Intelligent scheduling method, device and equipment of service resource and storage medium

Abstract

The application discloses a kind of intelligent scheduling methods, devices and equipment of service resource and storage medium, wherein the method comprises: receiving the service request submitted by user, and the service request at least includes service type, expected service time and special demand description;Dynamic demand urgency score of the service request is calculated based on multi-modal fusion model, including: extracting the multi-dimensional features associated with the service request, the multi-dimensional features at least include user portrait features, real-time environment features and order semantic features obtained by semantic analysis based on the special demand description;The multi-dimensional features are input into the pre-trained urgency evaluation model, and the quantitative dynamic demand urgency score is output;Real-time resource state of service area and future resource state predicted based on historical data are obtained, to form enhanced resource state view;Global scheduling decision is made based on reinforcement learning agent, to generate final scheduling sequence.

Description

Technical Field

[0001] This application relates to, but is not limited to, the field of software service technology, and in particular to an intelligent scheduling method, apparatus, device, and storage medium for service resources. Background Technology

[0002] With the rapid development of internet technology and the widespread adoption of mobile applications, home service platforms have become a common channel for users to conveniently book cleaning, repair, and other services. These platforms connect users and service providers, achieving initial resource integration and allocation, and improving the accessibility and efficiency of home services. However, with the expansion of the user base and the diversification of service scenarios, existing home service platforms have gradually revealed many limitations in resource allocation, making it difficult to meet increasingly complex user needs and dynamically changing resource conditions.

[0003] Existing technologies typically employ a "first-come, first-served" scheduling approach or a scheduling method based on fixed priority categories (such as user levels). This approach lacks the ability to dynamically assess the urgency of users' actual needs. For example, emergency leak repairs and routine home cleaning have inherently different levels of urgency, but existing systems often fail to effectively distinguish between them, resulting in high-urgency needs not receiving priority responses, impacting user experience and even causing property damage. Summary of the Invention

[0004] In view of this, embodiments of this application provide at least one intelligent scheduling method, apparatus, device, and storage medium for service resources.

[0005] The technical solution of this application embodiment is implemented as follows: On one hand, embodiments of this application provide an intelligent scheduling method for service resources, the method comprising: Receive service requests submitted by users, the service requests including at least the service type, expected service time and special requirements description; The dynamic urgency score of the service request is calculated based on a multimodal fusion model, including: extracting multi-dimensional features associated with the service request, wherein the multi-dimensional features include at least user profile features, real-time environment features, and order semantic features obtained by semantic analysis based on the special requirement description; inputting the multi-dimensional features into a pre-trained urgency assessment model, and outputting a quantified dynamic urgency score. It acquires the real-time resource status of the service area and the future resource status predicted based on historical data, forming an enhanced resource status view; Global scheduling decisions are made based on a reinforcement learning agent to generate a final scheduling sequence, wherein: the reinforcement learning agent uses the dynamic demand urgency score and the enhanced resource state view as state inputs; the reinforcement learning agent outputs scheduling decision actions according to its policy network, the scheduling decision actions include order allocation, order delay, or cross-regional deployment of service personnel.

[0006] On the other hand, embodiments of this application provide an intelligent scheduling device for service resources, the device comprising: The receiving module is used to receive service requests submitted by users, and the service requests include at least the service type, expected service time and special requirements description. The processing module is used to calculate the dynamic urgency score of the service request based on a multimodal fusion model, including: extracting multi-dimensional features associated with the service request, wherein the multi-dimensional features include at least user profile features, real-time environment features, and order semantic features obtained by semantic analysis based on the special requirement description; inputting the multi-dimensional features into a pre-trained urgency assessment model and outputting a quantified dynamic urgency score; and obtaining the real-time resource status of the service area and the future resource status predicted based on historical data to form an enhanced resource status view. The output module is used to make global scheduling decisions based on the reinforcement learning agent and generate a final scheduling sequence, wherein: the reinforcement learning agent uses the dynamic demand urgency score and the enhanced resource state view as state inputs; the reinforcement learning agent outputs scheduling decision actions according to its policy network, and the scheduling decision actions include order allocation, order delay, or cross-regional deployment of service personnel.

[0007] In another aspect, embodiments of this application provide a computer device, including a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the program, it implements some or all of the steps in the above-mentioned intelligent scheduling method for service resources.

[0008] In another aspect, embodiments of this application provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements some or all of the steps in the above-described intelligent scheduling method for service resources.

[0009] In another aspect, embodiments of this application provide a computer program including computer-readable code. When the computer-readable code is run in a computer device, the processor in the computer device executes some or all of the steps in the intelligent scheduling method for implementing the above-mentioned service resources.

[0010] In another aspect, embodiments of this application provide a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program. When the computer program is read and executed by a computer, it implements some or all of the steps in the above-mentioned intelligent scheduling method for service resources.

[0011] This application's embodiments accurately quantify the urgency of needs by integrating multi-dimensional features and combining real-time and predictive resource information, utilizing reinforcement learning for global optimization decisions. Ultimately, the system can intelligently prioritize high-urgency needs when resources are limited, while improving overall resource utilization efficiency and service response speed through proactive resource allocation and order scheduling. This achieves a balance and improvement in user satisfaction and platform operational efficiency in complex and ever-changing domestic service scenarios.

[0012] It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and are not intended to limit the technical solutions of this application. Attached Figure Description

[0013] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with this application and, together with the specification, serve to explain the technical solutions of this application.

[0014] Figure 1 A schematic diagram illustrating the implementation process of an intelligent scheduling method for service resources provided in this application embodiment; Figure 2 This application provides an overall architecture diagram of a service prioritization and resource allocation system based on multi-factor dynamic evaluation in a domestic service scenario. Figure 3 A schematic diagram of the structure of a multi-dimensional urgency assessment engine provided in an embodiment of this application; Figure 4 A flowchart illustrating the workflow of a dynamic priority scheduling decision engine provided in this application embodiment; Figure 5 A flowchart illustrating an urgency score calculation method provided in this application embodiment; Figure 6 A schematic diagram of a priority adjustment and compensation strategy provided in an embodiment of this application; Figure 7 A timing diagram of the scheduling process of a system in an emergency pipeline leakage scenario provided in an embodiment of this application; Figure 8 An example diagram illustrating the application of resource scheduling in a system during peak holiday periods, provided in this application embodiment; Figure 9 A schematic diagram illustrating the composition of another intelligent scheduling device for service resources provided in this application embodiment; Figure 10 This is a schematic diagram of the hardware entity of a computer device provided in an embodiment of this application. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application are further described in detail below with reference to the accompanying drawings and embodiments. The described embodiments should not be regarded as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0016] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0017] The terms “first / second / third” are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that “first / second / third” may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used herein is for descriptive purposes only and is not intended to limit the scope of this application.

[0019] This application provides an intelligent scheduling method for service resources, which can be executed by the processor of a computer device. The computer device can refer to a server, laptop computer, tablet computer, desktop computer, smart TV, set-top box, mobile device (such as a mobile phone, portable video player, personal digital assistant, dedicated messaging device, portable gaming device), or other similar computer equipment. Figure 1 This application provides a schematic diagram illustrating the implementation process of an intelligent scheduling method for service resources, as shown in the embodiments below. Figure 1 As shown, the method includes: Step 101: Receive a service request submitted by the user. The service request shall include at least the service type, expected service time and special requirements.

[0020] In this embodiment, a service request refers to a digital instruction submitted by a user through a home service platform, requesting specific home services. Service type refers to the specific category of home services, such as pipe repair, routine cleaning, and appliance cleaning. Expected service time refers to the specific time or time period specified by the user when submitting the request, indicating their desired service delivery. Special needs description refers to additional text information filled in by the user in the service request, describing their special circumstances or personalized requirements.

[0021] The system receives home service requests submitted by users via mobile applications or web pages through its user interface module. The system parses the request data packet, extracting the structured information contained within, including at least the service type selected by the user, the user's desired service time, and any special requirements specified by the user in text form. This information serves as the foundational input data for subsequent needs analysis and decision-making.

[0022] Step 102, calculate the dynamic demand urgency score of the service request based on the multimodal fusion model, including: extracting multi-dimensional features associated with the service request, wherein the multi-dimensional features include at least user profile features, real-time environment features, and order semantic features obtained by semantic analysis based on the special demand description; inputting the multi-dimensional features into a pre-trained urgency assessment model, and outputting a quantified dynamic demand urgency score.

[0023] In this embodiment, the multimodal fusion model refers to a computational model capable of processing and integrating data from different sources and in different formats. User profile features refer to a set of labeled features constructed based on user historical behavior data (such as order frequency, spending amount, review records, and cancellation rate) and registration information, used to describe user attributes and behavioral patterns. Real-time environmental features refer to the immediacy features related to time and location when calculating urgency, such as the current time, the user's location, and the historical order density of the current time period. Order semantic features refer to key semantic information reflecting the urgency of the need, extracted through natural language processing analysis of the "special needs description" text, such as whether the text contains keywords like "leaking water," "emergency," and "elderly person alone at home," along with their emotional tendencies. The pre-trained urgency assessment model refers to a machine learning model trained using historical order data and its corresponding results, capable of learning the complex mapping relationship between multi-dimensional features and the degree of urgency.

[0024] The system's multi-dimensional urgency assessment engine performs this step. First, the engine extracts relevant multi-dimensional features from the service request information received in step 101. This includes querying and calculating the user's profile features from the user database; obtaining current real-time environmental features; and performing semantic analysis on the special requirement description text to extract order semantic features. Then, the engine feeds these extracted multi-dimensional feature vectors as input to a pre-trained urgency assessment model (e.g., a neural network model). This model comprehensively calculates and fuses these features, ultimately outputting a quantified numerical value, namely the dynamic urgency score of the service request.

[0025] Step 103: Obtain the real-time resource status of the service area and the future resource status predicted based on historical data to form an enhanced resource status view.

[0026] In this embodiment, the service area refers to the specific geographical range set by the housekeeping service platform for resource scheduling and services. Real-time resource status refers to the instantaneous status information of all available service personnel within the service area at a given moment, including their geographical location, current task status, and skill tags. Future resource status refers to the prediction results of resource demand and supply in the service area over a future period based on data such as historical order patterns, holiday information, and weather forecasts. The enhanced resource status view refers to a more comprehensive and forward-looking resource status data view that integrates real-time status with future predictions.

[0027] The system's resource status monitoring module performs this step. The module uses the platform's location system and service personnel terminals to collect real-time dynamic information from all service personnel within the service area, creating a real-time resource status snapshot. Simultaneously, the module accesses the historical database and uses predictive algorithms (such as time series analysis) to forecast resource demand and availability for specific future time periods (such as the next hour or the next day), generating future resource status. Finally, the module integrates and correlates the real-time resource status with the future resource status predicted based on historical data, constructing an enhanced resource status view that not only reflects the current situation but also includes judgments on future trends.

[0028] Step 104: Make global scheduling decisions based on the reinforcement learning agent to generate the final scheduling sequence, wherein: the reinforcement learning agent uses the dynamic demand urgency score and the enhanced resource state view as state inputs; the reinforcement learning agent outputs scheduling decision actions according to its policy network, the scheduling decision actions include order allocation, order delay, or cross-regional deployment of service personnel.

[0029] In this embodiment, a reinforcement learning agent refers to an intelligent decision-making program that learns through interaction with its environment to maximize long-term returns. Global scheduling decision refers to a holistic and optimized scheduling arrangement made based on all pending orders and all available resources. The final scheduling sequence is an ordered list of the execution order of all pending orders and service personnel allocation schemes, determined after intelligent decision-making. State input refers to the set of all information describing the current decision-making environment within the reinforcement learning framework, serving as the basis for the agent's decisions. The policy network is a neural network within the reinforcement learning agent that directly outputs the optimal action to be taken based on the current state input. Scheduling decision actions refer to specific operational instructions that the agent can execute, such as assigning an order to a specific service personnel, delaying the service time of an order, or reassigning a service personnel from one service area to another.

[0030] The system's dynamic priority scheduling decision engine executes this step. The engine combines the calculated dynamic demand urgency score and the resulting enhanced resource state view as a state description of the current scheduling environment, inputting this into a reinforcement learning agent. The agent's internal policy network calculates based on the input state information and outputs a specific scheduling decision action. This action might be assigning the most skilled and nearest service personnel to a high-urgency order, appropriately delaying a low-urgency order to free up resources, or cross-regional deployment of service personnel to balance workloads across different regions. By repeatedly simulating and evaluating the long-term effects of different action combinations (such as overall user satisfaction and resource utilization), the agent ultimately generates a globally optimal final scheduling sequence.

[0031] This application's embodiments accurately quantify the urgency of needs by integrating multi-dimensional features and combining real-time and predictive resource information, utilizing reinforcement learning for global optimization decisions. Ultimately, the system can intelligently prioritize high-urgency needs when resources are limited, while improving overall resource utilization efficiency and service response speed through proactive resource allocation and order scheduling. This achieves a balance and improvement in user satisfaction and platform operational efficiency in complex and ever-changing domestic service scenarios.

[0032] Optionally, the user profile features include at least one of historical order frequency, user category tags, and historical complaint / cancellation rates; the real-time environmental features include at least one of weather conditions, regional events, and time segments; and the order semantic features include information extracted from the special requirements description through keyword matching or sentiment analysis.

[0033] In this application embodiment, user profile features refer to a set of structured tags that can describe user attributes, preferences and behaviors by analyzing the user's past behavior data on the platform.

[0034] Historical order frequency refers to the average or total number of times a user submits a housekeeping service order within a specific time period (such as the past month or quarter). User category tags are classification labels assigned by the system based on user behavior patterns (such as spending level, service preferences, and active time periods), such as "high-frequency household user," "corporate customer," and "new user." Historical complaint / cancellation rate refers to the ratio of the number of orders with complaints or orders that were actively cancelled by a user in the past to their total number of orders.

[0035] Real-time environmental characteristics refer to real-time factors in the external objective environment that may affect service scheduling or urgency assessment when a service request occurs. Weather conditions refer to the meteorological conditions in the service area at the time the service request occurs and during the expected service period, such as heavy rain, heavy snow, high temperatures, typhoons, etc. Regional events refer to specific activities or conditions occurring within the service area that may affect traffic, resource availability, or service demand, such as large gatherings, traffic control, regional power outages, or water outages. Time segments refer to dividing a 24-hour day into time periods with different levels of resource scarcity or service demand characteristics, such as weekday morning rush hour, holidays, and late-night hours.

[0036] Order semantic features refer to the semantic information extracted from unstructured text information (special request descriptions) submitted by users, reflecting the potential urgency or special circumstances of the situation. Keyword matching involves searching for and identifying keywords in the text using a pre-defined emergency keyword library (such as "water leak," "power outage," "elderly person alone at home," "immediately," "as soon as possible") to determine the urgency level expressed in the text. Sentiment analysis uses natural language processing technology to analyze the emotional tone and intensity expressed in the text, such as determining whether the tone is anxious, calm, or complaining, thereby indirectly assessing the urgency of the need. Special request descriptions are additional text descriptions filled in by users when submitting service orders, beyond the fixed options, to supplement the explanation of the service scenario, special requirements, or urgency.

[0037] The system retrieves historical user data associated with the current service request from the user database. It analyzes the user's historical order records to calculate their order frequency within a certain period. Simultaneously, based on preset classification rules or machine learning models, the system matches or updates one or more user category tags for the user. Furthermore, the system retrieves the user's historical order status, counts the number of complained orders and canceled orders, and calculates the complaint rate and cancellation rate. Finally, the system integrates these calculated and matched results (one or more of historical order frequency, user category tags, and historical complaint / cancellation rates) to form a user profile feature vector for subsequent calculations.

[0038] The system obtains real-time environmental information related to the current service request by accessing external data interfaces (such as weather APIs and traffic event APIs) or by using the built-in calendar and clock. The system queries real-time weather data for the service address's area and obtains weather forecasts for the next period. The system checks for reported or news-related regional events within the service area. Simultaneously, based on the current date and time, the system determines the predefined time segment to which the event belongs (e.g., "weekend afternoon" or "weekday commuting hours"). The system uses this acquired information (one or more of weather conditions, regional events, and time segments) as input for real-time environmental features.

[0039] The system receives the "Special Request Description" text field from a user's service request. First, the system uses keyword matching technology to compare the text with a pre-defined urgency keyword database, identifying and recording all appearing keywords and their frequencies. Simultaneously, the system uses a sentiment analysis model to process the text, outputting its sentiment polarity (positive, negative, neutral) and sentiment intensity score. The system then merges the keyword matching results (such as a list of matched keywords and their weights) with the sentiment analysis results (such as the sentiment polarity score) to generate structured order semantic features, used to quantify the urgency information implicit in the user's text.

[0040] This application's embodiments automatically extract structured, quantifiable key evaluation features from raw, unstructured user requests and the environment. These features serve as core inputs, which are then used by a "multi-dimensional urgency assessment engine" to calculate a comprehensive "urgency score." This process achieves a refined, multi-dimensional, and dynamic assessment of the urgency of user needs, overcoming the limitations of existing technologies that rely on only a single dimension. It provides an accurate and comprehensive data foundation for subsequent intelligent dynamic scheduling and resource allocation based on actual urgency and resource status, ultimately improving resource utilization efficiency and user satisfaction.

[0041] Optionally, the pre-trained urgency assessment model is a gradient boosting decision tree model, which is trained based on the time interval between the push and acceptance of historical orders and user feedback data after the order is completed.

[0042] In this embodiment, the gradient boosting decision tree model is a machine learning model that iteratively improves the overall prediction accuracy by combining multiple weaker decision tree models. The time interval from order push to order acceptance in historical service records refers to the length of time elapsed from when a service request is pushed to the service personnel by the system to when the service recipient confirms acceptance. User feedback data after order completion refers to the user's evaluation of the service experience after the service is completed.

[0043] Before deployment, the system pre-trains a gradient boosting decision tree model. The computer uses historical order data accumulated on the platform as training samples. The core objective of the training process is to teach the model how to predict the urgency of orders based on their characteristics. The labeled data used for training comes from two sources: first, the time interval between the system pushing a historical order and the service recipient accepting it (shorter intervals generally indicate more urgent needs or higher matching efficiency); second, user feedback data after order completion, such as satisfaction ratings or sentiment analysis results of review texts, which indirectly reflect user satisfaction with service timeliness. Through algorithmic iteration, the system continuously adjusts the model's internal parameters, enabling the model to learn complex patterns related to the "push-acceptance time interval" and "user feedback" from the multidimensional features of orders (such as service type, time, and user context), thus forming a predictive model capable of accurately assessing the urgency of new orders.

[0044] This application employs a gradient boosting decision tree model trained based on historical order response times and user feedback as the core of urgency assessment. This system achieves more accurate and adaptive quantification of demand urgency than traditional rules. The model automatically learns the non-linear relationship between complex features and urgency from massive amounts of historical data, making the urgency assessment results more closely reflect actual business scenarios and user experiences. This provides a scientific and reliable data foundation for subsequent dynamic priority scheduling, thereby improving the overall rationality of resource allocation and user satisfaction.

[0045] Optionally, the enhanced resource status view is constructed in the following manner: Step 201: Divide the service area into a grid.

[0046] In this embodiment, the service area refers to the physical geographic range within which the housekeeping service platform provides services, such as a city or a district. Grid partitioning is a data processing method that divides a continuous geographic area into multiple independent, regular, and non-overlapping units (i.e., grids) according to certain rules (such as latitude and longitude, administrative boundaries, or fixed dimensions). Each grid serves as an independent geographic analysis unit.

[0047] The system first determines the geographical boundaries of its operational coverage, i.e., the service area. Then, the system invokes a geographic information processing algorithm to automatically divide the service area into a series of continuous, uniformly sized grid cells based on preset grid sizes (e.g., a 500m x 500m square) or geographic features (such as main roads or street blocks). This process transforms continuous spatial geographic information into discrete, indexable, and computable data structures, providing a basic geographic framework for subsequent location-based resource monitoring, aggregation, and predictive analysis. Each grid cell is assigned a unique identifier (ID) and associated with its geographic coordinate range.

[0048] Step 202: Aggregate the status, skills, and location information of service personnel within each grid in real time.

[0049] In this embodiment, aggregation refers to the system collecting data from multiple data sources and summarizing, merging, and statistically analyzing it according to a specific dimension (here, a grid). Service personnel status refers to their current work status, such as idle, working, or off-duty. Service personnel skills refer to their service capabilities, such as plumbing repair or deep cleaning. Service personnel location information refers to their real-time geographic coordinates reported via their mobile device's GPS module.

[0050] The system continuously receives and updates real-time data streams from each service worker via their mobile application, including their GPS coordinates, current task status (e.g., "Idle," "Service," "Offline"), and skill tag list. Based on an established grid-based geographic mapping, the system resolves each service worker's real-time coordinates to their specific grid ID. Then, the system performs real-time aggregation calculations on a grid-by-grid basis for all service workers falling within that grid. Aggregation operations include, but are not limited to: counting the number of service workers in an "Idle" state within the grid, listing the skill sets of all available personnel, and calculating the average location or density distribution of service workers. These aggregation results form a real-time resource snapshot for each grid.

[0051] Step 203: Based on historical order data and service personnel movement patterns, predict the resource demand trend of each grid within a future set time period.

[0052] In this embodiment, historical order data refers to the platform's accumulated records of completed or cancelled service orders, including order time, service location (grid), service type, and other information. Service personnel movement pattern refers to the regularity of service personnel's location changes over time during task execution, such as a trajectory pattern from home to a service location, and then to the next service location. Future time frame refers to a pre-configured time window, such as the next two hours. Resource demand trend refers to the predicted direction and magnitude of changes in the number of service orders or the number of service personnel required for each grid within a specific future time period.

[0053] The system invokes a machine learning prediction model. This model uses a historical order database and a service personnel historical trajectory database as training and input data. The model first analyzes historical data to identify patterns in order volume, service type distribution, and spatiotemporal patterns of service personnel response and movement across different grids on different dates and at different times (e.g., weekday morning rush hour, weekend afternoon). Combining this with contextual information such as current time, date type (weekday / holiday), and weather, the model calculates a predicted number of orders or service personnel needed for each grid within a preset time period (e.g., the next 2 hours). This prediction quantifies the future resource scarcity or abundance for each grid, i.e., the resource demand trend.

[0054] Step 204: Integrate the real-time status, predicted trends, and queue information of pending orders from each grid.

[0055] In this embodiment, integration refers to the system associating and merging multi-source, heterogeneous data from different modules or computational processes to form a unified and comprehensive data view to support subsequent decision analysis. The queue of orders to be assigned refers to the collection of all user service requests that have been submitted to the platform but have not yet been matched with service personnel and scheduled for a specific service time. Each order includes information such as the type of service requested, the expected time, and the user's location (grid to which they belong).

[0056] The system establishes a comprehensive data view. This view uses grids as the basic unit, linking and presenting three core pieces of information for each grid: first, the grid's real-time resource status (e.g., the number of available personnel and their skills); second, the grid's projected future resource demand trends; and third, a list of orders with target service addresses located within the grid, along with their detailed information (e.g., urgency and service type). The system aligns and correlates these three types of information using the grid ID as a key index. This integrated panoramic view clearly shows what resources each grid currently has, how many resources it might need in the future, and how many demands are currently waiting to be met, providing a complete data foundation for the dynamic priority scheduling decision engine to perform cross-regional, forward-looking resource optimization and allocation.

[0057] This application embodiment constructs a refined, real-time spatial resource situation map with predictive capabilities by sequentially performing gridded division of service areas, real-time aggregation of resources within the grid, prediction of future grid demand based on historical data, and final integration of multi-source information.

[0058] Optionally, the reinforcement learning agent employs proximal policy optimization or deep network algorithms, and the reward function of the reinforcement learning agent is configured to comprehensively reward multiple indicators, including allocating high-urgency orders, shortening service distances, reducing order cancellations, improving user ratings, and balancing service personnel income.

[0059] In this embodiment, a reinforcement learning agent is a computer program or algorithm model that learns how to take actions through interaction with the environment to maximize the cumulative reward obtained from the environment. Proximal policy optimization is a reinforcement learning algorithm that ensures the stability of learning by limiting the magnitude of each policy update. Deep network algorithms, typically referring to deep neural networks, are machine learning models that mimic the connection structure of neurons in the human brain. They are capable of handling complex nonlinear relationships and are often used as the policy network or value network in reinforcement learning agents.

[0060] The reward function is a core component of a reinforcement learning agent. It is a mathematical function used to quantify the immediate gain or penalty an agent receives after taking a certain action in a specific state. The system uses the reward function to signal to the agent whether its behavior is good or bad, and the agent's learning objective is to maximize the long-term cumulative reward.

[0061] The system deploys a reinforcement learning agent within its dynamic priority scheduling decision engine. This agent is responsible for selecting actions (such as order allocation or priority adjustment) based on the current state (e.g., order urgency, resource distribution). The system provides this agent with two optional algorithm implementations: a proximal policy optimization algorithm or a deep network-based algorithm. The system trains the agent repeatedly in simulated or real-world domestic service scheduling environments to learn the optimal scheduling strategy. When using the proximal policy optimization algorithm, the system ensures that the agent's policy updates are smooth in each iteration, avoiding drastic fluctuations in policy performance. When using the deep network algorithm, the system leverages the powerful function fitting capabilities of deep neural networks to approximate complex policy functions or state-value functions, thereby handling high-dimensional state and action spaces.

[0062] The system designs and configures a composite reward function for the reinforcement learning agent. This function is not single-objective but integrates multiple key business metrics to form a comprehensive evaluation standard. When the agent makes a scheduling decision, the system immediately calculates the immediate reward. The calculation process is as follows: The system evaluates whether the decision successfully assigns high-urgency orders to suitable service personnel; if so, a positive reward is given. The system calculates the change in the distance service personnel travel to the service location due to the decision; a shorter distance results in a positive reward. The system monitors whether the decision reduces the probability of orders being canceled by users or service personnel; a decrease results in a positive reward. The system predicts the potential post-service ratings from the decision; an improvement in ratings results in a positive reward. The system also evaluates the impact of the decision on the income balance among service personnel; promoting income balance results in a positive reward. The system weights and sums the reward values ​​of these multiple metrics according to preset weights, ultimately deriving a comprehensive reward value, which is fed back to the reinforcement learning agent to guide subsequent strategy updates and optimizations.

[0063] This application embodiment enables a reinforcement learning agent, guided by a composite reward function, to autonomously learn and generate scheduling strategies. These strategies not only ensure that high-urgency domestic service needs are prioritized, but also take into account service efficiency (shortening distance), service stability (reducing cancellations), user experience (improving ratings), and ecosystem health (balancing income). This surpasses simple rule-based scheduling and achieves globally optimal resource allocation under multi-objective collaboration, thereby significantly improving the platform's overall operational efficiency, resource utilization, and user satisfaction in complex and dynamic domestic service scenarios.

[0064] Optionally, the method further includes: Step 310: Execute the scheduling decision action.

[0065] In this embodiment, the scheduling decision action refers to the final set of instructions generated by the dynamic priority scheduling decision engine regarding how to allocate domestic service orders to specific service personnel and when to execute them. This action is the optimal resource allocation scheme derived by the system after multi-dimensional urgency assessment, resource status analysis, and priority ranking.

[0066] The system, through its service execution and feedback module, transforms the optimal scheduling scheme generated by the dynamic priority scheduling decision engine into specific, executable operation instructions. These instructions are simultaneously pushed to relevant user terminals and service personnel terminals. For user terminals, the system pushes a notification containing the confirmed service time, service personnel information, and service details. For service personnel terminals, the system pushes task instructions containing detailed information about the order to be executed, the user's address, service requirements, and the system's suggested route and time. The system ensures the accurate delivery of instructions and initiates a tracking process for order execution status.

[0067] Step 320: Collect the execution result data of the scheduling decision action, the execution result data including at least the actual order response time and user rating.

[0068] In this embodiment, the actual order response time refers to the time interval from when the system issues a dispatch instruction to the service personnel to when the service personnel actually arrive at the user's designated service location and begin service. User rating refers to the numerical value by which a user quantitatively evaluates the overall experience of the service through a terminal application after the service is completed.

[0069] The system continuously monitors and records order lifecycle data through its service execution and feedback module. It automatically collects the check-in time and service start time reported by service personnel from their terminals, and calculates the actual order response time by combining this data with the dispatch instruction issuance time. Simultaneously, after service completion, the system pushes an evaluation request to the user's terminal through the user interface module and receives quantitative ratings and possible text feedback from the user. The system stores these collected execution result data, including response time, user ratings, and other relevant process data (such as on-time performance and service completion rate), in a structured manner to form data records for subsequent analysis.

[0070] Step 330: Use the execution result data as training samples to perform online learning and optimization on the policy network of the reinforcement learning agent.

[0071] In the embodiments of this application, a reinforcement learning agent is a computer program entity that can learn and improve its decision-making strategy by interacting with the environment and based on the reward or punishment signals it receives. The policy network is a mathematical model (usually a neural network) inside the reinforcement learning agent, which is responsible for outputting an action (i.e., a scheduling decision) based on the current environmental state (such as order urgency, resource status).

[0072] The system transforms collected execution result data (such as response time and user ratings) into "reward" or "penalty" signals for the previous scheduling decision. For example, shorter response times and higher user ratings generate positive reward signals. This signal, along with the environmental state at the time the scheduling decision was made (order urgency, resource availability, etc.), is used as a new set of training samples and fed into the reinforcement learning agent's learning algorithm. The algorithm uses these samples to update the internal parameters of the policy network online, enabling the policy network to make scheduling decisions that yield higher rewards (i.e., better user experience and resource efficiency) when facing similar environmental states in the future, thus achieving adaptive optimization of the scheduling strategy.

[0073] This embodiment of the application forms a complete "decision-execution-learning" closed loop by sequentially executing scheduling decisions, collecting execution feedback, and using the feedback data to optimize the decision model online. This closed loop enables the housekeeping service platform to transcend static, rule-based scheduling methods and transform into an intelligent system with self-learning and continuous optimization capabilities. The system can automatically adjust its priority evaluation weights and resource allocation strategies based on historical scheduling results, thereby continuously improving its response speed to high-urgency needs, the rationality of resource utilization, and overall user satisfaction in long-term operation, achieving a dynamic balance and continuous evolution of service scheduling efficiency and quality.

[0074] Optionally, when the scheduling decision action is an order delay, the method further includes: Step 410: Based on the dynamic urgency score of the delayed order, the delay duration, and the user's historical value, generate a personalized intelligent compensation plan.

[0075] In this embodiment, a delayed order refers to a user service request whose originally expected service time or processing order is postponed during dynamic priority scheduling due to limited system resources or the insertion of higher-priority orders. The dynamic demand urgency score is a numerical value calculated by a multi-dimensional urgency assessment engine to quantify the urgency of a user's current service request. This score integrates multi-dimensional characteristics such as time, service type, user context, and historical behavior. The delay duration refers to the time difference between the original expected service time (or the originally planned allocation time) and the new expected service time after system adjustment. User historical value is an indicator calculated based on a user's past consumption behavior data on the platform (such as cumulative consumption amount, order frequency, historical reviews, user level, etc.), used to measure the user's long-term value contribution to the platform. The intelligent compensation scheme refers to a combination of specific compensation measures automatically generated by the system to compensate for the inconvenience caused to users by service delays, such as service discounts, coupons, points rewards, or free service upgrades.

[0076] The compensation mechanism and balancing subsystem receives input from the dynamic priority scheduling decision engine, which includes a list of orders marked as delayed and their associated information. The system first extracts the dynamic urgency score, calculated delay duration, and the user's historical value data for each delayed order in the list. Then, based on a pre-defined compensation strategy model, the system uses these three core parameters as input variables. This model calculates the optimal compensation level and form by analyzing the correlation between delay duration and user dissatisfaction, the urgency score reflecting the current need, and the user's historical value representing the importance of user retention. For example, for an order with a long delay, a medium urgency score, but a high user historical value, the system might generate a combined compensation scheme including a higher discount and extra points. Finally, the system generates a personalized, structured, intelligent compensation scheme data object for each delayed order.

[0077] Step 420: Send the intelligent compensation scheme to the user terminal.

[0078] In this embodiment, the service execution and feedback module obtains the generated intelligent compensation scheme data packet from the compensation mechanism and balancing subsystem. This data packet is bound to a specific delayed order ID. The system uses a push notification service interface to send a notification message containing the specific content of the compensation scheme (such as compensation type, amount, usage conditions, etc.) and associated order information to the user's client of the order. The sending method can be online push notification, in-app message, or SMS. The system also records the status of the compensation scheme being sent, the timestamp, and the target user's device identifier to ensure message reachability and traceability.

[0079] Step 430: Quantify the cost corresponding to the intelligent compensation scheme into a negative reward signal and feed it back to the reinforcement learning agent.

[0080] In this embodiment, cost refers to the direct or indirect economic expenditure or resource consumption of the platform due to providing compensation solutions to users, such as the amount of discounts, the face value of coupons, and the potential redemption value of bonus points. In a reinforcement learning framework, a negative reward signal refers to a negative reward given to an agent by the environment after the agent performs an action (such as delaying an order). This reward indicates that the action has produced adverse consequences or costs, thereby guiding the agent to reduce such actions in future decisions. A reinforcement learning agent is a software entity that uses reinforcement learning algorithms. In this system, it specifically refers to the intelligent model within the dynamic priority scheduling decision engine responsible for learning and making the final scheduling decision. It optimizes its decision-making strategy based on the reward or penalty signals it receives by interacting with the environment (i.e., order and resource status).

[0081] After generating and sending a compensation plan, the compensation mechanism and balancing subsystem immediately performs a cost quantification assessment. Based on the specific content of the compensation plan, the system calculates its corresponding economic or resource cost. Then, the system converts this cost value into a negative reward signal. For example, if the compensation cost is 20 yuan, a reward signal of -20 is generated. Subsequently, the system constructs a feedback data packet containing the original scheduling decision context information that led to this compensation (such as the resource status and order queue status at the time), details of the executed delay action, and the calculated negative reward signal. This feedback data packet is sent to the reinforcement learning agent in the dynamic priority scheduling decision engine. Upon receiving the feedback, the reinforcement learning agent incorporates it into the experience replay buffer for subsequent policy gradient updates or Q-value updates, thereby learning that "delay decisions leading to high compensation are suboptimal," and adjusting its strategy in future scheduling to seek a better balance between meeting urgent needs and minimizing compensation costs.

[0082] This application's embodiments generate differentiated and personalized compensation plans for orders delayed due to scheduling adjustments, based on their urgency, delay details, and user value. These plans are then promptly communicated to users via the client, helping to soothe user emotions and maintain user satisfaction. This is particularly crucial during periods of resource scarcity, ensuring system fairness and user experience. More importantly, the system quantifies compensation actions and introduces them as negative feedback signals into the training loop of the reinforcement learning agent. This allows the scheduling decision model to learn the correlation between compensation costs and scheduling decisions. Consequently, in future resource allocation and priority ranking, it can proactively balance service response speed and operational costs, continuously optimizing itself and ultimately achieving a dynamic balance and sustained improvement between long-term platform resource utilization efficiency and overall user satisfaction.

[0083] In some embodiments, this application provides a service prioritization and resource allocation method based on multi-factor dynamic evaluation in a domestic service scenario, the system comprising: The multi-dimensional urgency assessment engine receives user service request information and calculates the urgency score of user needs based on time, service type, user context, and historical behavior dimensions. Service request information includes: service type, expected time, user information, and special requirements description. The resource status monitoring module is used to collect and analyze the status information of available resources within the service area in real time, including the distribution of service personnel, skill matching degree, and resource saturation prediction. The dynamic priority scheduling decision engine is used to perform priority sorting, resource constraint analysis, special resource allocation, and time window adjustment based on urgency score and resource status information to generate the optimal scheduling plan. Resource status information includes: service personnel availability, skill matching degree, geographical distribution, and traffic conditions. The compensation mechanism and balancing subsystem are used to automatically generate compensation plans for orders whose priority has been downgraded, ensuring the overall user satisfaction of the system is balanced. The service execution and feedback module is used to execute the final scheduling decision and collect service process feedback to provide data support for system optimization.

[0084] In some embodiments, this application also provides a corresponding dynamic priority scheduling method, including the following steps: Receive user requests for housekeeping services and extract the service type, expected time, user information, and special requirements from the request. The urgency score of the demand is calculated based on multidimensional features, including but not limited to: time urgency, service type urgency, user-specific context and historical behavior data. Obtain current resource status information, including service personnel availability, skill matching, geographical distribution, and traffic conditions; A dynamic priority sorting algorithm is executed, taking into account both urgency score and resource constraints, to determine the processing priority of each order; For orders whose priority has been adjusted, a corresponding compensation plan will be automatically generated and pushed to the relevant users for confirmation; Execute the final scheduling decision and push the adjustment results to relevant users and service personnel; Collect service execution feedback, update algorithm parameters, and optimize system performance.

[0085] Compared with the prior art, the embodiments of this application have the following beneficial effects: This system achieves accurate assessment of demand urgency based on multi-dimensional features, improving accuracy compared to existing single-dimensional assessments. A dynamic priority scheduling mechanism prioritizes high-urgency demands with limited resources, reducing emergency service response time during periods of resource scarcity. An innovative compensation mechanism and balancing system are introduced to maintain overall user satisfaction while ensuring high-urgency needs are met. An adaptive learning mechanism continuously optimizes urgency assessment and scheduling strategies as the system operates.

[0086] In some embodiments, the overall system architecture of this application can be found in [reference needed]. Figure 2 The service priority ranking and resource allocation system based on multi-factor dynamic evaluation in the domestic service scenario provided in this application includes: a user interface module S1, a multi-dimensional urgency evaluation engine S2, a resource status monitoring module S3, a dynamic priority scheduling decision engine S4, a compensation mechanism and balancing subsystem S5, and a service execution and feedback module S6.

[0087] The user interface module S1 is responsible for receiving service requests submitted by users through mobile applications and web pages, collecting information such as service type, expected service time, service address, and special requirement description contained in the request, and passing this information to the multi-dimensional urgency assessment engine S2.

[0088] The multi-dimensional urgency assessment engine S2 is one of the core components of this system, and its structure is as follows: Figure 3 As shown, it includes a data preprocessing unit M1, a feature extraction unit M2, a multi-dimensional scoring model M3, and an urgency calculation unit M4. This engine assesses the urgency of user needs based on the following four dimensions: The time dimension refers to assessing the urgency of time by analyzing the difference between the user's expected service time and the current time, the degree of service delay, etc. The service type dimension refers to the fact that different service types have different inherent urgency levels. For example, repairing a leaky pipe is more urgent than routine cleaning. The user context dimension refers to considering the special circumstances of the user's family, such as having infants, the elderly, or an important event to be held. Historical behavior dimension refers to analyzing users' historical order data, including usage frequency, ratings, cancellation rate, etc.

[0089] Real-time environmental features include, but are not limited to, weather, regional time, and time segments.

[0090] Order semantic features refer to the special requirements descriptions in the order request information.

[0091] The urgency score is calculated using the following formula (1): Urgency score = w1 × time urgency factor + w2 × service type factor + w3 × user context factor + w4 × historical behavior factor (1); Among them, w1, w2, w3, and w4 are dynamic weight coefficients. Their initial values ​​are determined based on historical data analysis and dynamically optimized based on system operation feedback through machine learning algorithms.

[0092] Resource status monitoring module S3 is responsible for real-time collection and analysis of service resource status, including: Service personnel information includes, but is not limited to, real-time location, skill level, and current task status; Regional resource distribution includes, but is not limited to, the density of service personnel in each region and the average response time. Traffic conditions include, but are not limited to, real-time traffic information and estimated arrival time; Resource forecasting refers to the prediction of resource demand for different time periods based on historical data.

[0093] The dynamic priority scheduling decision engine S4 is used to execute tasks based on urgency scores and resource status, such as... Figure 4The scheduling process shown includes: Initial sorting refers to the preliminary sorting of orders to be processed based on their urgency scores; Resource constraint analysis refers to adjusting the initial ranking based on the current availability of resources; Special resource allocation refers to assigning the most suitable service personnel to high-priority orders; Time window adjustment refers to adjusting the service time window when necessary; The cascading impact assessment refers to analyzing the impact of priority adjustments on other orders.

[0094] The compensation mechanism and balancing subsystem S5 are used to generate compensation schemes to ensure system fairness, wherein: Automatic compensation is provided for orders whose priority has been downgraded, such as service discounts, extra service time, coupons, etc. The compensation level is dynamically calculated based on the reduction amount and user value; a satisfaction balance strategy is implemented to ensure the overall user experience.

[0095] The service execution and feedback module S6 is used to: send the final scheduling decision to users and service personnel; collect status updates during service execution; collect user feedback and satisfaction evaluations after service completion; and use the feedback data for system parameter optimization.

[0096] In some embodiments, refer to Figure 5 The multidimensional urgency assessment method includes the following steps: Step 401: Receive user service request information, including service type, expected service time, service address, and description of special requirements.

[0097] Step 402: Extract time dimension features and calculate the time urgency factor T, which is calculated using the following formula (2): T = α1×(1 / time difference) + α2×service duration + α3×time period tension (2); Among them, the time difference is the difference between the expected service time and the current time, the service duration is the estimated service time, the time slot tension is the historical order density of the time slot, and α1, α2, and α3 are the internal weights of the time dimension.

[0098] Step 403: Extract service type dimension features and calculate the service type factor S, which is calculated using the following formula (3): S = β1×service basic urgency + β2×seasonal impact factor (3); Among them, the basic urgency of the service is preset according to the service type (e.g., 0.9 for leak repair and 0.3 for daily cleaning), the seasonal impact factor considers the impact of the season on the demand for different services (e.g., the urgency of heating maintenance increases in winter), and β1 and β2 are the internal weights of the service type dimension.

[0099] Step 404: Extract user context dimension features and calculate the user context factor C. The calculation formula is as follows: Formula (4): C = γ1×Family Special Circumstances + γ2×Activity Urgency + γ3×Health and Safety Impact (4); Among them, special family circumstances take into account factors such as the presence of infants / elders at home, the urgency of the activity takes into account situations such as important events to be held soon, and the impact on health and safety takes into account the health and safety risks that may be caused by service delays. γ1, γ2, and γ3 are the internal weights of the user context dimension.

[0100] Step 405: Extract historical behavior dimension features and calculate the historical behavior factor H, which is calculated using the following formula (5): H = δ1×User Loyalty + δ2×Historical Evaluation Score + δ3×(1-Historical Cancellation Rate) (5); Among them, user loyalty is based on the frequency and amount of historical orders, historical evaluation is the average score of users' historical evaluations, historical cancellation rate is the proportion of orders cancelled by users in the past, and δ1, δ2, and δ3 are the internal weights of the historical behavior dimension.

[0101] Step 406: Calculate the urgency score E, which is calculated using the following formula (6): E = w1×T + w2×S + w3×C + w4×H (6); Among them, w1, w2, w3, and w4 are the weight coefficients for each dimension, initially set to 0.3, 0.3, 0.25, and 0.15, and dynamically adjusted based on system operation feedback through machine learning algorithms.

[0102] Step 407: Normalize the urgency score to generate an urgency score of 0-100. The calculation formula is as follows: Formula (7): Normalized urgency score = (E - Emin) / (Emax - Emin) × 100 (7); Among them, Emin and Emax are the minimum and maximum values ​​of the historical urgency score, respectively.

[0103] In some embodiments, refer to Figure 4 The specific process of the dynamic priority scheduling decision method is as follows: Step 301: Based on the urgency score, sort all pending orders in descending order of urgency score to obtain the initial priority sequence.

[0104] Step 302: Resource constraint analysis, considering the current available resource status, including: availability of personnel with specific skills required for the service type; matching degree between service location and service personnel location; matching degree between service time and service personnel availability time.

[0105] Step 303: Priority Adjustment Decision. Based on the resource constraint analysis results, execute the following decisions: maintain the initial priority when resources are sufficient; upgrade orders with high urgency but low initial priority; and downgrade orders with low urgency that can be postponed when resources are tight.

[0106] Step 304: Special resource allocation. Assign the most suitable service personnel to high-priority orders, considering the following factors: select the service personnel with the best skills match; select the personnel closest in distance, provided that the skills match; give priority to personnel who have served the user before and have given good reviews.

[0107] Step 305: Adjust the time window. If necessary, adjust the service time window, including: adjusting to the earliest feasible time within the user's acceptable range; postponing to a period of time with relatively sufficient resources; and calculating and displaying the estimated arrival time in real time.

[0108] Step 306: Assessment and handling of cascading effects, analyzing the impact of priority adjustments on other orders, including: constructing an order dependency graph to identify the scope of cascading effects; calculating the overall impact cost to ensure that the overall benefits of the adjustment outweigh the costs; and implementing corresponding compensation measures for affected orders.

[0109] In some embodiments, refer to Figure 6 The specific methods of the compensation mechanism and balancing strategy are as follows: Step S1: For orders whose priority has been adjusted, the system calculates a compensation coefficient based on the adjustment range and user value: Compensation coefficient = Priority reduction range × (1 + User value coefficient); where the priority reduction range is the percentage change in queue position before and after the adjustment, and the user value coefficient is based on the user's historical contribution and potential value assessment.

[0110] In step S2, the system automatically generates one or more of the following compensation schemes based on the compensation coefficient and the specific scenario: calculate the discount ratio based on the compensation coefficient; provide a higher level of service or extend the service time; issue coupons with the amount proportional to the compensation coefficient; give extra points that can be redeemed for services or goods; temporarily upgrade the membership level or open specific benefits.

[0111] Step S3: The system uses a satisfaction balancing algorithm to ensure that the compensation plan can effectively offset the negative experience caused by the priority adjustment. The calculation formula is as follows: Satisfaction Balance Index = Original Satisfaction Expectation - Priority Adjustment Impact Value + Compensation Satisfaction Improvement Value.

[0112] The compensation strategy is considered effective when the satisfaction balance index is not lower than the preset threshold.

[0113] In some embodiments, refer to Figure 7 Taking emergency pipe leak service as an example, the application process of this system is explained as follows: Step 601: The user submits an emergency pipe leak service request through the APP, selects the "emergency repair" type, and explains in the remarks "the kitchen water pipe is broken and water is flowing continuously"; Step 602: The multi-dimensional urgency assessment engine analyzes the request information, as exemplified below: The user requests immediate service, with a time urgency factor T=0.95; the water pipe leak is a high-urgency service, with a service type factor S=0.90; the remarks mention that the water is flowing continuously, affecting normal life, with a situation factor C=0.85; the user is an active user on the platform with good reviews, with a historical behavior factor H=0.80; the urgency score is calculated as E=0.3×0.95+0.3×0.90+0.25×0.85+0.15×0.80=0.885; after normalization, the urgency score is 0.885 (high urgency).

[0114] Step 603: The resource status monitoring module detected that there are 3 plumbing technicians in the current area, all of whom have made appointments; Step 604: The dynamic priority scheduling decision engine analyzes and decides, for example, as follows: Select technician A from the 3 technicians whose current task is "routine maintenance" (low urgency); adjust technician A's work sequence and insert the leak repair as the next task; postpone the original "routine maintenance" task by 2 hours.

[0115] Step 605: The compensation mechanism automatically generates a service discount compensation plan for users whose "routine maintenance" has been postponed; Step 606: The system sends a confirmation message to the user with the emergency leak that "the repair personnel will arrive within 30 minutes" and sends a new work assignment to technician A at the same time. Step 607: The service execution and feedback module tracks the entire service process, collects user feedback, and updates algorithm parameters.

[0116] Through the above processing, the system successfully prioritized the allocation of limited technical personnel resources to emergency leak services, while ensuring the satisfaction of affected users through a compensation mechanism.

[0117] In some embodiments, refer to Figure 8 During the peak period of pre-holiday cleaning, the resource scheduling process of this system is as follows: Step 801: The system detected a surge in demand for cleaning services as the Spring Festival approaches, and the resource status monitoring module reported that the service personnel saturation rate has reached 95%. Step 802: The multi-dimensional urgency assessment engine scores all cleaning service requests based on the "before holidays" time factor, and takes into account the specific circumstances of each user (such as whether there are family gathering plans, elderly people, children, or other special circumstances). Step 803: The dynamic priority scheduling decision engine sorts all pending orders into three priority levels: high, medium, and low, based on urgency scores and user value. Step 804: For high-priority orders, ensure the original service time remains unchanged; for medium-priority orders, adjust the service time within an acceptable range; for low-priority orders, recommend alternative dates or refer to partner service providers. Step 805: The compensation mechanism provides different levels of compensation for medium and low priority orders: medium priority orders receive an x% discount, and low priority orders receive a y% discount plus additional service time; Step 806: The system pushes the adjustment plan and compensation instructions to the user through the APP. The user can choose to accept the adjustment or stick to the original plan (but may have to wait). Step 807: Based on user feedback, the system adjusts the resource allocation plan in real time and updates the urgency assessment and scheduling algorithm parameters.

[0118] This approach allows the platform to prioritize users with the most urgent needs during periods of resource scarcity, while maintaining overall user satisfaction through reasonable compensation measures.

[0119] In some embodiments, the goal of the multimodal demand urgency assessment model is to generate a quantified "Urgency Score" (US), the calculation of which no longer relies solely on the user's explicitly stated expected time, but rather on an algorithmic model that incorporates multidimensional data.

[0120] The model receives the three types of features listed in Table 1 below and performs standardization / normalization processing: Table 1

[0121] The model's algorithms and mechanisms include, but are not limited to, regression / classification models based on Gradient Boosting Decision Trees (GBDT). XGBoost or LightGBM can be chosen as models because they effectively handle heterogeneous features and are insensitive to missing values. The model's output is an urgency score (US) between 0 and 100. Training data can be historical order data, containing all the aforementioned features as input. Training labels include primary and secondary labels. The primary label is the time interval between order notification and order acceptance by the service provider. The shorter the interval, the more urgency the order has historically been "defaulted" to by the system. This is the target we want the model to learn. The secondary label is the user's rating and complaint status after the order is completed. If an order is processed quickly but the user is still dissatisfied, it indicates that our urgency judgment may be inaccurate, and this needs to be penalized in the loss function. The training process involves training the model using historical data to minimize the difference between the predicted urgency score and the actual order acceptance time interval (and user feedback).

[0122] In some embodiments, the model algorithm can also be a dynamic weight fusion algorithm, as detailed below: For each feature category (User U, Environment E, Semantics S), a sub-score is calculated. The final urgency score satisfies: US = α × U_score + β × E_score + γ × S_score. Here, the coefficients α, β, and γ are not fixed but dynamically adjusted through a lightweight online learning model. For example, the system continuously tracks whether the environmental feature E_score has a greater impact on the final order acceptance speed on rainy days. If so, the weight of β is automatically increased.

[0123] This embodiment combines user profiling analysis, environmental perception, and natural language processing technologies in a non-obvious way with the specific field of domestic service dispatch. Furthermore, the model weights or coefficients are derived from machine learning of historical data rather than human experience settings. This gives the model the ability to self-evolve, and its decision-making logic is not revealed by existing technologies. It also transforms the vague business concept of "urgency" into a quantifiable machine learning prediction problem with a real label (order acceptance time).

[0124] In some embodiments, the resource dynamic allocation strategy based on reinforcement learning in this application treats the entire scheduling system as an agent, which learns the optimal order sorting and personnel matching strategy through interaction with the environment.

[0125] The state space is defined as follows: State S_t, a multi-dimensional vector that comprehensively describes the system's state at time t; gridded resource state, dividing the city into 1km×1km grids, each grid containing the number of available service personnel, the number of ongoing orders, and the estimated number of orders to be completed in the next hour (based on service duration prediction); and order queue state, a list of orders currently awaiting allocation, with each order defined by its urgency score (US), service type, grid location, and expected service duration. Global metrics include the system's current overall load rate and average response time.

[0126] The Action Space is defined as follows: Actions A_t that an agent can perform include: order-person matching, selecting an order from the queue of orders to be assigned and assigning it to a specific idle service personnel; order delay decision, deciding to put a non-urgent order into the "delay pool" and not assign it for the time being; and cross-grid scheduling, instructing an idle service personnel in a certain grid to move to a nearby high-demand grid to stand by.

[0127] In some embodiments, the reward function is designed as follows: Rewards are key to guiding an agent's learning; they are a composite function balancing the interests of multiple parties, including: immediate rewards, basic rewards, and periodic rewards, among which: Instant rewards include: successfully assigning an order.

[0128] Basic rewards include: assigning a high urgency (US) order; rewards being proportional to urgency to encourage handling urgent orders; penalties for excessively long service personnel distances to user locations to optimize efficiency; and orders being canceled by users due to waiting timeouts.

[0129] Periodic rewards (settled every N time steps or hour) include: if the average order completion time decreases during the period; if the income balance of service personnel (Gini coefficient) improves during the period; if the average user rating increases during the period.

[0130] In some embodiments, the reinforcement learning algorithm and training method are as follows: Due to the large and continuous state and action space, deep reinforcement learning algorithms, such as Proximal Policy Optimization (PPO) or Deep Q-Network (DQN), are suitable. Training is performed using historical log playback or a simulation environment. The simulation environment can simulate the arrival of new orders based on historical order generation patterns. After initial training, the model can be deployed online for fine-tuning in an "exploration-exploitation" mode, continuously adapting to changes in order patterns.

[0131] In some implementations, the system workflow is as follows: The US (final urgency score) is calculated using the "Multimodal Demand Urgency Assessment Model". The reinforcement learning agent observes the current system state S_t; The agent selects an action A_t based on the policy π(a|s), for example, assigning order O to service personnel P; Once the action is executed, the system enters a new state S_t+1 and calculates the reward R_t; The experience (S_t, A_t, R_t, S_t+1) is stored in the experience replay pool and used to update the neural network parameters; It executes in a loop to achieve continuous, adaptive dynamic scheduling; This application's embodiment is a domestic service system based on multimodal emergency assessment and reinforcement learning dynamic scheduling. It abandons the traditional sorting and matching logic based on fixed rules, and achieves a high degree of dynamic adaptation of resources and needs through the collaborative work of a data-driven urgency perception brain and a scheduling decision brain with a long-term vision.

[0132] In some embodiments, the integrated process is as follows: P1. Receive user-submitted housekeeping service requests, specifically including: service type, expected service time, user information, and any special requirements. The system receives requests through a well-defined service interface, typically a RESTful API endpoint based on the HTTP protocol. The request payload is encapsulated in a structured data format, containing at least key fields such as service type, expected service time, user identifier, and unstructured special requirements description text. Upon receiving the request, the system first performs data validation and cleaning to ensure that required fields are complete and formatted correctly. Subsequently, the system generates a globally unique order identifier for the request and persists the complete request information to a transactional database. Simultaneously, the request event is published to a message middleware to asynchronously trigger downstream processing pipelines.

[0133] For example, the system receives a service request whose structured data indicates: service type is "pipe repair", expected time is "2024-05-20 14:00:00", user ID is "U12345", and special requirement description is "a water pipe bursts at home, an elderly person is home alone, and urgent help is needed". After the system verifies the request, it assigns an order ID "REQ_20240520_001" and stores it in the database.

[0134] P2. Based on the multimodal fusion model, calculate the dynamic demand urgency score, as follows: This step is no longer a simple calculation, but rather initiates a multimodal demand urgency assessment model to extract user profile features from user information, such as historical order frequency and user category tags. It obtains real-time environmental features from external APIs, such as weather and regional events. Semantic analysis is performed on the "Special Request Description" to extract keywords and conduct sentiment analysis. These features are then input into a pre-trained GBDT model (or a dynamic weight fusion algorithm). This generates a precise, quantified dynamic urgency score. This score comprehensively reflects the personalized, environmentally sensitive, and semantically understood urgency level.

[0135] This step initiates the multimodal demand urgency assessment model. The model first extracts features from three independent dimensions: user profile features, which extracts static and dynamic attributes related to the user identifier from the user database, such as historical order frequency and user category tags (e.g., "families with infants," "elderly living alone"). Real-time environmental features, which obtain real-time contextual information related to the order's geographical location by calling external application programming interfaces (APIs), such as current and forecast weather conditions and the presence of any major regional events.

[0136] Order semantic features involve natural language processing of the specific request description text submitted by the user. This process includes pattern matching based on a predefined keyword library and using sentiment analysis models to determine the polarity (positive, negative, neutral) and intensity of the emotions expressed in the text.

[0137] Subsequently, the extracted multi-dimensional feature vectors are input into a pre-trained machine learning model (such as Gradient Boosting Decision Tree, GBDT) for inference. This model interacts with the features through non-linear transformations and outputs a normalized scalar score, namely the dynamic urgency score, which comprehensively reflects the urgency of the order at three levels: personalization, environmental sensitivity, and semantic understanding.

[0138] For example, for the order "REQ_20240520_001", the model extracts the user as an "elderly person living alone" (user profile), the current weather as "heavy rain" (environmental feature), and the text contains the keyword "burst water pipe" with a sentiment polarity of "strongly negative" (semantic feature). After integrating this information, the model calculates a dynamic urgency score of 85.6 points (out of 100).

[0139] P3. Obtain real-time and predicted resource status of the service area. The specific process is as follows: Obtain the distribution of service personnel, skill matching degree, current location and status (idle, working). Based on digital twins or historical data, predict the resource saturation trend of each grid area in the future. The current queue of orders to be assigned and their dynamic urgency scores.

[0140] This application's embodiments construct an enhanced resource status view in the following way: The system continuously collects and aggregates information such as the geographical location, current status (idle, working, offline), and skill tags of service personnel from their clients. Through spatial computing, personnel and orders are mapped to pre-divided geographical grids, and resource supply and demand indicators within each grid are calculated in real time, such as idle personnel density, number of pending orders, and average urgency. Based on historical order data, personnel movement patterns, and known future events (such as holidays), time series forecasting algorithms or digital twin technology are used to predict the resource saturation trend of each grid within a specific future time window (such as the next hour).

[0141] In this embodiment, the above-mentioned real-time and predicted status information, together with all currently pending order queues and their respective dynamic urgency scores, constitute a global, spatiotemporal snapshot of resource status.

[0142] For example, the system detected that within "Grid A03," there are currently 2 available pipeline repair technicians but 5 pending orders, including high-urgency orders. Meanwhile, the predictive model indicates that due to heavy rain, pipeline repair requests in this area will increase by 50% within the next hour. This information is integrated into the resource status.

[0143] P4. The reinforcement learning agent makes global decisions and generates the final scheduling sequence. The specific process is as follows: The reinforcement learning agent receives enhanced resource states (including prediction information) from P3 and dynamic urgency scores of orders from P2, which together constitute its state space S_t. Based on its trained policy network, the agent analyzes the current state and outputs an action A_t. This action is not a simple ranking, but a complex decision, such as: Assign "Order A" (urgency level 95) to "Technician Wang Wu" in "Grid X"; Delay the allocation of "Order B" (urgency level 60) for 10 minutes; The idle technician in the "Grid Y" command is moved to the adjacent high-demand "Grid Z"; The agent's series of decisions essentially directly generate a final order execution sequence and personnel matching scheme that takes into account global efficiency and long-term benefits.

[0144] The reinforcement learning agent takes the enhanced resource state snapshot output by S3 as input to its state space. This agent encapsulates a deep neural network policy trained on massive amounts of historical interaction data. After evaluating the current state, the policy network outputs a specific scheduling decision action. This action space is designed to include multiple decision types, specifically: order allocation, immediately assigning specific high-urgency orders to a specific service personnel within the grid with matching skills and optimal distance; order delay, deciding the specific duration for delaying the allocation of low-urgency orders to wait for better resource matching or to give way to higher-priority orders; resource allocation, instructing idle service personnel in low-demand grids to pre-move to neighboring high-demand or soon-to-be-high-demand grids to stand by. The series of decision actions output by the agent logically constitutes a final scheduling sequence and resource allocation scheme that considers global efficiency and long-term benefits.

[0145] For example, based on the current state, the agent decides: "Immediately assign order 'REQ_20240520_001' (urgency level 85.6) to technician 'W789' in grid A03", while "instructing idle technician 'W456' in grid B10 to move to grid A03 within 10 minutes".

[0146] P5. Execute scheduling decisions and perform closed-loop learning. The specific process is as follows: The system sends the scheduling instructions (order allocation, personnel movement) generated by S4 to the corresponding service personnel and users. The system records the results of this decision (such as whether the order is accepted, completion time, user rating) and stores it as experience data in the reinforcement learning experience replay pool for continuous optimization and online fine-tuning of the agent's decision-making strategy.

[0147] The scheduling execution engine parses the decision actions output by the reinforcement learning agent and transforms them into specific system instructions. For order allocation decisions, the system updates the order status in the database, binds it to the designated service personnel, and notifies both parties via push notifications. For personnel movement decisions, the system sends scheduling instructions to the clients of the relevant service personnel. After execution, the system initiates a closed-loop learning process: it continuously tracks and records the actual execution effect of each decision action, forming "experience data." This data includes key performance indicators such as the actual response time of orders, user ratings after service completion, and the work efficiency of service personnel. This experience data is stored in a first-in, first-out experience replay pool. The system periodically samples data batches from the pool to incrementally update the policy network parameters of the reinforcement learning agent, thereby achieving online optimization and adaptive adjustment of the decision model.

[0148] For example, the system executes the assignment instruction, notifying technician 'W789' to accept the order. After the order is completed, the user gives a 5-star review, and the system records this "status-action-reward" data, which is used for subsequent policy network fine-tuning, making the agent more inclined to make similar successful decisions in the future.

[0149] P6. Intelligent compensation and feedback based on decision value, the specific process is as follows; This step is triggered when the reinforcement learning agent makes the decision to "delay an order." A compensation calculation engine generates a personalized, optimal compensation plan (such as a coupon of a specific amount, service upgrade, or future priority) in real time based on the order's delay duration, its dynamic urgency score, the user's historical value, and the company's cost structure. The cost of this compensation is fed back to the reinforcement learning agent as a negative reward signal. In this way, the agent learns to weigh the "system resources saved by delaying the order" against the "compensation cost and loss of customer satisfaction" in the next decision, thus making a more economical overall decision. The generated compensation plan is automatically sent to the user. When the reinforcement learning agent makes the decision to "delay an order," the system triggers the intelligent compensation engine. This engine, based on a predefined compensation strategy model, takes into account the order's dynamic urgency score, the delay duration, the user's historical value (such as loyalty and spending power), and the platform's service cost structure, and outputs a personalized, optimal compensation plan. The compensation can take the form of a fixed-amount coupon, service upgrade, or priority for future orders.

[0150] This compensation plan is automatically sent to the user's end via the messaging system. Simultaneously, the economic or user experience cost associated with this compensation is quantified as a negative reward signal and fed back to the reinforcement learning agent. In subsequent learning processes, the agent will incorporate this cost into its calculations, thereby learning to make a refined economic trade-off between "optimizing system resources through delayed orders" and "the costs incurred due to compensation and user dissatisfaction."

[0151] For example, for an order "REQ_20240520_002" (urgency level 60) that is delayed by 30 minutes, the compensation engine calculates that providing a 25 yuan coupon is the optimal solution. The system issues this coupon to the user and simultaneously associates the cost of "-25" as a negative reward with this delay decision, feeding it back to the reinforcement learning model for learning.

[0152] Based on the foregoing embodiments, this application provides an intelligent scheduling device for service resources. The device includes various units and modules included in each unit, which can be implemented by a processor in a computer device; of course, it can also be implemented by specific logic circuits. In the implementation process, the processor can be a central processing unit (CPU), a microprocessor unit (MPU), a digital signal processor (DSP), or a field programmable gate array (FPGA), etc.

[0153] Figure 9 This is a schematic diagram illustrating the structural composition of an intelligent scheduling device for service resources provided in an embodiment of this application, as shown below. Figure 9 As shown, the intelligent scheduling device 20 for service resources includes: The receiving module 201 is used to receive service requests submitted by users, wherein the service requests include at least the service type, expected service time, and special requirements description; Processing module 202 is used to calculate the dynamic urgency score of the service request based on a multimodal fusion model, including: extracting multi-dimensional features associated with the service request, wherein the multi-dimensional features include at least user profile features, real-time environment features, and order semantic features obtained by semantic analysis based on the special requirement description; inputting the multi-dimensional features into a pre-trained urgency assessment model, and outputting a quantified dynamic urgency score; and obtaining the real-time resource status of the service area and the future resource status predicted based on historical data to form an enhanced resource status view; Output module 203 is used to make global scheduling decisions based on reinforcement learning agents and generate a final scheduling sequence, wherein: the reinforcement learning agent uses the dynamic demand urgency score and the enhanced resource status view as state inputs; the reinforcement learning agent outputs scheduling decision actions according to its policy network, and the scheduling decision actions include order allocation, order delay, or cross-regional deployment of service personnel.

[0154] Optionally, the user profile features include at least one of historical order frequency, user category tags, and historical complaint / cancellation rates; the real-time environmental features include at least one of weather conditions, regional events, and time segments; and the order semantic features include information extracted from the special requirements description through keyword matching or sentiment analysis.

[0155] Optionally, the pre-trained urgency assessment model is a gradient boosting decision tree model, which is trained based on the time interval between the push and acceptance of historical orders and user feedback data after the order is completed.

[0156] Optionally, the processing module 202 is further configured to: The service area is divided into grids; Real-time aggregation of status, skills, and location information of service personnel within each grid; Based on historical order data and service personnel movement patterns, predict the resource demand trends of each grid within a future set time period; It integrates the real-time status, predicted trends, and queue information of pending orders from each grid.

[0157] Optionally, the reinforcement learning agent employs proximal policy optimization or deep network algorithms, and the reward function of the reinforcement learning agent is configured to comprehensively reward multiple indicators, including allocating high-urgency orders, shortening service distances, reducing order cancellations, improving user ratings, and balancing service personnel income.

[0158] Optionally, the processing module 202 is further configured to: Execute the aforementioned scheduling decision action; Collect the execution result data of the scheduling decision action, and the execution result data includes at least the actual order response time and user rating; The execution result data is used as training samples to perform online learning and optimization of the policy network of the reinforcement learning agent.

[0159] Optionally, when the scheduling decision action is an order delay, the processing module 202 is further configured to: Based on the dynamic urgency score of delayed orders, delay duration, and user historical value, a personalized intelligent compensation plan is generated. Send the intelligent compensation scheme to the user terminal; The cost corresponding to the intelligent compensation scheme is quantified as a negative reward signal and fed back to the reinforcement learning agent.

[0160] This application implements a system that accurately quantifies the urgency of needs by integrating multi-dimensional features and combines real-time and predictive resource information with reinforcement learning for global optimization decisions. Ultimately, the system can intelligently prioritize high-urgency needs when resources are limited, while improving overall resource utilization efficiency and service response speed through proactive resource allocation and order scheduling. This achieves a balance and improvement in user satisfaction and platform operational efficiency in complex and ever-changing domestic service scenarios.

[0161] The descriptions of the apparatus embodiments above are similar to those of the method embodiments above, and have similar beneficial effects. In some embodiments, the functions or modules included in the apparatus provided in this application can be used to perform the methods described in the method embodiments above. For technical details not disclosed in the apparatus embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0162] If the technical solution of this application involves personal information, the product using this technical solution has clearly informed the user of the personal information processing rules and obtained the user's voluntary consent before processing the personal information. If the technical solution of this application involves sensitive personal information, the product using this technical solution has obtained the user's separate consent before processing the sensitive personal information, and also meets the requirement of "express consent". For example, at personal information collection devices such as cameras, clear and prominent signs are set up to inform users that they have entered the scope of personal information collection and that personal information will be collected. If an individual voluntarily enters the collection scope, it is deemed that they have agreed to the collection of their personal information; or on the personal information processing device, the personal information processing rules are clearly informed through signs / information, and authorization is obtained through pop-up information or by asking the individual to upload their personal information; wherein, the personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the types of personal information processed.

[0163] It should be noted that, in the embodiments of this application, if the intelligent scheduling method for the above-mentioned service resources is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of this application, or the part that contributes to the related technology, can be embodied in the form of a software product. This software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, mobile hard drives, read-only memory (ROM), magnetic disks, or optical disks. Thus, the embodiments of this application are not limited to any specific hardware, software, or firmware, or any combination of hardware, software, and firmware.

[0164] This application provides a computer device including a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the program, it implements some or all of the steps in the above-described method.

[0165] This application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements some or all of the steps in the above-described method. The computer-readable storage medium can be transient or non-transient.

[0166] This application provides a computer program including computer-readable code, wherein when the computer-readable code is executed in a computer device, a processor in the computer device performs some or all of the steps in the above-described method.

[0167] This application provides a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program. When the computer program is read and executed by a computer, it implements some or all of the steps in the above-described method. This computer program product can be implemented specifically through hardware, software, or a combination thereof. In some embodiments, the computer program product is specifically embodied as a computer storage medium; in other embodiments, the computer program product is specifically embodied as a software product, such as a software development kit (SDK), etc.

[0168] It should be noted that the descriptions of the various embodiments above tend to emphasize the differences between them, while their similarities or commonalities can be referred to interchangeably. The descriptions of the above embodiments of the device, storage medium, computer program, and computer program product are similar to the descriptions of the above method embodiments and have similar beneficial effects. For technical details not disclosed in the embodiments of the device, storage medium, computer program, and computer program product of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0169] It should be noted that, Figure 10 This is a schematic diagram of a hardware entity of a computer device in an embodiment of this application, such as... Figure 10 As shown, the hardware entity of the computer device 700 includes: one or more processors 701, a communication interface 702, and a memory 703, wherein: Processor 701 typically controls the overall operation of computer device 700.

[0170] Communication interface 702 enables computer devices to communicate with other terminals or servers over a network.

[0171] The memory 703 is configured to store instructions and applications executable by the processor 701, and can also cache data to be processed or already processed (e.g., image data, audio data, voice communication data, and video communication data) in the processor 701 and various modules in the computer device 700. It can be implemented using flash memory or random access memory (RAM). Data transfer between the processor 701, the communication interface 702, and the memory 703 can be performed via bus 704. Only one processor is shown in the figure; each processor 701 includes one or more cores.

[0172] It should be noted that the computer device may include multiple processors 701, and each processor 701 can interact with each other through aggregated communication methods such as all-to-all, all-gather, or all-reduce. The processors 701 may be central processing units (CPUs), graphics processing units (GPUs), embedded neural network processing units (NPUs), tensor processing units (TPUs), data processing units (DPUs), accelerated processing units (APUs), floating-point processing units (FPUs), or application-specific integrated circuits (ASICs). The processors may also be single-core or multi-core processors. The processor may consist of a CPU and hardware chips. The hardware chips may be ASICs, PLDs, or combinations thereof. The PLDs may be complex programmable logic devices (CPLDs), FPGAs, generic array logic (GALs), or any combination thereof. The processor can also be implemented using logic devices with built-in processing logic, such as FPGAs or digital signal processors (DSPs).

[0173] The communication interface 702 can be a wired interface or a wireless interface, used to communicate with other modules or devices. The wired interface can be an Ethernet interface, a local interconnect network (LIN), etc., and the wireless interface can be a cellular network interface or a wireless LAN interface, etc.

[0174] Memory 703 can be non-volatile memory, such as read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Memory 703 can also be volatile memory, which can be random access memory (RAM) used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synclink dynamic random access memory (SLDRAM), and direct rambus RAM (DRRAM), direct rambus DRAM (DRDRAM), and rambus DRAM.

[0175] The 704 bus can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The bus can be divided into address bus, data bus, control bus, etc.

[0176] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above steps / processes do not imply a sequential order of execution; the execution order of each step / process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the above embodiments of this application are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0177] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0178] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0179] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0180] In addition, each functional unit in the various embodiments of this application can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0181] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as mobile storage devices, read-only memory (ROM), magnetic disks, or optical disks.

[0182] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, or the part that contributes to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, magnetic disks, or optical disks.

[0183] The above description is merely an embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for intelligent scheduling of service resources, characterized in that, The method includes: Receive service requests submitted by users, the service requests including at least the service type, expected service time and special requirements description; The dynamic urgency score of the service request is calculated based on a multimodal fusion model, including: extracting multi-dimensional features associated with the service request, wherein the multi-dimensional features include at least user profile features, real-time environment features, and order semantic features obtained by semantic analysis based on the special requirement description; inputting the multi-dimensional features into a pre-trained urgency assessment model, and outputting a quantified dynamic urgency score. It acquires the real-time resource status of the service area and the future resource status predicted based on historical data, forming an enhanced resource status view; Global scheduling decisions are made based on a reinforcement learning agent to generate a final scheduling sequence, wherein: the reinforcement learning agent uses the dynamic demand urgency score and the enhanced resource state view as state inputs; the reinforcement learning agent outputs scheduling decision actions according to its policy network, the scheduling decision actions include order allocation, order delay, or cross-regional deployment of service personnel.

2. The method according to claim 1, characterized in that, The user profile features include at least one of historical order frequency, user category tags, and historical complaint / cancellation rates; the real-time environmental features include at least one of weather conditions, regional events, and time segments; and the order semantic features include information extracted from the special requirements description through keyword matching or sentiment analysis.

3. The method according to claim 1 or 2, characterized in that, The pre-trained urgency assessment model is a gradient boosting decision tree model, which is trained based on the time interval between the push and acceptance of historical orders and user feedback data after the order is completed.

4. The method according to claim 1, characterized in that, The enhanced resource status view is constructed in the following way: The service area is divided into grids; Real-time aggregation of status, skills, and location information of service personnel within each grid; Based on historical order data and service personnel movement patterns, predict the resource demand trends of each grid within a future set time period; It integrates the real-time status, predicted trends, and queue information of pending orders from each grid.

5. The method according to claim 1, characterized in that, The reinforcement learning agent employs proximal policy optimization or deep network algorithms. The reward function of the reinforcement learning agent is configured to comprehensively reward multiple indicators, including allocating high-urgency orders, shortening service distances, reducing order cancellations, improving user ratings, and balancing service personnel income.

6. The method according to claim 1, characterized in that, The method further includes: Execute the aforementioned scheduling decision action; Collect the execution result data of the scheduling decision action, and the execution result data includes at least the actual order response time and user rating; The execution result data is used as training samples to perform online learning and optimization of the policy network of the reinforcement learning agent.

7. The method according to claim 1, characterized in that, When the scheduling decision action is an order delay, the method further includes: Based on the dynamic urgency score of delayed orders, delay duration, and user historical value, a personalized intelligent compensation plan is generated. Send the intelligent compensation scheme to the user terminal; The cost corresponding to the intelligent compensation scheme is quantified as a negative reward signal and fed back to the reinforcement learning agent.

8. An intelligent scheduling device for service resources, characterized in that, The device includes: The receiving module is used to receive service requests submitted by users, and the service requests include at least the service type, expected service time and special requirements description. The processing module is used to calculate the dynamic urgency score of the service request based on a multimodal fusion model, including: extracting multi-dimensional features associated with the service request, wherein the multi-dimensional features include at least user profile features, real-time environment features, and order semantic features obtained by semantic analysis based on the special requirement description; inputting the multi-dimensional features into a pre-trained urgency assessment model and outputting a quantified dynamic urgency score; and obtaining the real-time resource status of the service area and the future resource status predicted based on historical data to form an enhanced resource status view. The output module is used to make global scheduling decisions based on the reinforcement learning agent and generate a final scheduling sequence, wherein: the reinforcement learning agent uses the dynamic demand urgency score and the enhanced resource state view as state inputs; the reinforcement learning agent outputs scheduling decision actions according to its policy network, and the scheduling decision actions include order allocation, order delay, or cross-regional deployment of service personnel.

9. A computer device comprising a memory and a processor, the memory storing a computer program executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent scheduling method for service resources according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent scheduling method for service resources according to any one of claims 1 to 7.

Images