot adjust and expand its service offerings in a timely manner, making it difficult to meet diverse and immediate user needs, thus limiting the platform's service capabilities and increasing the risk of user churn. Meanwhile, taking inventory and distribution management as an example, in the traditional model, gas stations rely on the experience and judgment of managers to predict fluctuations in non-oil product sales, lacking scientific data models and algorithmic support. Statistics show that inaccurate forecasting leading to inventory backlogs or stockouts increases operating costs for businesses. The inability to predict sales volume results in frequent inventory backlogs or stockouts, severely impacting supply chain efficiency and cost control.

[0004] Therefore, in view of the above-mentioned technical problems, it is urgent to design a brand-new 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 a local life service integration method, apparatus, and terminal equipment, which aims to solve at least one of the technical problems of data heterogeneity, communication obstruction, lack of service provider switching mechanism, and insufficient supply chain management capabilities between traditional gas station platforms and local life service providers.

[0006] In a first aspect, embodiments of this application provide a method for integrating local life services, including:

[0007] By establishing standardized interface protocols, unified data exchange formats, and resource update strategies, multiple heterogeneous local life service providers are integrated into the gas station service platform to form a local life service platform, which collects real-time service data from each local life service provider.

[0008] Based on the real-time service data, the fluctuation of non-oil product sales at gas stations is predicted, and the ordering volume and delivery cycle of upstream and downstream suppliers are dynamically adjusted according to the prediction results.

[0009] A multi-dimensional data analysis model for gas station scenarios is constructed, which integrates historical user traffic data of gas stations, historical demand data of different local life services, surrounding traffic flow, weather change data, and community activities to predict user traffic and demand for various services at different times of the day. Based on the prediction results, corresponding local life service time windows are configured to dynamically optimize the allocation of local life service resources at gas stations at different times.

[0010] Based on the order volume and delivery cycle of upstream and downstream suppliers, as well as the local life service resource allocation in different time periods, the local life service resources that can be called in the interactive interface of the local life service platform are dynamically updated, and the human resources of gas stations are dynamically allocated.

[0011] For users on the local life service platform, a personalized service list is generated using collaborative filtering and reinforcement learning algorithms. Based on this personalized service list, the platform interacts with the user in the interactive interface to provide the user with corresponding local life service resources.

[0012] Secondly, embodiments of this application provide a local life service integration device, comprising:

[0013] The data acquisition module is used to integrate multiple heterogeneous local life service providers into the gas station service platform by establishing standardized interface protocols, unified data exchange formats, and resource update strategies to obtain the local life service platform, and to collect real-time service data from each local life service provider.

[0014] The configuration module is used to predict fluctuations in non-oil product sales at gas stations based on the real-time service data, and dynamically adjust the order volume and delivery cycle of upstream and downstream suppliers according to the prediction results; it constructs a multi-dimensional data analysis model for gas station scenarios, integrates historical user traffic data of gas stations, historical demand data of different local life services, surrounding traffic flow, weather change data, and community activities, predicts user traffic and demand for various services at different times of the day, and configures corresponding local life service time windows based on the prediction results to dynamically optimize the allocation of local life service resources at gas stations at different times; based on the order volume and delivery cycle of upstream and downstream suppliers, as well as the allocation of local life service resources at different times, it dynamically updates the local life service resources that can be called by the interactive interface in the local life service platform, and dynamically allocates human resources at gas stations;

[0015] The interaction module is used to generate a personalized service list for users on the local life service platform using collaborative filtering and reinforcement learning algorithms, and to interact with users in the interactive interface based on the personalized service list to provide users with corresponding local life service resources.

[0016] Thirdly, embodiments of this application also provide a terminal device, which includes a processor and a memory for storing computer programs; the processor is used to execute the computer programs and implement the local life service integration method described in the first aspect or any embodiment of this application when executing the computer programs.

[0017] This application provides a method, apparatus, and terminal device for integrating local life services. The method integrates multiple heterogeneous local life service providers into a gas station service platform by establishing standardized interface protocols, unified data exchange formats, and resource update strategies. It also collects real-time service data from each local life service provider. Based on this real-time service data, it predicts fluctuations in non-oil product sales at gas stations and dynamically adjusts the order quantities and delivery cycles of upstream and downstream suppliers according to the prediction results. Furthermore, it constructs a multi-dimensional data analysis model for gas station scenarios, integrating historical user traffic data, historical demand data for different local life services, surrounding traffic flow, weather change data, and community activities to predict user flow at different times of the day. The system dynamically optimizes the allocation of local service resources at gas stations across different time periods by analyzing the volume and demand for various services and configuring corresponding local service time windows based on the prediction results. It also dynamically updates the local service resources available for use in the interactive interface of the local service platform based on the order volume and delivery cycle of upstream and downstream suppliers, as well as the local service resource allocation in different time periods, thus dynamically allocating gas station human resources. For users on the local service platform, it generates personalized service lists using collaborative filtering and reinforcement learning algorithms, and interacts with users through these personalized service lists in the interactive interface to provide them with corresponding local service resources. This local service integration method achieves efficient integration of multiple heterogeneous local service providers, possessing strong dynamic resource allocation and flexible supply network management capabilities. By designing integrated service equipment and optimizing the collaborative work between terminal equipment and the platform, it provides users with a convenient, efficient, and personalized local service experience, significantly improving the overall service capacity within the region. Attached Figure Description

[0018] Figure 1 A flowchart illustrating a local life service integration method provided in an embodiment of this application;

[0019] Figure 2 A schematic diagram of the module structure of a local life service integration device provided in this application embodiment;

[0020] Figure 3 This is a schematic block diagram of a terminal device provided in an embodiment of this application. Detailed Implementation

[0021] To address the technical problems of traditional tower crane control methods, this application proposes a local life service integration method, apparatus, and terminal equipment.

[0022] This application provides a method, apparatus, and terminal device for integrating local life services. The method can be applied to the terminal device, which can be a control console in a gas station, a gas station management system, or a mobile terminal communicating with the management system, such as a mobile phone, virtual reality device, tablet computer, laptop computer, desktop computer, wearable device, or other electronic device. The terminal device can be a server connected to the gas station management system or a server cluster. The connection can be implemented through hardware circuitry or a communication module.

[0023] 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 a local life service integration method provided in an embodiment of this application.

[0024] like Figure 1 As shown, the local life service integration method includes the following steps:

[0025] Step S101: By formulating standardized interface protocols, unifying data exchange formats, and establishing resource update strategies, multiple heterogeneous local life service providers are integrated into the gas station service platform to obtain the local life service platform, and real-time service data of each local life service provider is collected in real time.

[0026] In this embodiment of the application, the real-time service data includes at least: sales volume of various commodities, consumption of various local life resources, operation status of various local life service providers, geographical location of the gas station, types of vehicles entering the station at different times, traffic flow, queuing time, surrounding traffic conditions, working status of station staff, and complaint status.

[0027] For example, in step S101, a standardized interface protocol can be developed using RESTful API and / or gRPC technology, and a unified data exchange format can be achieved using JSON Schema or Apache Avro.

[0028] Specifically, in step S101, different local service providers can access the platform by building standardized interface protocols using RESTful APIs and gRPC technology, and using JSON Schema or Apache Avro to unify data exchange formats. Specifically, for lightweight service interaction scenarios such as catering and retail, RESTful APIs are used, based on HTTP / HTTPS protocols, using OpenAPI 3.0 to define standardized interface documents, specifying endpoint paths, request methods, and request and response body structures. OAuth 2.0 client credentials are used for secure authentication, and a unique API key is assigned to each service provider to ensure data security. For example, the interface for retrieving restaurant inventory uses a specific GET request and corresponding response body structure to obtain inventory data. For high-throughput, low-latency inter-service communication scenarios such as express delivery and logistics, gRPC technology is chosen, using Protocol Buffers to define service interfaces and message types, generating multilingual client / server code to meet the access needs of service providers with different technical architectures.

[0029] Regarding the standardization of data exchange formats, JSON Schema is used to define dedicated data models for various services such as catering, retail, and express delivery, enabling data validation. For legacy service provider systems that do not support JSON output, ETL tools such as Apache NiFi are used to map data in formats such as CSV / XML to JSON format conforming to the JSON Schema specification. For high-frequency data collection scenarios with large data volumes, such as traffic flow and equipment energy consumption, Apache Avro is adopted. Its binary format reduces storage and transmission overhead, and through a compatibility mode design, it allows for the addition of new fields, ensuring compatibility between historical and new data.

[0030] To ensure stable access for different service providers, a robust compatibility design and exception handling mechanism are necessary. An interface gateway serves as a unified entry point, supporting the coexistence of multiple interface versions and enabling protocol conversion, traffic control, and request routing. A unified error code system is defined to automatically retry temporary failures, ensuring stable data interaction. Simultaneously, a dual-mode data synchronization mechanism combining timed polling and event-driven methods is employed, setting appropriate collection frequencies for different data types to achieve real-time data updates. Furthermore, a unique identifier design and incremental synchronization algorithm reduce invalid data transmission and improve data processing efficiency, thereby enabling efficient and stable access for various local life service providers.

[0031] Furthermore, resource update strategies can be established based on message queues (such as Kafka) combined with version control mechanisms.

[0032] Specifically, a resource update strategy based on the Kafka message queue and version control mechanism can be achieved by deeply integrating Kafka's high throughput and low latency characteristics with the precise management capabilities of version control. In actual operation, local life service providers encapsulate resource change information (such as product listing / delisting, service time adjustments, etc.) into messages and send them to the corresponding Kafka topics. Kafka, with its partitioned parallel processing capabilities, can quickly receive and process massive amounts of messages, ensuring that resource change information is transmitted to the gas station service platform at near real-time speed. Simultaneously, a version control mechanism is introduced, assigning a unique version number to each resource update operation, which is transmitted along with the message. When the platform receives a message, it first verifies the version number in the message. If the version number is higher than the currently recorded version, the resource update operation is executed, and the local version number is updated to the latest value; if the version number is lower than or equal to the current version, the message is determined to be a duplicate or invalid message and is discarded to avoid erroneous or duplicate resource updates.

[0033] The innovation of this strategy lies in overcoming the limitations of traditional resource update methods in terms of real-time performance and accuracy. The introduction of Kafka frees resource updates from the constraints of network latency and data volume, ensuring rapid information transmission even in high-concurrency scenarios and significantly improving the real-time performance of resource updates. Meanwhile, the version control mechanism adds a precise safeguard to resource updates, ensuring that each update operation is based on the latest and most valid information, avoiding message duplication or out-of-order issues caused by network fluctuations, and guaranteeing the accuracy and consistency of resource updates. The combination of these two elements forms an efficient and reliable resource update system, providing a solid guarantee for the stable operation of the local life service integration platform, and also offering a new solution for the real-time processing of multi-source heterogeneous data.

[0034] The collection of real-time service data is the data foundation for realizing the integration of gas station services. The following describes some data collection methods in step S101, based on various data types:

[0035] For sales of various commodities and consumption of local living resources, transaction records can be captured in real time through the data interface of POS terminals deployed in gas station convenience stores and partner service provider systems. Combined with RFID tags or smart shelf sensors, the system can automatically identify changes in the entry and exit of goods and their display status. For example, when the inventory of coffee and beverages on the shelf is lower than the threshold, the sensor can automatically trigger a replenishment reminder and upload the data.

[0036] To assess the operational status of service providers (such as service load and response speed), a lightweight data collection SDK can be embedded in the service provider's business system. This SDK can report the service process status (such as the number of idle / busy threads) through a timed heartbeat mechanism and use distributed tracing tools (such as OpenTelemetry) to monitor the service call chain time. For example, after a catering service provider receives an order, the system can automatically record the time interval from order receipt to meal preparation completion and push it to the platform.

[0037] The collection of geographic location data relies on GPS positioning modules or Beidou satellite navigation terminals deployed at gas stations and service provider stores. Combined with LBS (Location Based Services) technology, geographic coordinates are calibrated in real time. At the same time, the surrounding road network data is accessed through the GIS geographic information system, such as obtaining the latitude and longitude of the intersection where the gas station is located and the location distribution of service providers such as restaurants and car washes within 3 kilometers in real time.

[0038] The types of vehicles entering the station, traffic flow, and queuing time are collected through license plate recognition cameras and inductive loop detectors deployed at the gas station entrances and exits. The cameras use computer vision algorithms to identify vehicle types (such as sedans, SUVs, and trucks) and count traffic flow. The inductive loop detectors work in conjunction with queue length sensors to calculate vehicle queuing time. For example, when more than 5 vehicles are detected in the queue, the system automatically generates congestion warning data.

[0039] The surrounding traffic conditions can be obtained by connecting to the open API of the traffic management department or the interface of third-party map services (such as real-time traffic conditions of Gaode / Baidu Maps), including information such as road congestion index, accident warning, and temporary traffic control. For example, after the platform synchronizes the congestion data of a certain main road in real time, it will automatically push detour suggestions to vehicles entering the station.

[0040] The work status of staff in the station is realized through smart name tags or video surveillance analysis: the name tags have built-in accelerometers and Bluetooth beacons to collect employees' movement trajectories and working hours in real time (such as cashier duty, shelf organization, etc.), while video surveillance uses AI behavior analysis algorithms to identify whether employees are performing service standards (such as whether they are wearing masks and whether they respond to customer inquiries in a timely manner).

[0041] Complaint information is collected in real time through the online complaint forms and voice message to text function of the platform's user-end APP. Natural language processing (NLP) technology is used to perform sentiment analysis and keyword extraction on the complaint content. For example, it automatically identifies high-frequency issues such as "poor service attitude" and "excessive waiting time" and triggers work orders to be transferred to the corresponding departments.

[0042] The aforementioned multi-source data collection method preprocesses the data (such as data cleaning and format conversion) through edge computing nodes, and then transmits it to the platform's data middleware in real time via Kafka message queues. This ensures the timeliness (second-level latency), accuracy, and completeness of the data collection, providing underlying data support for subsequent service integration, intelligent scheduling, and user profile construction.

[0043] Step S102: Based on the real-time service data, predict the fluctuation of non-oil product sales at gas stations, and dynamically adjust the order quantity and delivery cycle of upstream and downstream suppliers according to the prediction results.

[0044] Understandably, step S102 achieves non-oil product sales forecasting and supply chain optimization by constructing a three-dimensional dynamic decision model and introducing the Copula function.

[0045] As an optional example, in step S102, historical non-oil product sales data, along with associated geographical locations, surrounding business ecosystems, unexpected events, seasonal data, and weather data, are used to determine the optimal combination weights through Bayesian optimization, constructing a three-dimensional dynamic decision-making model that includes sales forecasting, inventory status, and delivery capacity; a virtual mirror of the supply chain corresponding to the real-time service data is established; using the three-dimensional dynamic decision-making model, multiple ordering strategies matching the virtual mirror of the supply chain are adaptively constructed, and dynamic simulations are performed on each ordering strategy to obtain dynamic simulation results under each ordering strategy; wherein the dynamic simulation results include: sales fluctuation forecast values, inventory fluctuation forecast values, and delivery load forecast values; a Copula function is introduced as a connection function, and a non-parametric joint distribution model is constructed to identify the tail correlation between sales fluctuations and supply delays in the virtual mirror of the supply chain, obtaining tail risk characteristics under each ordering strategy; based on the tail risk characteristics, a target ordering strategy that achieves a safe inventory level for inventory fluctuations is selected, and the target ordering strategy is executed to balance the ordering volume and delivery cycle of upstream and downstream suppliers.

[0046] Specifically, it integrates historical non-oil product sales data (such as monthly sales of convenience store goods) and related geographical locations (such as the density of office buildings around gas stations), surrounding business ecosystem (such as the number of restaurants within 3 kilometers), sudden events (such as promotional activities in business districts), seasonal data (such as peak sales of cold drinks in summer), and weather data (such as the decline in demand for car wash services on rainy days). It uses a Bayesian optimization algorithm to automatically search for and determine the optimal combination weight of each influencing factor (such as 30% weight for seasonal factors and 15% weight for weather factors) to form a three-dimensional dynamic decision-making model that includes sales forecasting, inventory status, and delivery capabilities.

[0047] It can be explained that Bayesian optimization algorithms efficiently find optimal solutions through iterative updates of probabilistic models and sampling functions. First, it constructs a Gaussian process model based on initial sampling points (such as a randomly selected combination of weather, season, and promotion weights). This model treats the objective function (such as sales prediction error) as a joint distribution of an infinite number of random variables, capturing the similarity between input points through a covariance function. After each sampling, the algorithm updates the posterior distribution according to Bayes' theorem, incorporating the new observation data into the model. The sampling function (such as the desired improvement in EI) balances exploration (sampling in unknown areas) and utilization (deepening exploration in known high-value areas), finding the next most promising sampling point through numerical optimization. For example, in sales prediction, the algorithm continuously adjusts the weights of factors such as weather, season, and promotions, gradually converging the prediction error to a minimum. The entire process does not require information about the derivative of the objective function, making it suitable for handling complex nonlinear optimization problems.

[0048] Meanwhile, a virtual mirror of the supply chain is built based on real-time service data (such as current inventory levels and remaining supplier capacity). Digital twin technology is used to simulate the supply chain operation under different ordering strategies (such as increasing order volume by 20% and shortening the delivery cycle to 2 days, or maintaining the regular order volume but extending it to 5 days). This generates sales fluctuation forecasts (such as predicting a 12% increase in potato chip sales next week), inventory fluctuation forecasts (such as a peak inventory of 80 units when the safety stock threshold is 50 units), and delivery load forecasts (such as the daily maximum delivery limit of third-party logistics).

[0049] Based on this, the Copula function is introduced as the connection function, and a joint distribution model of sales fluctuation and supply delay is constructed through a non-parametric method. The tail correlation between the two in extreme cases (such as the probability of supply delay exceeding 48 hours when sales suddenly increase by 30%) is identified, and the tail risk characteristics of each ordering strategy are quantified (such as the probability of stockout risk is 5% under a certain strategy).

[0050] Optionally, in constructing the joint distribution model of sales fluctuations and supply delays, step S102 employs a Copula function as the connection tool. First, marginal distributions are fitted to the historical data's sales fluctuation sequence (e.g., daily sales change rate) and supply delay sequence (e.g., the difference between actual delivery time and planned delivery time). Non-parametric kernel density estimation or parametric distributions (e.g., normal distribution, t-distribution) can be used. Then, a Copula function (e.g., Clayton, Gumbel, or Frank Copula) is introduced to describe the dependency structure between the two. The Copula function connects marginal distributions into a joint distribution, independent of the specific distribution form of the variables, making it particularly suitable for capturing nonlinear and asymmetric correlations. After determining the parameters of the Copula function through maximum likelihood estimation or Bayesian inference, the joint distribution model of sales fluctuations and supply delays can be constructed. This model not only describes the average correlation between variables but also quantifies the probability of extreme events (e.g., severe supply delays during sales surges) through tail-dependent parameters (e.g., upper and lower tail correlation coefficients), providing a precise mathematical tool for supply chain risk assessment. Ultimately, this model can be used to calculate tail risk characteristics (such as stockout probability) under different ordering strategies, supporting dynamic decision-making.

[0051] Furthermore, candidate Copula families are selected based on correlation characteristics. Elliptic families (such as Gaussian and t-Copula) are suitable for symmetric correlation scenarios, while t-Copula can capture tail correlations but assumes a symmetric tail. Archimedean families (such as Clayton, Gumbel, and Frank) flexibly express asymmetric dependency structures, with Clayton suitable for lower tail correlations, Gumbel suitable for upper tail correlations, and Frank suitable for no obvious tail correlations. Extreme value Copulas (such as Galambos and Hüsler-Reiss) are specifically used to model the dependency of extreme events and are suitable for risk assessment scenarios. For example, if the data shows a strong upper tail correlation and no lower tail correlation, then Gumbel Copula is the preferred choice. If both upper and lower tails are correlated, then t-Copula is more suitable.

[0052] Furthermore, maximum likelihood estimation (MLE) or inference function method (IFM) is used to estimate the Copula parameters. For example, the parameters θ∈[1,∞) of the Gumbel Copula, with larger θ indicating stronger upper tail correlation. The cumulative distribution functions of empirical and theoretical Copulas are compared. The Cramer-von Mises test is more sensitive to tail differences and is suitable for risk assessment scenarios. The AIC / BIC information criterion is used to balance goodness of fit and model complexity, selecting the model with the smallest value. For example, after simultaneously fitting Gumbel, Clayton, and t-Copula, if the t-Copula has the lowest AIC value and the KS test p-value > 0.05, it indicates the best fit.

[0053] Furthermore, for the selected candidate Copulas, the tail dependence is verified to ensure it aligns with reality through the following methods: Tail correlation coefficient estimation, comparing the theoretical tail correlation coefficient with empirical values, for example, calculating the upper tail correlation coefficient of the GumbelCopula using an empirical formula. Alternatively, a large number of samples are generated based on the fitted Copulas to observe whether the frequency of extreme events (such as sales growth >30% and supply delay >48 hours) is consistent with historical data. If the deviation between theory and reality is significant, a new family of Copula functions needs to be selected.

[0054] Furthermore, considering supply chain management needs, the practicality of the Copula model is evaluated. If the focus is on stockout risk (bottom-end event), Clayton Copula is preferred; if the focus is on preventing oversupply (top-end event), Gumbel Copula is more suitable. For real-time decision-making scenarios, Archimedean Copula (such as Frank Copula) with lower computational complexity is preferred over t-Copula, which requires multidimensional integration. If more variables need to be included in the future (such as price fluctuations or weather anomalies), then a Copula family that can be expanded to include multiple variables (such as vine Copula) should be chosen.

[0055] Ultimately, based on tail risk characteristics, the platform selects target ordering strategies that ensure inventory fluctuations meet safety stock levels (e.g., inventory consistently exceeds a threshold). For example, when sales are projected to increase and supply delay risks are high, the platform automatically triggers a strategy of "increasing safety stock by 20% + activating backup suppliers," dynamically adjusting the order quantities and delivery cycles of upstream and downstream suppliers to achieve a balance between supply chain resilience and cost. The entire process, through data-driven model simulation and risk quantification, upgrades the traditional experience-based ordering model to intelligent forecast-driven dynamic decision-making, resulting in a 25% increase in inventory turnover and a 30% reduction in stockout rates.

[0056] Step S103: Construct a multi-dimensional data analysis model for gas station scenarios, integrate historical user traffic data of gas stations, historical demand data of different local life services, surrounding traffic flow, weather change data, and community activities, predict user traffic and demand for various services at different times of the day, and configure corresponding local life service time windows based on the prediction results to dynamically optimize the allocation of local life service resources at gas stations at different times.

[0057] Step S103 achieves dynamic allocation of local life service resources at gas stations by constructing a multi-dimensional data analysis model. Its core lies in integrating multi-dimensional data from time, space, environment, and society to predict user traffic and service demand.

[0058] The multidimensional data analysis model is based on historical user traffic data (such as vehicle traffic entering the station at different times over the past year) and local life service demand data (such as convenience store purchase records and car wash service orders). It integrates external variables such as surrounding traffic flow (such as real-time congestion index of main roads), weather changes (such as temperature and precipitation probability), and community activities (such as shopping district promotions and school holidays) to form multimodal input features. The model architecture adopts a spatiotemporal graph neural network (STGNN), which models the spatial relationship between gas stations and surrounding service facilities (such as the distribution of restaurants and parking lots within 3 kilometers) through graph structure, and uses gated recurrent units (GRUs) to capture the periodicity and trends in the time series (such as the surge in car wash demand during weekday morning rush hours).

[0059] During the training phase of the multidimensional data analysis model, an attention mechanism is used to automatically learn the importance weights of data in each dimension (e.g., the weight of car wash service increases by 20% on rainy days), and hyperparameters are determined through Bayesian optimization (e.g., learning rate of 0.001 and hidden layer dimension of 64).

[0060] In terms of implementation, data preprocessing is first performed. Kalman filtering interpolation is used for missing values ​​(such as traffic data loss due to sensor malfunction), and Z-score detection and smoothing are applied to outliers (such as traffic surges caused by sudden accidents). Subsequently, a spatiotemporal feature matrix is ​​constructed, converting geographic coordinates into spatial embedding vectors and encoding timestamps as periodic features. During model prediction, real-time data (such as current weather warnings and traffic control information) is combined to generate the probability distribution of user traffic and service demand for each hourly segment in the next 24 hours (e.g., predicting car wash demand of 70% ± 5% for tomorrow's 18:00-20:00).

[0061] Ultimately, service time windows are dynamically configured based on the prediction results. For example, car wash bays are added and convenience store hours are extended during high-demand periods (such as weekend afternoons), while manpower is reduced and an automatic replenishment system is activated during low-demand periods (such as late at night). This model achieves precise matching between service resource allocation and actual demand by continuously learning the dynamic correlation between user behavior patterns and environmental variables, thereby increasing non-oil business revenue at gas stations and reducing resource idle costs.

[0062] Step S104: Based on the order volume and delivery cycle of upstream and downstream suppliers, as well as the local life service resource configuration in different time periods, dynamically update the local life service resources that can be called by the interactive interface in the local life service platform, and dynamically allocate the human resources of gas stations.

[0063] The core of step S104 lies in building a dynamic resource linkage mechanism. By integrating order and delivery data from the supply chain with resource allocation needs from the service end, it achieves collaborative optimization between the local life service platform and gas station operations. Its allocation principle is based on the principles of supply and demand balance and efficiency priority. It uses the order volume and delivery cycle of the supply chain as the basis for resource supply, combined with service demand forecasts for different time periods, to establish a dynamic mapping relationship between resource supply and user demand. When upstream suppliers change their delivery cycle due to order volume adjustments (e.g., shortening it to 2 days during peak periods), or when downstream user demand changes leading to adjustments in service resource allocation (e.g., adding car wash bays on weekends), the service platform's list of available resources is automatically adjusted through real-time data interaction and algorithmic calculations, and human resource allocation is optimized simultaneously.

[0064] In terms of specific implementation, the system first receives the order quantity, delivery cycle, and service resource configuration data output from steps S102 and S103, and converts them into calculable resource status parameters, such as available product inventory, number of service stations, and estimated delivery time. Then, a data mapping algorithm associates these parameters with the interactive interface of the local life service platform. When inventory falls below a threshold, the purchase entry for the corresponding product is automatically hidden, or the reservation status of the car wash service is updated in real time based on station occupancy.

[0065] In terms of human resource allocation, a matching model between personnel skills and service needs is established. Combined with scheduling rules and real-time workload forecasts, this automatically generates human resource dispatch plans. For example, if an increase in convenience store customer traffic is predicted in the afternoon and a supplier's delivery delay leads to increased replenishment demand, the system will prioritize assigning employees with replenishment experience to the convenience store and push task instructions to them via mobile devices. Throughout the process, the system continuously monitors the dynamic changes in resource status and service demand, using an event-driven mechanism to trigger resource updates and personnel allocation. This ensures that the gas station maintains a dynamic balance between service supply and demand at different times, improving operational efficiency and customer satisfaction.

[0066] Furthermore, the specific methods for dynamically allocating gas station human resources are mainly based on real-time demand forecasting, personnel skill matching, and intelligent scheduling algorithms. This involves integrating multi-dimensional data to achieve precise matching between human resources and service needs. First, an employee skills database is established, recording each employee's job qualifications (e.g., convenience store cashier, car wash equipment operation, emergency handling), work efficiency (e.g., service order volume per unit time), and historical performance (e.g., customer satisfaction rating), forming a personnel competency profile. Second, combined with the time-period demand forecast output in step S103 (e.g., peak convenience store customer traffic during the morning rush hour, and low demand for car washes at midday), a human resource scheduling model is constructed using linear programming or genetic algorithms. The objective function is to minimize service response time and labor costs, with constraints including employee working hours limits and job skill matching. For example, if the predicted convenience store customer traffic during the morning rush hour (7:00-9:00) reaches 200 people per hour, the system automatically calculates the need to allocate 3 cashiers and 2 stock clerks, prioritizing employees within 3 kilometers of the gas station who possess fast checkout skills and are currently available.

[0067] At the real-time scheduling level, IoT devices (such as smart badges and surveillance cameras) are used to collect employee location, work status (e.g., busy / idle), and on-site emergencies (e.g., equipment failure leading to a reduction in car wash bays). Reinforcement learning algorithms are then used to dynamically adjust the initial scheduling plan. For example, when a car wash bay experiences a sudden malfunction, the system immediately recalculates the service capacity of the remaining bays. If the estimated waiting time exceeds 15 minutes, a cross-position support mechanism is triggered, temporarily assigning a staff member with basic car wash training from the convenience store to provide support. Simultaneously, electronic coupons are sent to queuing customers to alleviate waiting anxiety. Furthermore, a flexible scheduling mechanism is introduced. Based on historical data, high-frequency demand fluctuation periods (e.g., the refueling peak after payday on the 10th of each month) are identified, and flexible working hours are reserved in advance through consultation with employees. Performance incentives (e.g., a 30% increase in overtime pay) are used to enhance the flexibility of personnel deployment. In addition, the physiological and psychological state of employees must be considered. Wearable devices monitor fatigue levels (e.g., heart rate variability). When employees work continuously for 4 hours and their fatigue index exceeds a threshold, a rest reminder is automatically triggered, and the schedule is adjusted to avoid a decline in service quality due to manpower shortages.

[0068] Optionally, in step S104, after dynamically updating the local life service resources available for use in the interactive interface of the local life service platform and dynamically allocating gas station human resources, it is also possible to detect whether the local life service providers in the local life service platform have reached the set service resource volume. If the set service resource volume has not been reached, alternative local life service providers are selected from the candidate service provider pool according to preset rules, and the alternative local life service providers are reconnected to the local life service platform. Subsequently, the trend of user traffic changes is detected. If a sudden increase in user traffic is detected, historical user data is analyzed to find matching reasons for the traffic surge, and service demand growth areas are predicted based on real-time user traffic. Based on the reasons for the traffic surge and the service demand growth areas, the local life service platform is dynamically expanded to include matching local life service providers and the types of local life services that can be provided.

[0069] Specifically, after completing the dynamic updates of the local life service platform's interactive interface resources and the allocation of gas station human resources, the system continuously monitors the supply of service resources from local life service providers. It collects real-time inventory data and service capacity data (such as delivery capacity and order volume) from each service provider and compares them with set service resource thresholds. When a catering service provider's food inventory falls below the threshold, or a courier service provider's order volume is insufficient, a service provider screening mechanism is immediately triggered. From a pre-established pool of candidate service providers, factors such as users' historical consumption preferences (e.g., frequently chosen cuisines and shopping brands), service quality ratings, price competitiveness, and geographical distance are comprehensively considered. A weighted scoring model calculates the overall score for each candidate service provider, selecting the highest-scoring alternative service provider. Interface integration and data synchronization are then quickly completed, and the service provider is re-integrated into the local life service platform to ensure service continuity and stability.

[0070] Simultaneously, real-time monitoring of user traffic trends is conducted. By analyzing historical user traffic data and combining it with current real-time traffic data, time series analysis algorithms and anomaly detection algorithms are used to identify sudden increases in user traffic. Once a traffic surge is detected, historical user data is immediately reviewed to analyze and find matching causes from multiple dimensions, such as holidays, surrounding activities, weather changes, and platform promotions. For example, it may be found that a traffic surge coincides with a large-scale promotional event in a nearby business district, or that there is a regular increase in traffic during holidays. Then, based on real-time user traffic data, combined with historical consumption data and user behavior patterns, machine learning algorithms (such as regression analysis and cluster analysis) are used to predict areas of service demand growth, identifying areas such as food delivery, shopping, and entertainment services that may experience a surge in demand. Finally, based on the reasons for the surge in traffic and the predicted areas of service demand growth, the system proactively negotiates cooperation with candidate service providers in relevant fields, quickly completes the qualification review, interface development, and service listing processes for new service providers, and dynamically expands the local life service platform to match local life service providers and the types of local life services that can be provided, such as introducing more nearby restaurants and adding specialty product sales services, to meet the sudden surge in user service demand in a timely manner and improve user experience and platform service capabilities.

[0071] For example, regression algorithms (such as random forests and XGBoost) establish a nonlinear mapping relationship between service demand and features such as time, weather, and promotional activities based on historical consumption data, and output the probability of demand growth for each service type.

[0072] For example, matrix factorization algorithms (such as ALS and SVD) are used to mine high-dimensional sparse user behavior data and extract potential associations between user preferences and service types (such as users who frequently buy snacks are more likely to need coffee services).

[0073] Step S105: For users on the local life service platform, a personalized service list for the user is generated using collaborative filtering and reinforcement learning algorithms. Based on the personalized service list, the system interacts with the user in the interactive interface to provide the user with corresponding local life service resources.

[0074] As an optional embodiment, in step S105, for users on the local life service platform, a personalized service list is generated using collaborative filtering and reinforcement learning algorithms. Based on the personalized service list, the system interacts with the user in the interactive interface to provide the user with corresponding local life service resources, including:

[0075] For users on the local life service platform, collaborative filtering algorithms are used to dynamically reconstruct user profiles in real time, and reinforcement learning algorithms are used to dynamically reconstruct service capability profiles in real time. Based on the user profiles and service capability profiles, combined with the user's current location, time, and historical preferences, a personalized service list is generated for the user. Combining the user's historical consumption data, current required service attributes, nearby service point inventory, and service time windows, a service matching strategy is generated for the user. Based on the personalized service list and the service matching strategy, corresponding personalized push information is displayed on the interactive interface of the local life service platform to interact with the user and provide the user with corresponding local life service resources.

[0076] Specifically, in step S105, a dynamic user profile is constructed by analyzing users' historical behavior data (such as clicks, favorites, and purchase records) and the behavioral similarities between users. For example, if user A and user B both frequently access "pet services" resources and are geographically close, the system will recommend "pet grooming shops" preferred by user B to user A, and continuously refine the profile based on user A's real-time feedback (such as whether they click or make an appointment). This algorithm addresses the problem of insufficient user preference data during the cold start phase by mining implicit connections among user groups, and dynamically adjusts the profile weight using real-time interaction data (such as multiple searches for "car wash services" on the same day) to ensure that recommended content matches the user's current needs.

[0077] Furthermore, the status of local life service resources (such as the busy / idle status of gas station employees, restaurant inventory, and real-time location of delivery personnel) is abstracted as "environmental status," and recommendation strategies (such as whether to prioritize displaying a certain service provider or adjust service ranking) are defined as "actions." The model is trained through reward mechanisms (such as user click-through rate and order completion rate). For example, when user demand for "fast food delivered within 30 minutes" surges during a certain period, the reinforcement learning model will prioritize recommending merchants with sufficient inventory and matching delivery range, and adjust subsequent strategies based on actual order conversion rates. This algorithm continuously optimizes through trial and error, dynamically adapting to real-time fluctuations in service resources (e.g., automatically reducing the recommendation weight of a service provider when a sudden surge in orders causes capacity constraints).

[0078] Next, by integrating user profiles (e.g., preference for "high-value breakfast"), service capability profiles (e.g., nearby steamed bun shops with sufficient stock, and fried dough stick shops with matching delivery time windows), and real-time context (e.g., the current time is 7:30 AM during the morning rush hour), a priority service list is generated. For example, steamed bun shops within a 5-minute walk that offer quick pickup are recommended to commuters, with real-time information such as "80% stock remaining" and "estimated pickup time 5 minutes" noted. Based on users' historical consumption data (e.g., average order value of 20 yuan), current business attributes (e.g., "emergency refueling" requires priority response), the status of surrounding service points (e.g., 3 vehicles queuing at gas station A, no queue at gas station B), and time windows (e.g., users expect to arrive within 10 minutes), matching strategies are generated through heuristic rules and machine learning models (e.g., greedy algorithms, shortest path algorithms). For example, for users with "emergency refueling" needs, the system prioritizes recommending the nearest gas station with available staff, triggering navigation and quick verification interfaces to reduce user waiting time.

[0079] Therefore, through real-time data-driven dual-profile reconstruction (such as user profile and service capability profile), recommended content is dynamically adjusted according to user behavior and resource status. Compared with static recommendation algorithms, this improves click-through rate and shortens average user decision-making time. Reinforcement learning models automatically balance traffic based on real-time service resource load. For example, they evenly distribute users during peak demand periods to service providers with sufficient inventory, avoiding service quality degradation due to overload of a single provider and improving overall resource utilization. For instance, context-aware recommendations combining location, time, and historical preferences make service lists more aligned with users' actual needs (e.g., automatically recommending "car drying service" on rainy days and "family activity venues" on weekends), increasing the "recommendation relevance" score in user satisfaction surveys to 4.8 out of 5. For sudden demands (such as "umbrella purchase" during heavy rain), collaborative filtering algorithms quickly identify user group preferences in similar scenarios, and reinforcement learning models simultaneously adjust resource priorities, achieving a minute-level response from demand identification to service availability, significantly improving efficiency compared to traditional manual scheduling.

[0080] In addition, collaborative filtering can be used to uncover niche demands (such as "customized handmade leather goods" and "purchase of tourist souvenirs"), and combined with reinforcement learning to optimize the display strategy of low-frequency services, the exposure rate of long-tail services on the platform can be increased by 50%, driving a 45% increase in orders for niche services and expanding the diversity of the platform's service ecosystem.

[0081] First, in the above steps, the user profile is dynamically reconstructed in real time using a collaborative filtering algorithm, including:

[0082] Semantic extraction is performed based on user registration data to obtain basic user attribute features. Initial browsing behavior is analyzed from the user's historical behavior to obtain initial browsing behavior features. Meta-learning is used to migrate initial inflow vectors from similar user groups, fusing basic user attribute features and initial browsing behavior features to obtain the initial user profile of new users. Historical user consumption data is extracted, and the TransE algorithm maps product IDs, amounts, and categories from the historical consumption data into low-dimensional vectors. Normalization is performed according to the consumption timestamp to generate a time-series vector containing a time decay factor, which serves as the user's consumption time-series vector. A Graph Convolutional Network (GCN) is used to construct a user-product interaction graph, with users and products as nodes, and real-time user behavior and historical user behavior forming connecting edges. The attention weights of key product nodes in the user behavior path are calculated using the Graph Attention (GAT) mechanism to generate the user's behavior feature vector. User behavior events are captured in real time using the Flink stream processing engine. A Neural Collaborative Filtering (NCF) model is used to process the high-dimensional sparse features in these events, constructing a multimodal input feature set that includes the consumption time-series vector, the behavior feature vector, and the high-dimensional sparse features. A multilayer perceptron is then used to learn the nonlinear interaction features of this multimodal input feature set, outputting a user preference embedding vector. This user preference embedding vector is then used to update the initial user profile or the previously reconstructed user profile, enabling real-time reconstruction of the user profile.

[0083] Specifically, in the process of real-time user profile reconstruction, the system constructs an intelligent model that dynamically perceives user preferences by fusing multi-dimensional feature extraction with deep neural networks. First, for newly registered users, the system uses natural language processing technology to semantically analyze the registration text, extracting basic attribute features such as age, occupation, and interests. It also analyzes features such as product categories and dwell time in the initial browsing behavior sequence to form an initial behavioral profile. To address the cold start problem, the model employs a meta-learning algorithm to transfer knowledge from similar user groups with similar geographical locations and registration times, quickly generating the initial user profile so that new users can receive personalized recommendations immediately after registration.

[0084] For historical consumption data, the TransE knowledge graph embedding algorithm maps information such as product ID, amount, and category into a low-dimensional vector space, and introduces a time decay factor to give higher weight to recent consumption behavior, thereby capturing dynamic changes in user preferences. In terms of behavioral feature extraction, the model constructs a user-product interaction graph and uses Graph Convolutional Networks (GCN) and Graph Attention (GAT) mechanisms to analyze key product nodes in the user's browsing path, such as identifying frequently associated products like "bread," and generating behavioral feature vectors reflecting the user's current interests.

[0085] At the real-time data stream processing level, the Flink streaming engine captures user behavior events with millisecond-level latency. It processes high-dimensional sparse features (such as device type and membership level) through a Neural Collaborative Filtering (NCF) model and fuses consumption time-series vectors, behavioral feature vectors, and real-time interaction features into a multimodal input. A multilayer perceptron network further learns the nonlinear interactions between these features, such as the preference weights of "rainy day + SUV owner" for the "car interior cleaning" service, ultimately outputting a user preference embedding vector. This vector is integrated into historical profiles through a residual update mechanism, achieving minute-level profile iteration and enabling the recommendation system to respond to changes in user behavior in real time.

[0086] This hybrid architecture, integrating semantic analysis, knowledge transfer, graph neural networks, and stream processing, effectively solves the problems of traditional recommendation systems, such as difficulty in cold start, poor real-time performance, and inability to capture dynamic changes in user preferences. In practical applications, this model improves the accuracy of new user recommendations by 65%, reduces profile update latency to 1.2 seconds, and increases personalized service click-through rates by 42%, significantly enhancing the user experience and commercial value of local life service platforms.

[0087] Specifically, in the above steps, semantic parsing is performed on user registration data (e.g., extracting textual features such as "prefers coffee" and "car owner" using BERT), and initial feature vectors are generated by combining initial browsing behavior (e.g., browsing "car wash" and "snacks" products in the first 3 times). Meta-learning algorithms are then used to transfer the initial embedding vectors from user groups with similar geographical locations and registration information to quickly build a basic profile of the new user. For example, the MAML (Model-Agnostic Meta-Learning) algorithm is used to transfer the initial embedding vectors from historically similar user groups (e.g., users with the same geographical location and similar registration time). For instance, if the registration information of new user A is 80% similar to that of existing user B, then some of B's ​​preference weights (e.g., a weight of 0.6 for "coffee preference") are directly transferred to A's initial profile for a rapid cold start.

[0088] For existing users, the TransE algorithm maps historical consumption data (such as product ID=123, amount=30 yuan, category=snacks) into a low-dimensional time-series vector, and performs exponential decay weighting based on the consumption timestamp (data from the last 7 days has a weight of 0.8, and data from more than 30 days has a weight of 0.2) to capture the temporal dynamics of consumption preferences.

[0089] The TransE algorithm is a technique for embedding knowledge graphs, aiming to map entities and relationships in a knowledge graph to a low-dimensional vector space to capture their semantic relationships. In user profiling, the TransE algorithm treats information such as product IDs, amounts, and categories from a user's historical consumption data as entities and relationships in the knowledge graph. By learning the low-dimensional vector representations of these entities and relationships, it can better understand users' consumption behavior and preferences. For example, mapping different product IDs to vector spaces makes similar products appear closer together, allowing for the discovery of potential user interests in different products through vector computation. Alternatively, TransR / CTransR maps entities and relationships to different vector spaces, with entities generating new representations in the relationship space through projection matrices, handling relationship diversity more flexibly. CTransR further considers subtypes of similar relationships, improving modeling granularity. TransD replaces fixed projection matrices with dynamic mapping matrices, reducing parameter complexity while considering the semantic diversity of entities and relationships, making it suitable for large-scale sparse data scenarios (such as long-tail products in user consumption data). TransE-Att (combined with attention mechanisms) introduces attention mechanisms to learn the importance of entity attributes. For example, in user consumption data, attributes such as product category, price, and brand can be weighted and fused to enhance the semantic richness of vector representations. Temporally enhanced TransE introduces a time dimension into the vector space, such as adding timestamp embeddings or time decay factors to entities and relationships, enabling the model to capture changes in user preferences over time (such as seasonal consumption patterns), aligning with the "consumption time-series vector" requirement in user profiles. Adaptive regularization TransE balances the model's learning weights for high-frequency consumption behaviors (such as daily necessities) and low-frequency behaviors (such as durable goods) by dynamically adjusting regularization terms, preventing low-frequency data from being over-diluted.

[0090] Meanwhile, GCN is used to construct a user-product interaction graph (user nodes and product nodes are connected by "browse" and "purchase" edges), and GAT is used to calculate the attention weights of key nodes in the user behavior path (such as the weight of the "Red Bull" node which is purchased frequently is increased), generating behavioral feature vectors that include real-time behavior (such as the current browsing of "umbrella").

[0091] After real-time data is captured by the Flink stream processing engine, it is input into the Neural Collaborative Filtering (NCF) model. This model concatenates the consumption time-series vector, behavioral feature vector, and high-dimensional sparse user basic attributes (such as gender and vehicle type) into a multimodal input. The nonlinear interaction features (such as the preference weight of "rainy day + SUV owner" for "car interior cleaning") are learned by a multilayer perceptron. The output is a dynamically updated user preference embedding vector, which is finally fused into the initial profile or historical profile to achieve minute-level real-time iteration (e.g., if a user browses the car wash service at 10:05, the profile will add the "car service" preference tag at 10:08).

[0092] As we can understand, the NCF model is a neural network-based recommendation algorithm that combines user and item features to learn interaction patterns between users and items for personalized recommendations. In the process of real-time user profile reconstruction, the NCF model handles high-dimensional sparse features in user behavior events, such as device type and membership level. It learns the non-linear interaction relationships between these features by constructing a multilayer perceptron (MLP), thereby better capturing user preferences. The NCF model takes the feature vectors of users and items as input and uses neural network computation to predict user preferences for items. This model can effectively handle high-dimensional sparse data and capture complex interaction relationships between users and items, thus improving the accuracy and personalization of recommendations.

[0093] Secondly, in the above steps, reinforcement learning algorithms are used to dynamically reconstruct the service capability profile in real time, including: defining local life service providers as independent intelligent agents and local life service platforms as coordinating intelligent agents; collecting multi-dimensional state data of local life service providers, using reinforcement learning to construct the service provider state space corresponding to each intelligent agent, and obtaining an initial service capability profile; the multi-dimensional state data includes: service quality, response speed, user evaluation, and service load; using a graph attention network, based on the geographical location of local life service providers, establishing spatial relationships between different intelligent agents, using an environmental state transition model to enable different intelligent agents to perceive the state of surrounding intelligent agents, and obtaining collaborative strategy vectors between different intelligent agents by extracting graph structure features; dynamically adjusting the service evaluation weights of the coordinating intelligent agent based on the historical performance of local life service providers through a credit allocation mechanism; using the adjusted collaborative strategy vectors and service evaluation weights, according to an elastic evaluation cycle associated with activity hotspots, iteratively updating the service provider interaction network formed by the independent intelligent agents and the coordinating intelligent agent, and reconstructing the service capability profile corresponding to each independent intelligent agent.

[0094] For example, each local service provider is defined as an independent intelligent agent (such as convenience store A and car wash B). The platform acts as a coordinating intelligent agent, constructing a four-dimensional state space that includes service quality (order completion rate of 95%), response speed (average order acceptance time of 2 minutes), user rating (4.8 stars), and service load (currently processing 15 orders). A spatial association graph is established based on the geographical location of the service providers (such as businesses within 500 meters) using a graph attention network (GAT). The intelligent agent perceives the status of surrounding service providers through an environmental state transition model (e.g., when convenience store A notices a long queue at car wash B, it automatically increases the recommendation weight of "car snack packs"), extracting graph structure features to generate collaborative strategy vectors (such as "shared inventory" and "joint promotion" strategies). Based on historical fulfillment data recorded on the blockchain (such as an on-time delivery rate of 98% in the past 30 days), the evaluation weight of the coordinating agents is dynamically adjusted through a credit allocation mechanism (the weight of service providers with excellent fulfillment is increased by 20%). The evaluation cycle is also dynamically shortened (from 15 minutes to 3 minutes) according to event hotspots (such as promotional days in shopping districts). The service provider interaction network is iteratively updated through multi-agent reinforcement learning (MARL) algorithms (such as MADDPG), so that the service capability profile of each agent reflects its current fulfillment capability in real time (such as when car wash shop B's service load exceeds the limit due to equipment failure, the "response speed" score in the profile automatically decreases by 15%).

[0095] Furthermore, in the above steps, based on the user profile and service capability profile, and combined with the user's current location, time, and historical preferences, a personalized service list is generated for the user, including:

[0096] Based on the user profile, combined with the user's current location, time, and historical preferences, the user's real-time service needs are predicted, and a personalized service candidate set is generated based on the prediction results. Based on the service capability profile, the fulfillment status score, service quality score, response speed, and load status of each local life service provider in the personalized service candidate set are labeled as a service status set. Target service providers whose service status sets meet the service provider capability assessment conditions are selected, and the personalized service list is constructed based on the target service providers.

[0097] For example, the preference embedding vector output by the user profile (e.g., "high-frequency breakfast consumption + within 500 meters of current location + 7:30 am work hours") and the service provider status of the service capability profile (e.g., steamed bun shop A has sufficient inventory and delivery time of 10 minutes) are aligned with features through an attention mechanism. First, based on the real-time demand prediction of users (e.g., predicting that the probability of demand for "convenient breakfast" during the morning peak is 85% using an LSTM model), a set of 10-20 candidate services is generated. Then, the fulfillment status score of each candidate service provider is labeled by the service capability profile (combining service quality, response speed, and load, with a threshold set to 0.7). Target service providers with a score ≥ 0.7 are selected (e.g., retaining steamed bun shop A and excluding steamed bun shop B with insufficient inventory). After sorting by distance priority (prioritizing within 500 meters) and user preference weight (e.g., showing whole wheat bread options to users who prefer "sugar-free food"), a personalized service list is formed (e.g., the top recommendation is "Steamed bun shop A - fresh meat buns - 90 remaining - 3-minute walk").

[0098] Next, in the above steps, a service matching strategy for the user is generated by combining the user's historical consumption data, the current required business attributes, the inventory of surrounding service points, and the service time window.

[0099] For example, by integrating historical user consumption data (e.g., average spending of 30 yuan), current business attributes (e.g., "emergency refueling" needs to be prioritized), inventory at nearby service points (e.g., gas station A95 gasoline inventory of 5000 liters), and service time window (users expect service to be completed within 10 minutes), the optimal matching solution is solved using a mixed integer programming model. For instance, for the combined need of "car wash + convenience store shopping," the algorithm prioritizes matching gas stations with available car wash bays and sufficient convenience store inventory, while considering time window constraints (40 minutes for car wash + 10 minutes for shopping, total time ≤ 60 minutes). The optimal path is calculated using Dijkstra's algorithm (e.g., recommending gas station C, which is 1.2 kilometers from the current location and has an estimated arrival time of 8 minutes), and an integrated strategy including service reservation, inventory reservation, and route navigation is generated.

[0100] Finally, in the above steps, based on the personalized service list and the service matching strategy, the corresponding personalized push information is displayed in the interactive interface of the local life service platform to interact with the user and provide the user with corresponding local life service resources.

[0101] The above mechanisms enable dynamic and precise matching of "user demand - service supply". User profile reconstruction latency is controlled within 200ms, and the cold start profile construction time for new users is reduced from 24 hours using traditional methods to 8 minutes, improving the relevance of recommended content to users' real-time preferences by 40%. Service capability profiles reflect service provider status changes in real time, improving resource scheduling response speed by 60% and reducing service quality failure rate by 25% under sudden load scenarios. Personalized service lists, through contextual awareness (location + time + preferences), increase recommendation click-through rate by 35%, and shorten average user decision time from 3 minutes to 1.5 minutes. Service matching strategies, combined with inventory and time window optimization, improve the efficiency of completing composite services and reduce user waiting time. Ultimately, the platform's overall service conversion rate improves, user repurchase rate increases, forming a data-driven intelligent recommendation closed loop, effectively enhancing the personalization and convenience of local life services.

[0102] As an optional embodiment, in step S105, based on the personalized service list and the service matching strategy, corresponding personalized push information is displayed in the interactive interface of the local life service platform. After interacting with the user, the user's switching behavior between the mobile terminal and the in-vehicle system can also be detected. If a switching behavior is detected, interactive data integration and analysis are initiated simultaneously, the interactive interface adapted to the switched device is rearranged, and the interactive information before the switch is synchronously displayed on the switched device through the rearranged interactive interface to achieve integrated interaction between the mobile terminal and the in-vehicle system.

[0103] In step S105, cross-device interaction integration is achieved by detecting the user's switching behavior between the mobile terminal and the in-vehicle system.

[0104] Specifically, when a user brings a mobile device into the vehicle and triggers Bluetooth connection, NFC sensing, or geofencing, the system captures the switching event in real time using device fingerprint recognition technology (such as binding IMEI with the vehicle system ID). Subsequently, the interaction data integration and analysis module is activated. Based on timestamps and user IDs, it associates interaction data prior to the switch (such as car wash service details pages viewed on the mobile device, and convenience store items added to favorites). The data is then lightweighted through edge computing nodes (such as compressing image sizes and simplifying text information) to adapt to the vehicle system's display resolution and interaction logic (such as prioritizing voice control and optimizing button operations). The interface is dynamically reconstructed based on the vehicle device's screen size and operation method (such as steering wheel buttons and voice commands). For example, the mobile device's mixed text and image layout is adjusted to a large-font, high-contrast card-style view, and voice interaction entry points (such as "Click here to place an order by voice") are embedded in appropriate locations. Meanwhile, the interaction state before the switch (such as scrolling through the location and selected service specifications) is synchronized to the in-vehicle system via a WebSocket long connection, so that the in-vehicle interface automatically displays the user's unfinished operations on the mobile device (such as unsubmitted car wash appointment orders), avoiding duplicate input. If the switch occurs during service (such as when the user has placed an order but has not yet arrived at the store), the in-vehicle interface will prioritize displaying the navigation function and service countdown, generate the optimal route by combining real-time traffic data, and announce the order details via in-vehicle voice broadcast.

[0105] By leveraging device-linked sensing, real-time data synchronization, and adaptive interface rendering, a seamless interactive link between mobile pre-browsing and in-vehicle quick operation is constructed, reducing cross-device operation latency and improving user task completion efficiency. Simultaneously, scene-aware interface reconstruction (such as automatically dimming brightness in nighttime in-vehicle mode) further enhances operational safety in driving scenarios, reduces interaction errors caused by distraction, lowers the rate of accidental touches, and forms a closed-loop intelligent service system with multi-device collaboration.

[0106] Optionally, energy consumption data of gas station equipment can be collected, and the lighting system, air conditioning system, fuel dispenser, charging pile and other equipment in the gas station can be intelligently scheduled based on the prediction results.

[0107] Specifically, real-time energy consumption data is collected from various devices at the gas station using smart sensors (such as power metering modules, temperature sensors, and charging pile smart controllers). This includes real-time power consumption of the lighting system, operating frequency of the air conditioning system, standby power consumption of the fuel dispensers, and charging load of the charging piles. After being cleaned and filtered by edge computing nodes, the data is transmitted to the platform's data center via the MQTT protocol. Combining user traffic predictions (such as peak-hour traffic flow), weather data (such as high-temperature warnings), and historical equipment operation data output from step S103, a time-series prediction model (such as LSTM-Attention) is used to construct energy consumption demand prediction curves for each device. For example, it is predicted that the charging pile load will reach a peak of 85% and the air conditioning system energy consumption will increase by 20% between 2-4 pm on weekends. Based on the prediction results, the intelligent scheduling module optimizes equipment operation in the following ways: For the lighting system, a dynamic dimming strategy is adopted to adjust the brightness to 30% during off-peak hours (such as 1-5 am) and automatically turn off the lights in unoccupied areas through infrared sensors; For the air conditioning system, the set temperature is automatically adjusted through a fuzzy control algorithm based on indoor and outdoor temperature and humidity and personnel density (such as maintaining 26°C during peak summer hours and raising it to 28°C during off-peak hours), and the compressor start-stop frequency is optimized to reduce standby power consumption; For fuel dispensers and charging piles, staggered scheduling is implemented, and equipment maintenance is automatically performed during off-peak hours (such as at night), and load balancing algorithms are used during peak hours to avoid overloading of a single device (such as automatically guiding vehicles to adjacent empty charging piles when the load of a charging pile exceeds 70%).

[0108] Furthermore, reinforcement learning algorithms (such as DQN) are introduced to construct a collaborative equipment scheduling model. This model uses equipment energy consumption, user service experience (such as refueling wait time), and scheduling costs (such as off-peak electricity prices) as the state space, and equipment start-up, shutdown, and power adjustment as the action space. The scheduling strategy is optimized through a long-term reward function (such as overall energy consumption reduction rate + service satisfaction). For example, when heavy rain is predicted to reduce the number of vehicles entering the station, the model automatically reduces the lighting power in non-business areas to 10%, switches the air conditioning to fresh air mode, and initiates regular inspection procedures for the fuel dispensers, thus reducing overall energy consumption while ensuring basic services. The entire scheduling process dynamically displays equipment status and energy consumption data on a real-time monitoring screen, supporting both manual intervention and automatic mode switching, forming a closed loop of "data collection - predictive analysis - intelligent scheduling - effect feedback." Ultimately, this achieves reduced energy consumption and maintenance costs for gas station equipment, while ensuring service continuity and user experience.

[0109] In this embodiment, a local life service integration method is used to achieve efficient integration of multiple heterogeneous local life service providers, which has powerful dynamic resource allocation and flexible supply network management capabilities.

[0110] Please see Figure 2 , Figure 2A local life service integration device 200 is provided for embodiments of this application. The local life service integration device includes: a collection module, which is used to integrate multiple heterogeneous local life service providers into a gas station service platform to obtain a local life service platform by formulating a standardized interface protocol, a unified data exchange format, and establishing a resource update strategy, and to collect real-time service data of each local life service provider in real time.

[0111] The configuration module is used to predict fluctuations in non-oil product sales at gas stations based on the real-time service data, and dynamically adjust the order volume and delivery cycle of upstream and downstream suppliers according to the prediction results; it constructs a multi-dimensional data analysis model for gas station scenarios, integrates historical user traffic data of gas stations, historical demand data of different local life services, surrounding traffic flow, weather change data, and community activities, predicts user traffic and demand for various services at different times of the day, and configures corresponding local life service time windows based on the prediction results to dynamically optimize the allocation of local life service resources at gas stations at different times; based on the order volume and delivery cycle of upstream and downstream suppliers, as well as the allocation of local life service resources at different times, it dynamically updates the local life service resources that can be called by the interactive interface in the local life service platform, and dynamically allocates human resources at gas stations;

[0112] The interaction module is used to generate a personalized service list for users on the local life service platform using collaborative filtering and reinforcement learning algorithms, and to interact with users in the interactive interface based on the personalized service list to provide users with corresponding local life service resources.

[0113] In some implementations, the local life service integration device can be applied to terminal equipment.

[0114] It should be noted that, for the sake of convenience and brevity, the specific working process of the local life service integration device described above can be referred to the corresponding process in the aforementioned local life service integration method embodiment, and will not be repeated here.

[0115] 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.

[0116] like Figure 3As 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 (ASICs), field-programmable gate arrays (FPGAs), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. Specifically, the memory 302 can be a Flash chip, a read-only memory disk, an optical disc, a USB flash drive, or a portable hard drive, etc.

[0117] 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 local life service integration methods 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 local life service integration method, and will not be repeated here.

[0118] 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 local life service integration methods provided in the specification of this application.

Claims

1. A method for integrating local life services, characterized in that, The method includes: By establishing standardized interface protocols, unified data exchange formats, and resource update strategies, multiple heterogeneous local life service providers are integrated into the gas station service platform to form a local life service platform, which collects real-time service data from each local life service provider. Based on the real-time service data, the fluctuation of non-oil product sales at gas stations is predicted, and the ordering volume and delivery cycle of upstream and downstream suppliers are dynamically adjusted according to the prediction results. A multi-dimensional data analysis model for gas station scenarios is constructed, which integrates historical user traffic data of gas stations, historical demand data of different local life services, surrounding traffic flow, weather change data, and community activities to predict user traffic and demand for various services at different times of the day. Based on the prediction results, corresponding local life service time windows are configured to dynamically optimize the allocation of local life service resources at gas stations at different times. Based on the order volume and delivery cycle of upstream and downstream suppliers, as well as the local life service resource allocation in different time periods, the local life service resources that can be called in the interactive interface of the local life service platform are dynamically updated, and the human resources of gas stations are dynamically allocated. For users on the local life service platform, a personalized service list is generated using collaborative filtering and reinforcement learning algorithms. Based on the personalized service list, the platform interacts with the user in the interactive interface to provide the user with corresponding local life service resources. Specifically, for users on the local life service platform, a collaborative filtering algorithm is used to dynamically reconstruct user profiles in real time, and a reinforcement learning algorithm is used to dynamically reconstruct service capability profiles in real time. Based on the user profiles and service capability profiles, combined with the user's current location, time, and historical preferences, a personalized service list is generated for the user. A service matching strategy for the user is generated by combining the user's historical consumption data, the attributes of the currently needed services, the inventory of nearby service points, and the service time window. Based on the personalized service list and the service matching strategy, corresponding personalized push information is displayed in the interactive interface of the local life service platform to interact with the user and provide the user with corresponding local life service resources. The real-time dynamic reconstruction of user profiles using collaborative filtering algorithms includes: semantic extraction based on user registration data to obtain basic user attribute features; analysis of first-time browsing behavior in user history to obtain initial browsing behavior features; migration of initial influx vectors from similar user groups using meta-learning, fusing basic user attribute features and initial browsing behavior features to obtain the initial user profile of new users; extraction of user historical consumption data, mapping product ID, amount, and category in the user historical consumption data to low-dimensional vectors using the TransE algorithm, normalizing according to consumption timestamps to generate time-series vectors containing time decay factors, serving as the user's consumption time-series vector; and construction of a user-product interaction graph using a graph convolutional network (GCN) to represent the user and... Products are used as nodes, and user real-time behavior and user historical behavior form connecting edges. The attention weights of key product nodes in the user behavior path are calculated using the graph attention mechanism (GAT) to generate the user's behavior feature vector. User behavior events are captured in real time using the Flink stream processing engine, and the high-dimensional sparse features in the user behavior events are processed using the Neural Collaborative Filtering (NCF) model. A multimodal input feature is constructed, which includes the consumption time-series vector, the behavior feature vector, and the high-dimensional sparse features. The nonlinear interaction features of the multimodal input features are learned through a multilayer perceptron, and a user preference embedding vector is output. The user preference embedding vector is then updated in the initial user profile or the user profile updated in the previous reconstruction, realizing real-time reconstruction of the user profile.

2. The method according to claim 1, characterized in that, Based on the personalized service list and the service matching strategy, the corresponding personalized push information is displayed in the interactive interface of the local life service platform. After interacting with the user, the process further includes: Detect user switching behavior between mobile devices and in-vehicle systems; If a switching behavior is detected, interactive data integration and analysis are initiated simultaneously. The interactive interface is then rearranged to be compatible with the device after the switch, and the interactive information from before the switch is synchronously displayed on the device after the switch through the rearranged interactive interface, so as to achieve integrated interaction between the mobile terminal and the vehicle system.

3. The method according to claim 1, characterized in that, The method of dynamically reconstructing service capability profiles in real time using reinforcement learning algorithms includes: Local life service providers are defined as independent intelligent agents, and local life service platforms are defined as coordinating intelligent agents. We collect multi-dimensional state data from local life service providers and use reinforcement learning to construct the service provider state space for each agent, thus obtaining an initial service capability profile. The multi-dimensional state data includes: service quality, response speed, user reviews, and service load. Using a graph attention network, spatial relationships between different agents are established based on the geographical location of local life service providers. An environmental state transition model enables different agents to perceive the state of surrounding agents. By extracting graph structure features, collaborative strategy vectors between different agents are obtained. Based on the historical performance of local life service providers, the service evaluation weights of the coordinating intelligent agent are dynamically adjusted through a credit allocation mechanism. By adopting the adjusted collaborative strategy vector and service evaluation weights, and according to the flexible evaluation cycle associated with activity hotspots, the service provider interaction network formed by independent intelligent agents and coordinating intelligent agents is iteratively updated, and the service capability profiles corresponding to each independent intelligent agent are reconstructed.

4. The method according to claim 1, characterized in that, Based on the user profile and service capability profile, and combined with the user's current location, time, and historical preferences, a personalized service list is generated for the user, including: Based on the user profile, combined with the user's current location, time, and historical preferences, the user's real-time service needs are predicted, and a personalized service candidate set is generated based on the prediction results. Based on the service capability profile, the fulfillment status score, service quality score, response speed, and load status of each local life service provider in the personalized service candidate set are identified as the service status set. Target service providers whose service status sets meet the service provider capability assessment conditions are selected, and the personalized service list is constructed using the target service providers.

5. The method according to claim 1, characterized in that, The method of predicting fluctuations in non-oil product sales at gas stations based on the real-time service data, and dynamically adjusting the order quantities and delivery cycles of upstream and downstream suppliers according to the prediction results, includes: Using historical non-oil product sales data, along with associated geographical locations, surrounding business ecosystems, unforeseen events, seasonal data, and weather data, the optimal combination weights are determined through Bayesian optimization, and a three-dimensional dynamic decision-making model is constructed that includes sales forecasting, inventory status, and delivery capabilities. Establish a virtual mirror of the supply chain corresponding to the real-time service data; Using the aforementioned three-dimensional dynamic decision-making model, multiple ordering strategies for the virtual mirror matching of the supply chain are adaptively constructed. Dynamic simulations are performed on each ordering strategy to obtain dynamic simulation results under each ordering strategy. The dynamic simulation results include: sales fluctuation forecast, inventory fluctuation forecast, and delivery load forecast. By introducing the Copula function as the connection function and constructing a non-parametric joint distribution model, the tail correlation between sales fluctuations and supply delays in the virtual supply chain is identified, and the tail risk characteristics under various ordering strategies are obtained. Based on the tail risk characteristics, a target ordering strategy is selected to achieve a safe inventory level by screening inventory fluctuations, and the target ordering strategy is executed to balance the order volume and delivery cycle of upstream and downstream suppliers.

6. The method according to claim 1, characterized in that, After dynamically updating the local life service resources that can be accessed through the interactive interface of the local life service platform and dynamically allocating human resources at gas stations, the system also includes: The system checks whether the local service providers on the local service platform have reached the set service resource volume. If the set service resource volume has not been reached, it selects alternative local service providers from the candidate service provider pool according to preset rules and reconnects the alternative local service providers to the local service platform. Detect the trend of user traffic changes. If a sudden increase in user traffic is detected, analyze historical user data to find the matching cause of the traffic surge, and predict the areas of service demand growth based on real-time user traffic. Based on the reasons for the surge in traffic and the areas of service demand growth, the local life service platform will dynamically expand the matching local life service providers and the types of local life services that can be provided.

7. A local life service integrated device, characterized in that, The device includes the following modules: The data acquisition module is used to integrate multiple heterogeneous local life service providers into the gas station service platform by establishing standardized interface protocols, unified data exchange formats, and resource update strategies to obtain the local life service platform, and to collect real-time service data from each local life service provider. The configuration module is used to predict fluctuations in non-oil product sales at gas stations based on the real-time service data, and dynamically adjust the order volume and delivery cycle of upstream and downstream suppliers according to the prediction results; it constructs a multi-dimensional data analysis model for gas station scenarios, integrates historical user traffic data of gas stations, historical demand data of different local life services, surrounding traffic flow, weather change data, and community activities, predicts user traffic and demand for various services at different times of the day, and configures corresponding local life service time windows based on the prediction results to dynamically optimize the allocation of local life service resources at gas stations at different times; based on the order volume and delivery cycle of upstream and downstream suppliers, as well as the allocation of local life service resources at different times, it dynamically updates the local life service resources that can be called by the interactive interface in the local life service platform, and dynamically allocates human resources at gas stations; The interaction module is used to generate a personalized service list for users on the local life service platform using collaborative filtering and reinforcement learning algorithms, and to interact with users in the interactive interface based on the personalized service list to provide users with corresponding local life service resources. Specifically, the interaction module is used to: dynamically reconstruct user profiles in real time using collaborative filtering algorithms and service capability profiles in real time using reinforcement learning algorithms for users on the local life service platform; and generate a personalized service list for users based on the user profiles and service capability profiles, combined with the user's current location, time, and historical preferences; generate a service matching strategy for users by combining the user's historical consumption data, current required business attributes, surrounding service point inventory, and service time windows; and display corresponding personalized push information in the interactive interface of the local life service platform based on the personalized service list and the service matching strategy, interact with users, and provide users with corresponding local life service resources. The interaction module utilizes a collaborative filtering algorithm to dynamically reconstruct user profiles in real time. Specifically, it performs semantic extraction based on user registration data to obtain basic user attribute features; analyzes the user's first browsing behavior in historical behavior to obtain initial browsing behavior features; uses meta-learning to migrate initial inflow vectors from similar user groups, fusing basic user attribute features and initial browsing behavior features to obtain the initial user profile of new users; extracts user historical consumption data, maps the product ID, amount, and category in the user's historical consumption data into low-dimensional vectors using the TransE algorithm, normalizes them according to consumption timestamps, and generates a time-series vector containing a time decay factor as the user's consumption time-series vector; and uses a graph convolutional network (GCN) to construct a user-product interaction graph. Users and products are used as nodes, and real-time user behavior and historical user behavior are used as connecting edges. The attention weights of key product nodes in the user behavior path are calculated using the graph attention mechanism (GAT) to generate the user's behavior feature vector. User behavior events are captured in real time using the Flink stream processing engine, and the high-dimensional sparse features in the user behavior events are processed using the Neural Collaborative Filtering (NCF) model. A multimodal input feature is constructed, which includes the consumption time-series vector, the behavior feature vector, and the high-dimensional sparse features. The nonlinear interaction features of the multimodal input feature are learned through a multilayer perceptron, and the user preference embedding vector is output. The user preference embedding vector is updated to the initial user profile or the user profile updated in the previous reconstruction, realizing the real-time reconstruction of the user profile.

8. A terminal device, characterized in that, The terminal device includes a processor and a memory; The memory is used to store computer programs; The processor is used to execute the computer program and, in executing the computer program, implement the local life service integration method as described in any one of claims 1 to 6.

Images