Hotel customer demand task management method and system

Abstract

The invention discloses a hotel customer demand task management method and system, and relates to the technical field of data service value evaluation and decision optimization, and the method comprises the steps: collecting guest room sensor data, collecting employee positioning data, collecting external environment data, extracting time-space correlation features, and constructing a multi-source heterogeneous data set; predicting task demand distribution in a future period by using a long short-term memory network, and generating a virtual task sequence with probability weight; performing space-time matching on the real-time task and the virtual sequence based on a mixed integer programming algorithm to generate a dynamic scheduling scheme of rolling updating; and tracking a task execution track, calculating a prediction deviation, and dynamically correcting model parameters and a cross-department resource coordination rule. Through the LSTM prediction model, mixed integer programming optimization scheduling, task trajectory tracking and cross-department cooperation, the intelligent level of hotel task management is improved, resource allocation is optimized, execution deviation is reduced, and service efficiency and customer satisfaction are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data service value evaluation and decision optimization, and in particular to a hotel guest demand task management method and system. Background Art

[0002] With the rapid development of information technology, data services play a vital role in all walks of life. Through the collection, processing and analysis of massive data, data services provide enterprises and organizations with decision support, business optimization and innovation-driven value. In recent years, artificial intelligence technology, especially deep learning models such as multi-layer perceptron (MLP), has demonstrated strong capabilities in the field of data analysis and prediction. MLP can capture complex nonlinear relationships and achieve accurate mapping from input to output through the connection of multiple neuron levels. This capability has made it widely used in fields such as user behavior prediction and personalized recommendation.

[0003] However, existing data service value assessment methods still have many shortcomings when facing diverse user needs and complex application scenarios. First, traditional methods often rely on single or limited indicators for evaluation, which fails to fully reflect the actual effect of data services, resulting in one-sided evaluation results. Secondly, there is a lack of in-depth analysis of user characteristics and failure to consider the differentiated needs of different user groups for data services, resulting in insufficient accuracy in personalized service recommendations. In addition, existing methods show limitations when dealing with nonlinear and high-dimensional data, making it difficult to effectively capture complex data patterns, affecting the accuracy of evaluation and decision-making. Finally, there is a lack of dynamic update mechanism, which makes it impossible to respond to user needs and environmental changes in a timely manner, resulting in a lag in decision optimization. Therefore, there is an urgent need for an innovative method based on a multi-layer perceptron that can comprehensively consider user characteristics and data service effects, build a scientific evaluation index system, and achieve accurate evaluation of data service value and decision optimization. Summary of the invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problems solved by the present invention are: the existing data service value assessment methods have single assessment indicators, lack of user feature analysis, insufficient ability to process nonlinear and high-dimensional data, and how to achieve accurate assessment of data service value and decision optimization.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: In a first aspect, an embodiment of the present invention provides a hotel guest demand task management method, including collecting room sensor data, collecting employee location data and collecting external environment data, extracting spatiotemporal correlation features to construct a multi-source heterogeneous data set; Use long short-term memory networks to predict the distribution of task requirements in future periods and generate virtual task sequences with probability weights; Based on the mixed integer programming algorithm, real-time tasks and virtual sequences are matched in time and space to generate a dynamic scheduling solution with rolling updates; Track task execution trajectories and calculate prediction deviations, dynamically correct model parameters and cross-departmental resource coordination rules.

[0007] As a preferred solution of the hotel guest demand task management method described in the present invention, wherein: collecting guest room sensor data includes: deploying temperature and humidity sensors in the guest rooms, periodically collecting environmental data and transmitting via wireless communication protocols; Install door magnetic sensors to monitor the open and closed status of guest room doors and the corresponding timestamps; Setting a weight sensor to monitor weight changes of items in the minibar and generating a raw weight data stream; Collecting employee location data includes: configuring ultra-wideband location tags for employees and obtaining real-time three-dimensional coordinate data through deployed positioning base stations; Linked with the elevator control system to record the target information of employees moving across floors; The collection of external environmental data includes: accessing the meteorological data interface to obtain the rainfall and temperature forecast values ​​for the future period; Obtain real-time traffic congestion level data around the hotel; Parse the local event calendar file and extract the number of participants and time information.

[0008] As a preferred solution of the hotel guest demand task management method described in the present invention, wherein: the extraction of spatiotemporal correlation features to construct a multi-source heterogeneous data set includes uniformly converting the timestamps of the room sensor data, employee location data and external environment data to the time axis of the same time zone, and using an interpolation method to complete the missing data; Map employee location coordinates and guest room locations to a unified floor plan coordinate system; Aggregate sensor data by room number to form a time series dataset; Group employee location data into movement trajectory data sets, recording time, coordinates, and floor information; Parse the unstructured text of event calendars to extract event types, number of participants, and time ranges; Calculate sliding window statistics in the time dimension for room data, including operation frequency and consumption rate; Generate a spatial service density heat map based on employee trajectories to quantify the service coverage intensity of each area; Convert time series data into image features and input them into the pre-trained model to extract spatiotemporal correlation feature vectors; Detect logical conflicts between multi-source data and trigger manual review processes; Abnormal data beyond a reasonable range are eliminated, and valid data are retained to construct the final data set.

[0009] As a preferred solution of the hotel guest demand task management method described in the present invention, wherein: a long short-term memory network model is constructed, a spatiotemporal correlation feature vector is input, and a probability distribution of each task type in the future time period is output; Regularly train the model incrementally and monitor prediction deviations, triggering emergency retraining to update model parameters; Identify the peak period of the task volume curve, and output the time window and task type distribution.

[0010] As a preferred solution of the hotel guest demand task management method described in the present invention, wherein: the generating of a virtual task sequence with probability weights comprises: generating a virtual task timestamp sequence based on a predicted task arrival rate, binding task types and confidence weights; Dynamically insert real-time tasks and remove low-confidence conflicting tasks; Pre-allocate resources and simulate to verify the feasibility of task sequences, and adjust weights to optimize execution efficiency.

[0011] As a preferred solution of the hotel guest demand task management method described in the present invention, wherein: the time-space matching of real-time tasks and virtual sequences based on a mixed integer programming algorithm includes establishing an optimization model including task allocation variables, time window constraints and resource conflict detection; Set trigger conditions to dynamically solve the model and prioritize high-confidence task allocation; The branch and bound method is combined with heuristic rules to achieve efficient solution.

[0012] As a preferred solution of the hotel guest demand task management method described in the present invention, wherein: the generation of a rolling updated dynamic scheduling solution includes dividing the scheduling time window and defining a task adjustment strategy; Dynamically balance employee load and task priorities to optimize resource allocation; Generate visual navigation instructions and push them to the execution terminal in real time.

[0013] As a preferred solution of the hotel guest demand task management method described in the present invention, wherein: the tracking of task execution trajectory and calculating prediction deviation includes collecting employee location data and task execution records to construct a spatiotemporal trajectory data set; Calculate the deviation between the actual task execution time and the predicted time and the path matching degree; Set deviation thresholds to trigger alerts and initiate the root cause analysis process.

[0014] As a preferred solution of the hotel guest demand task management method described in the present invention, the dynamic correction model parameters and cross-department resource coordination rules include training the prediction model according to the deviation data increment and adjusting the network parameters in layers; Optimize inter-departmental task coordination agreements and use reinforcement learning to update coordination rules; The stability of the update solution is ensured through simulation verification and grayscale release.

[0015] In a second aspect, an embodiment of the present invention provides a hotel guest demand task management system, including: Multi-source data fusion acquisition module: collects guest room sensor data, employee location data and external environment data, extracts spatiotemporal correlation features and constructs multi-source heterogeneous data sets; Intelligent demand forecasting module: uses long short-term memory network to predict the distribution of task demand in future periods and generate virtual task sequences with probability weights; Dynamic scheduling optimization module: Based on the mixed integer programming algorithm, real-time tasks and virtual sequences are matched in time and space to generate a dynamic scheduling solution with rolling updates; Execution feedback self-optimization module: tracks task execution trajectory and calculates prediction deviation, dynamically corrects model parameters and cross-departmental resource coordination rules.

[0016] The beneficial effects of the present invention are as follows: the present invention realizes accurate prediction of customer demand tasks and forward-looking resource allocation through multi-source data fusion and LSTM prediction model, reduces manual scheduling delays, and improves task response efficiency; based on mixed integer programming and rolling scheduling strategy, dynamically balances employee load and task priority, reduces resource idle rate, and avoids peak service congestion; through execution tracking and reinforcement learning feedback mechanism, continuously optimizes prediction models and cross-departmental collaboration rules, enables the system to adapt to environmental changes and improves operational stability; combines intelligent recommendation with active intervention in abnormalities, significantly reduces customer complaint rate, improves personalized service coverage, and enhances hotel brand competitiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work, among which: Figure 1 The present invention provides an overall flow chart of a hotel guest demand task management method according to the first embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0019] Example 1, reference Figure 1 , is an embodiment of the present invention, and provides a hotel guest demand task management method, comprising: S1: Collect guest room sensor data, employee location data, and external environment data, extract spatiotemporal correlation features, and construct a multi-source heterogeneous data set.

[0020] S11: Collect guest room sensor data, employee location data, and external environment data.

[0021] Collecting guest room sensor data includes: deploying temperature and humidity sensors in guest rooms, periodically collecting environmental data and transmitting them through wireless communication protocols; installing door magnetic sensors to monitor the opening and closing status of guest room doors and corresponding timestamps; setting weight sensors to monitor the weight changes of mini bar items and generate raw weight data streams; specifically: A temperature and humidity sensor (model SHT35) is fixedly installed in the center of the guest room ceiling, configured to collect environmental data every 5 minutes, and transmit the temperature (unit: °C) and humidity (unit: %) values ​​to the central server via the LoRaWAN protocol; Install a door magnetic sensor (Zigbee 3.0 protocol) on the guest room entrance door frame to record the door opening and closing status and the corresponding timestamp (UTC format), and store the original data directly in the database; A weight sensor (range 0-5kg, accuracy ±2g) is installed in the mini bar cabinet to monitor the weight changes of items in real time. The original weight data (unit: gram) is reported to the message queue through the MQTT protocol.

[0022] Collecting employee positioning data includes: configuring ultra-wideband positioning tags for employees, obtaining real-time three-dimensional coordinate data through deployed positioning base stations; linking with the elevator control system to record the target information of employees moving across floors; specifically: Each employee is assigned a UWB positioning tag (Decawave DW3000 chipset), and the tag ID is bound to the employee's work number; UWB positioning base stations (anchor points) are deployed every 15 meters in the hotel corridors, and the base station coordinates (3D Cartesian coordinate system) are pre-entered into the system; The signal transmission time difference between the tag and the base station is obtained through the TDoA (Time Difference of Arrival) algorithm, and the original time difference data (unit: nanoseconds) is uploaded to the positioning engine.

[0023] The collection of external environmental data includes: accessing the meteorological data interface to obtain the rainfall and temperature forecast values ​​for the future period; obtaining the real-time traffic congestion level data around the hotel; parsing the local event calendar file to extract the number of participants and time information; specifically: Call the Meteorological Bureau's open API (interface address: api.weather.gov) to obtain the rainfall intensity forecast value (data type: floating point number, unit: mm / h) and temperature forecast data for the next 6 hours every hour; Through the traffic interface (interface number: AMAP_TRAFFIC_001) of AMAP (or other map apps), the real-time congestion level of the 3 km road network around the hotel is obtained every 10 minutes (integer data, ranging from 0-10); Regularly download ICal format event calendar files from the local department's public data platform and parse the event title, number of participants (integer), and start time (ISO 8601 format) fields.

[0024] S12: Take spatiotemporal correlation features to construct multi-source heterogeneous datasets.

[0025] The timestamps of the guest room sensor data, employee location data, and external environment data are uniformly converted to the time axis of the same time zone, and the missing data is supplemented by interpolation method; specifically: Convert guest room sensor data (UTC format), employee location data (Unix millisecond timestamp), and external environment data (local time GMT+8) to the time zone of the hotel (such as Beijing time).

[0026] The time conversion formula is: ; in, is the timestamp of the original data, is the time zone offset (for example, UTC+8 corresponds to 28800 seconds).

[0027] The cubic spline interpolation method is used to fill the missing values ​​of the guest room sensor data. The interpolation formula is as follows: ; in, are the coefficients of the piecewise polynomial, which are solved according to the boundary continuity conditions; For the time interval [ , ), an interpolation function to fill in the missing data of the guest room sensor; , They are respectively and timestamp of a known data point, for example: if the sensor =10:00 and =10:05, if there are missing data, then interpolation is performed within this interval.

[0028] Through cubic polynomial fitting, we ensure that missing data is smoothly connected with the previous and next data after completion (for example, there is no sudden change in the temperature and humidity change curve), and avoid LSTM model prediction errors caused by data breakpoints.

[0029] Map the staff location coordinates and guest room locations to a unified floor plan coordinate system; specifically: Using rotation + translation transformation, the conversion formula is: ; in, , The original coordinates of the employee or device in the positioning base station coordinate system; , The coordinates in the converted unified floor plan coordinate system (aligned with the hotel architectural drawings) ensure that all positioning data (staff, guest rooms, equipment) are in the same spatial reference system to facilitate path planning and resource allocation; The rotation angle between the base station coordinate system and the architectural drawing (for example, the base station is not strictly aligned with the north direction during installation and needs to be corrected). is the translation offset.

[0030] According to the hotel layout, establish Grid resolution, generate guest room space index table.

[0031] The sensor data is aggregated by room number to form a time series dataset. Specifically, the data structure is in JSON format, with room number as the index and data arranged in timestamp order to generate a standardized time series dataset.

[0032] The employee location data is grouped into mobile trajectory data sets, recording time, coordinates, and floor information. Specifically, the data structure is in CSV format, grouped by tag ID, sorted by timestamp, and stored in employee trajectory data.

[0033] Parse unstructured text of event calendars to extract event type, number of participants, and time range.

[0034] Calculate the sliding window statistics of the time dimension for the guest room data, including operation frequency and consumption rate. Specifically: Calculation of door magnetic switch times: ; in, For the time window The number of times the guest room door status changes from closed to open; is the current time point, Window is the length of the time window, is a specific moment in the time window; middle, For time The door status (0=closed, 1=open); is the gate state at the previous moment (used to detect state transition); is the indicator function, defined as: ; Only count valid door actions from closed to open (such as guests entering the guest room) to avoid double counting (such as doors not fully closed).

[0035] Minibar item consumption rate: ; in, is the weight change, is the time interval.

[0036] Generate a spatial service density heat map based on employee trajectories to quantify the service coverage intensity of each area. Specifically: Divide the floor plan into 50cm×50cm grids and count the percentage of employees’ stay time in each grid; Gaussian smoothing filter, Gaussian smoothing of adjacent grids: ; in is the value after Gaussian smoothing, corresponding to the original data The result after Gaussian filtering is For the original data The value of the position; in this formula is the radius of the smoothing window, which determines the range of the Gaussian filter; Gaussian weight for: ; in, is the index variable of the Gaussian kernel function, which determines the contribution of adjacent pixels in the smoothing operation; σ is the standard deviation of the Gaussian function, which controls the degree of smoothing and is set to 1.5 grid units. Furthermore, this formula is the basic expression of the two-dimensional Gaussian function, and the specific derivation process is: First, we start with a one-dimensional Gaussian function: ; After being extended to two-dimensional space, in two directions (denoted as and Axis) respectively apply one-dimensional Gaussian function and multiply them, and we can get: ; If we assume that the variance of the Gaussian function in the two-dimensional space is the same in both directions, that is, , the formula is simplified to: ; The above formula is the derivation source of the two-dimensional Gaussian kernel function used in the present invention.

[0037] It should be noted that the two-dimensional Gaussian function is a classic filter kernel function, widely used in signal processing, image processing and other fields, and is derived based on the probability density function (PDF) of the normal distribution in statistics. To control the diffusion degree of the Gaussian function, the present invention selects = 1.5 grid units, suitable The value ensures sufficient smoothing of neighborhood information, avoiding over-smoothing that leads to feature loss, while not being too narrow to effectively remove noise. It is used in the Gaussian kernel to achieve exponential decay of distance, ensuring that the farther the data point is in space, the smaller its contribution is. This is consistent with the intuitive logic that the weight of adjacent data is greater than that of distant data when actual data is smoothed.

[0038] Convert time series data into image features and input them into the pre-trained model to extract spatiotemporal correlation feature vectors. Specifically: Normalize the time series: ; in, is the normalized time series value, is the original time series data, and are the maximum and minimum values ​​of the time series data, respectively.

[0039] Polar coordinate transformation: ; In this formula, For the The angle variable of the data point is used for polar coordinate conversion.

[0040] Generate the Gram matrix: ; in, is a Gram matrix element, defined as , used to calculate the correlation of time series data; here For the The angle variable for each data point.

[0041] The generated Gram matrix is ​​input into the ResNet-18 pre-trained model to extract high-level feature vectors, so as to extract high-level spatiotemporal feature vectors containing periodic evolution relationships, providing a structured input basis for subsequent LSTM network modeling and prediction.

[0042] Detect logical conflicts between multi-source data and trigger the manual review process; specifically, when the door status is "closed" but the infrared sensor detects human activity, it is marked as a data conflict and triggers manual review.

[0043] Abnormal data beyond a reasonable range are eliminated, and valid data are retained to construct the final data set. Specifically: Dynamic Threshold Detection: ; in, For the current time point Sensor data, The average value for the past 24 hours. is the standard deviation.

[0044] It should be noted that considering that the data collected by hotel sensors is easily affected by sudden factors or equipment failures, resulting in frequent abnormal data, thus affecting the accuracy of the overall analysis, a dynamic threshold detection mechanism based on the data of the past 24 hours is set. Specifically: each newly collected data point The average of the past 24 hours By comparing, the standard deviation σ24h is calculated, and the mean ±3 times the standard deviation is used as the upper and lower thresholds, thus forming a reasonable data range that is dynamically updated over time. If the newly collected data deviates significantly from this range, it satisfies the formula When the system detects data anomalies, it will automatically determine that the data is abnormal and trigger manual review to confirm the cause of the data anomaly.

[0045] Furthermore, on the one hand, it can adapt to data fluctuation characteristics in real time and dynamically to avoid false alarms or omissions caused by fixed thresholds; on the other hand, the authenticity of abnormal data can be further confirmed through manual review, and maintenance measures can be taken in time to ensure the data quality and accuracy of hotel guest demand task decision analysis.

[0046] S2: Use the long short-term memory network to predict the distribution of task requirements in the future period and generate a virtual task sequence with probability weights.

[0047] S21: Construct a long short-term memory network model, input the spatiotemporal correlation feature vector, and output the probability distribution of each task type in the future period; regularly incrementally train the model and monitor the prediction deviation, triggering emergency retraining to update the model parameters; identify the peak period of the task volume curve, and output the time window and task type distribution.

[0048] Specifically: the input vector dimension is set to 64, including temperature and humidity sliding window statistics, service density heat map values, GAF encoding features, etc.; the time step is set to 24 hours, with each hour as a time unit; standardization is used to normalize the features of each dimension to ensure consistent distribution of input data.

[0049] The LSTM network architecture design adopts a two-layer LSTM structure, with each layer containing 128 hidden units to avoid the gradient vanishing problem; "Dropout (dropout rate = 0.2)" is used to prevent overfitting; The hidden state transfer formula is: ; The formula represents the time step The hidden state of The hidden state of the previous time step and cell status Calculated; represents the number of LSTM layers, is the cell state; is the input feature vector (dimension 64) of the current time step; through this formula, LSTM gradually learns the long-term dependencies in time series data, ensuring that the task prediction is time-sensitive.

[0050] Add a multi-head attention mechanism (number of heads = 4) at the output of the second layer LSTM to calculate the historical state weights: ; in, , represents the dimension of the attention head; this mechanism can give different weights to different historical time steps, improve the long-term dependency ability of LSTM, and ensure more accurate task volume prediction; is the query matrix, which is composed of the current time step Hidden state vector By learning the parameter matrix The linear transformation is obtained and expressed as: ; is the key matrix, is the value matrix, which is the historical hidden state matrix of the past 24 hours (i.e. 24 time steps, 128 dimensions per step) The learnable parameter matrix , The linear transformation is obtained and expressed as: , ; in, The matrix dimensions are 24×128.

[0051] Furthermore, the specific calculation process of the historical state weight is as follows: First, calculate the similarity score between the query matrix and the key matrix: ; Among them, the denominator is the scaling factor (the dimension of the attention head) to avoid the dot product value being too large and causing softmax saturation, thus making the training more stable.

[0052] The similarity score Perform softmax normalization to obtain the normalized attention weights: ; Finally, the attention weights are multiplied by the value matrix VVV to obtain the output vector that integrates past historical state information: ; The above process is the derivation process of the historical state weight calculation formula. It should also be noted that the Softmax function ensures that the sum of the attention weights of all historical states is 1, which effectively expresses the relative importance of different historical states. The scaling factor comes from the classic practice in the Transformer attention mechanism, which prevents numerical instability caused by excessive inner products, thereby ensuring the stability of weight calculation and training convergence. By introducing a multi-head attention mechanism, the dependency of the long short-term memory network (LSTM) model on historical spatiotemporal data, especially the long-range dependency of time series data, can be effectively captured, ensuring that historical information is effectively integrated into the current prediction decision.

[0053] Incremental training is performed every 6 hours, and the training data window is updated in a rolling manner (retaining the data for the last 30 days); The loss function uses time-decayed weighted mean square error (MSE): ; in, is the actual task volume, The prediction task volume. Give higher weight to recent predictions, and the time step The closer to the current moment, the greater its weight The larger the value, the more attention the model pays to recent changes in task volume and improves prediction accuracy.

[0054] When the mean square error (MAE) of the task volume exceeds the threshold (default 20%) for three consecutive times, emergency retraining is triggered. Specifically, while the original LSTM model structure remains unchanged, the parameters of the first two layers in the network are frozen (that is, they are temporarily not involved in the training process), and only the parameters of the fully connected layer of the last layer of the network are re-fine-tuned and trained. At the same time, GPU parallel computing technology is used to improve computing efficiency and shorten training time. After training, the model weights are immediately updated online and deployed to the production environment in real time, so as to quickly correct the task volume prediction error and ensure that the accuracy of task demand prediction is always maintained within a reasonable range.

[0055] The final output layer of LSTM uses the Softmax activation function to generate the probability distribution of each task type in the next 6 hours: ; in, For the The logits value of the class task represents the model's task The degree of confidence in the probability of occurrence; is the logits value of the cleaning task, indicating the confidence level of the model in the task. Further, its derivation process is: The Softmax function first converts the logits values ​​of various tasks output by the model After nonlinear transformation through exponential function, the value obtained after transformation is always positive, and the expression is: ; Sum the index values ​​for each task type: ; By dividing the index value of a specific task type (such as cleaning task) Divide by the sum of the index values ​​of all task types, that is: ; The above process is the derivation process of the probability distribution of each task type. The probability value of each task type is obtained (between 0 and 1), and the sum of the probability values ​​of all task types is 1, which meets the basic requirements of probability distribution. It should also be noted that the use of exponential functions can effectively widen the gap between the logits values ​​of different task types, significantly enhance the probability weights of tasks corresponding to larger logits values, and reduce the probability weights of tasks corresponding to smaller logits values, making the probability output of the model clearer and more explicit, which is convenient for subsequent task decision-making and resource scheduling. By dividing by the sum of all exponential values, it is ensured that the sum of the output task probability values ​​is always 1. This normalization method is a commonly used method in statistics and probability theory, and is widely used in classification problems in the field of machine learning.

[0056] By analyzing the second-order derivative of the predicted task volume, we can identify the task-intensive periods: ; When the conditions are met, mark the time point as the task peak period. The growth rate is positive , but the growth rate turned negative , indicating that the task volume is about to reach its peak, making it easier to schedule resources in advance.

[0057] S22: Generate a virtual task timestamp sequence based on the predicted task arrival rate, bind the task type and confidence weight; dynamically insert real-time tasks and remove low-confidence conflicting tasks; pre-allocate resources and simulate to verify the feasibility of the task sequence, and adjust the weight to optimize the execution efficiency; specifically: Virtual task time modeling: Poisson process is used to model the task arrival time, which is reflected as follows: ; in, For the The arrival time of each task; is the task arrival rate (unit: task / hour), random variable It follows a uniform distribution between (0,1) to ensure that the task interval conforms to the Poisson distribution characteristics. This formula comes from the definition of the Poisson process, which describes the random occurrence of tasks or events, in which the time interval between the arrival of any two tasks follows an exponential distribution. The probability density function of the exponential distribution is: ; By integrating the probability density function of the exponential distribution, we get the cumulative distribution function (CDF): ; Let it be equal to a uniformly distributed random number ,Pick , then: ; After algebraic transformation, we get: ; because ,but It also obeys uniform distribution, so the expression is further simplified to: ; The above is the time interval formula for tasks. Therefore, the arrival time of each task is Can be expressed as the arrival time of the last task Add a random variable that follows an exponential distribution It should also be noted that the exponential distribution and Poisson process are used to describe the arrival process of hotel tasks because the hotel's guest tasks (such as cleaning and delivery) are usually uncertain and independent of each other, and the interval between any two tasks is approximately random. This Poisson process assumption is very consistent with the real scene.

[0058] Bind the task type and sample the task type based on the Softmax probability, for example: cleaning task: 65%; delivery task: 20%.

[0059] Confidence weight calculation, adjust the weight according to the predicted peak time: ; in, For the The confidence weight of each task, To predict the peak time of the task, . Mission time The closer to the mission peak , the higher the confidence, the higher the priority of peak-time tasks. The weight formula uses an exponential function to characterize the sensitivity between task arrival time and peak time. The exponential function is ( ) grows rapidly when it approaches zero (i.e., when the task is close to the predicted peak), and slows down when it is far from the peak, which intuitively reflects the logic that the closer to the task peak period, the higher the confidence; the exponential decay function form is consistent with the actual demand forecast task management scenario. The closer the task is to the peak period, the more priority the manager or system needs to pay attention to, and the weight should naturally be greater. The confidence of tasks far from the peak period is lower, which is also in line with actual decision-making preferences.

[0060] The conflict between computing tasks and virtual tasks: ; If the conflict degree is ≥ 0.3, the virtual task with the lowest confidence is eliminated.

[0061] For high confidence tasks , pre-allocate resources in advance, for example: cleaning tasks: pre-allocate vacuum cleaners and linen carts; delivery tasks: pre-load the robot cargo hold.

[0062] It should be noted that " "refers to the overlapping time of two tasks (for example, cleaning task A is scheduled to be executed from 10:00 to 10:30, and virtual task B is scheduled to be executed from 10:15 to 10:45, and the overlapping time is 15 minutes);" ” refers to the actual execution time required for a single task (for example: cleaning a room takes 20 minutes); ” refers to the minimum interval between tasks predicted by the system (for example, the time interval between adjacent tasks in a virtual task sequence is 15 minutes).

[0063] To verify the task sequence optimization, 1000 random sampling simulations were performed on the generated task sequence to evaluate the following indicators: stability of task completion rate; balance of employee load (standard deviation ≤ 15%).

[0064] If the simulation result task accumulation probability is >10%, the weight of subsequent tasks is increased according to the following formula: ; in, For the task The adjusted confidence weights, For the task The original confidence weight of " refers to the probability that a task is not completed on time in the simulation (for example, in 1000 simulations, task accumulation occurs 120 times, and the accumulation probability is 12%); "0.5" is the weight adjustment coefficient, which controls the adjustment range (experience value, balancing sensitivity and stability); "0.1" is the accumulation probability threshold (10%), exceeding this value triggers weight adjustment.

[0065] S3: Based on the mixed integer programming algorithm, real-time tasks are matched with virtual sequences in time and space to generate a dynamic scheduling solution with rolling updates.

[0066] S31: Establish an optimization model including task allocation variables, time window constraints and resource conflict detection; specifically: Decision variable definition, binary variable is , indicating the task Is it assigned to employees? ; Continuous variables , indicating the task The start time of the

[0067] Minimize the total cost, including task delay penalty, employee movement energy consumption and resource conflict cost. The objective function is: ; in, They are task delay penalty weight, mobile energy consumption weight and resource conflict penalty weight respectively; For the task The actual time when execution started. For the task Required deadlines; is the moving distance (meters) from the employee's current location to the task; For the task With the task Resource conflict variable (conflict = 1, otherwise = 0); is the total number of tasks, Index for tasks; is the total number of employees, Indexing employees; Is the conflicting task index, indicating Conflicting other tasks.

[0068] The constraint condition is set as: time window constraint, the task start time must be within the valid range: ; in, For the task The earliest possible start time, For the task Latest possible start time; Employee competence matching: tasks can only be assigned to employees with corresponding skills; Resource exclusivity: the same equipment (such as a cleaning vehicle) cannot perform multiple tasks at the same time.

[0069] Set trigger conditions to dynamically solve the model and prioritize high-confidence task allocation. Specifically: The trigger condition is that when any of the following conditions is met, task matching optimization is started: New tasks arrive and their priority is ≥ level 2; The actual completion time of the task deviates from the forecast by >15%; The employee's status is abnormal (the employee has been stationary for more than 10 minutes).

[0070] Use the Branch and Bound (BnB) method to solve the model: Set the maximum solution time to 30 seconds. If the timeout exceeds the limit, switch to heuristic rule solution. Prioritize high confidence tasks ( ≥0.7); assign employees nearby (moving distance ≤ 50m).

[0071] The branch and bound method is combined with heuristic rules to achieve efficient solution. Specifically: Branch and Bound (BnB): uses depth-first search (DFS) + pruning optimization to recursively solve the optimal task allocation solution; combined with "Lagrangian Relaxation" to accelerate convergence.

[0072] Heuristic solution rules (when BnB times out): Based on Greedy Matching, tasks are sorted by the minimum waiting time; employees are assigned according to the shortest moving path.

[0073] Based on Simulated Annealing (SA): Set the initial temperature ,Iteratively optimize the task exchange scheme; Objective Function for: ; In this formula, and is a weight parameter that controls the impact of task delay and moving distance.

[0074] S32: Divide the scheduling time window and define the task adjustment strategy; specifically: The scheduling period is set to 15-minute windows, and each window performs the following operations: Task freezing area (tasks in the next 5 minutes cannot be adjusted); optimizable area (tasks in the next 5-30 minutes can be reallocated); prediction area (tasks after 30 minutes are only virtually reserved).

[0075] The update trigger logic is to globally refresh the scheduling plan every 15 minutes; when the task queue change rate is >20% (compared to the upper window), rescheduling is triggered immediately.

[0076] Dynamically balance employee load and task priorities to optimize resource allocation; specifically: Calculate the standard deviation of employee load. If it exceeds the threshold (default 15%), adjust the task allocation according to the following formula: ; in, is the average load; For employees Current load, here Used to traverse all employees and calculate the load of all employees to balance the task distribution.

[0077] Tasks that meet the following conditions will have their priority increased: Waiting time ; Customer level ≥ member.

[0078] Generate visual navigation instructions and push them to the execution terminal in real time, specifically: Generate optimal movement path and comprehensive elevator waiting time model: ; in, is the total waiting time for the elevator, is the number of elevator stops; "8" is the average time consumed for each elevator stop (in seconds), including door opening, door closing, acceleration / deceleration time. The average time consumed for a single stop of the hotel elevator is 8 seconds (including the time for passengers to enter and exit). "60" is the coefficient for converting seconds to minutes (1 minute = 60 seconds).

[0079] Real-time navigation instructions are displayed through employee smart terminals; path update frequency = 1Hz to ensure accurate navigation information.

[0080] S4: Track task execution trajectories and calculate prediction deviations, dynamically modify model parameters and cross-departmental resource coordination rules.

[0081] S41: Collect employee location data and task execution records to build a spatiotemporal trajectory dataset; specifically: The UWB (ultra-wideband) positioning system records the three-dimensional coordinates of employees at a frequency of 10Hz; Combined with the elevator control system, the time of moving across floors is recorded to construct a complete movement trajectory dataset.

[0082] The following data is recorded for each task in the task ticket system: Plan time: ; The actual time is: (Time stamp of employee clocking in and out) Resource consumption (such as detergent usage, number of linen changes): automatically counted through RFID shelves.

[0083] Calculate the deviation between the actual task execution time and the predicted time and the path matching degree. Specifically: Calculate the absolute time deviation for each task : ; Aggregate statistics of task types: mean average error (MAE) and standard deviation : ; ; Use the Dynamic Time Warping (DTW) algorithm to match the actual movement path of employees with the theoretical scheduling path: ; In this formula, is the optimal bending path, is the number of path points; The actual path Position coordinates, The theoretical scheduling path Position coordinates.

[0084] Set the deviation threshold to trigger an alarm and start the root cause analysis process. Specifically: The alarm is triggered when any of the following conditions is met: Single-task time deviation > 30 minutes and priority ≥ level 2; The MAE of the same type of tasks exceeded the historical average by 20% for three consecutive times; Path similarity < 65% (DTW error exceeds the set threshold).

[0085] When a deviation alarm is triggered, the system will perform root cause analysis based on predefined diagnostic rules. The specific process is as follows: First check if there is a time skew: If time deviation exists, the employee's movement path is further checked to determine whether the employee completes the task according to the optimal path.

[0086] If time deviation does not exist, continue to check other possible influencing factors.

[0087] If there is no abnormality in the moving path: Continue to check resource consumption to determine whether there is any resource usage beyond the normal range during task execution (for example, abnormal consumption of cleaning supplies, linens, etc.).

[0088] If resource consumption is normal: Further check the employee skill match and analyze whether the task is delayed due to insufficient or unskilled employees.

[0089] S42: Incrementally train the prediction model according to the deviation data and adjust the network parameters in layers; specifically: The parameter adjustment strategy is: when MAE>20%: trigger fine-tuning of all parameters of the LSTM model (learning rate reduced to 0.0001). When MAE ≤ 20% but > 15%: only adjust the parameters of the attention mechanism layer. A / B test of the new model version: verify the model update effect in a 50% traffic environment.

[0090] Optimize inter-departmental task coordination agreements and use reinforcement learning to update coordination rules. Specifically: Cross-departmental collaborative optimization is triggered when the following scenarios occur: the engineering department's maintenance task has not started for >25 minutes; the conflict rate between the catering department's meal delivery task and the room service task is >30%.

[0091] Using reinforcement learning (PPO algorithm) to optimize rules: ; in, Assign strategy parameters to the new version of the task, Assign strategy parameters to old versions of tasks, is the learning rate; is the cumulative reward function, which means that in the state Take action The reward value obtained after , and sum it up for the entire task sequence; is the time step The system status at the moment, is the time step Moment-by-moment decision making; The reward function considers: task connection time, conflict rate, customer rating, etc.

[0092] The stability of the update solution is ensured through simulation verification and grayscale release. Specifically: Simulate the new model in the digital twin system and run it for 24 hours: Task completion rate increased by >5%; The time deviation of task connection between departments was reduced by >15%.

[0093] First test on 10% of the guest rooms and employees for 48 hours; if there are no major failures, the full system will be launched online.

[0094] Embodiment 2 is the second embodiment of the present invention, which is different from the previous embodiment in that: If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention can be essentially or partly embodied in the form of a software product that contributes to the prior art or partly in the current technical solution. The current computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

[0095] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0096] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0097] Embodiment 3 is an embodiment of the present invention, which provides a hotel guest demand task management system, including a multi-source data fusion collection module, an intelligent demand prediction module, a dynamic scheduling optimization module and an execution feedback self-optimization module.

[0098] Multi-source data fusion acquisition module: collects guest room sensor data, employee location data and external environment data, extracts spatiotemporal correlation features to construct a multi-source heterogeneous data set.

[0099] Intelligent demand forecasting module: Use long short-term memory network to predict the distribution of task demand in future periods and generate virtual task sequences with probability weights.

[0100] Dynamic scheduling optimization module: Based on the mixed integer programming algorithm, real-time tasks and virtual sequences are matched in time and space to generate a dynamic scheduling solution with rolling updates.

[0101] Execution feedback self-optimization module: tracks task execution trajectory and calculates prediction deviation, dynamically corrects model parameters and cross-departmental resource coordination rules.

[0102] Example 4 is an embodiment of the present invention, which provides a hotel guest demand task management method. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation / comparative experiments.

[0103] The test object of this embodiment is a five-star hotel with 400 rooms and 150 employees. The test period is from August 1 to October 31, 2024, and the performance of the system of the present invention in a real production environment is verified in three stages.

[0104] Guest Room Sensors: Temperature and humidity sensors are installed in all guest rooms, collecting data every 5 minutes (accuracy ±0.2℃, ±2%RH); The door magnetic sensor records the door status with a timestamp accuracy of ±0.5 seconds; Minibar smart shelves monitor item consumption in real time with an error of ±3g.

[0105] Employee Positioning: UWB positioning tags cover all employees, and base stations are deployed every 20 meters in the corridor, with a positioning accuracy of ±15cm; The elevator system integrates a real-time floor scheduling interface.

[0106] External data: API of the meteorological bureau in the region (probability of rainfall in the next 6 hours, temperature); Amap real-time traffic congestion index (2km road network around the hotel); Local large-scale event data (ICal schedule of the convention and exhibition center).

[0107] LSTM prediction model: Input features: 72-dimensional space-time vector (including room activity, employee heat density, and external event coefficient); Training data: historical task records from August 2023 to July 2024 (128,000 tasks in total for cleaning, delivery, and maintenance); Incremental updates: Automatically executed every 4 hours, training takes ≤ 8 minutes (NVIDIA A10G cluster).

[0108] Mixed Integer Programming Engine: Optimization goal: minimize the task delay cost (weight 0.6) + employee movement energy consumption (weight 0.4); Constraints: skill matching, equipment exclusivity, task time window (±10 minutes tolerance); Solver: CPLEX 22.1, timeout threshold 45 seconds, after which the heuristic nearest allocation rule is used.

[0109] Baseline data collection (August 1-August 14): Run the traditional manual scheduling system to record indicators such as task response time and resource utilization; calibrate sensors and positioning systems to ensure that the data error rate is <1.5%.

[0110] System trial run (August 15-September 30): Generate task forecasts for the next 18 hours at 5:00 every day, and output virtual sequences (confidence threshold 0.4); update plans every 20 minutes, and push AR navigation instructions to employees' smart watches (Garmin Instinct 2); sudden ward maintenance tasks (such as water pipe rupture) automatically trigger cross-departmental collaboration (Engineering Department + Housekeeping Department).

[0111] Optimization iteration (October 1-October 31): Based on the data from the first 6 weeks, the model parameters were adjusted: the number of LSTM attention heads increased from 4 to 6 to improve the external event modeling capability; the scheduling model added the "VIP customer task priority" constraint (weight coefficient +0.3). Reinforcement learning updated department collaboration rules (12 iterations), and the connection delay decreased by 23%.

[0112] The experimental reference data is shown in Table 1.

[0113] Table 1 Experimental data records

[0114] Experimental data show that the cleaning task prediction accuracy of the present invention is improved by 34.2%: thanks to the correlation modeling between external events (such as a meeting of 10,000 people at the convention and exhibition center) and guest room activity, the prediction error of the check-out cleaning task on rainy days is reduced from 42% to 11%; the peak response time of the delivery task is compressed to 5.1 minutes (originally 21.5 minutes), and elevator congestion is avoided in real time due to dynamic scheduling (elevator waiting time during peak hours is reduced by 68%).

[0115] The utilization rate of maintenance resources increased by 77.3%: the mixed integer programming model accurately avoided equipment conflicts (for example, the same vacuum cleaner was only assigned to adjacent floors within 30 minutes), and the equipment idling rate dropped from 24.7% to 9.3%; the monthly energy consumption cost dropped by 40.2%: employee path optimization reduced invalid movement (the average daily walking distance dropped from 8.2 kilometers to 3.1 kilometers), and the lighting and air-conditioning energy consumption was reduced simultaneously.

[0116] Customer satisfaction reached 91.6%: the system proactively identified VIP customers (accounting for 15%), and the response time for their delivery tasks was further compressed to 2.3 minutes, and the complaint rate was close to 0; the delay in cross-departmental collaboration was reduced by 62.5%: after reinforcement learning optimization, the connection time from the completion of maintenance by the engineering department to the start of cleaning by the housekeeping department was shortened from 25.1 minutes to 9.4 minutes, and the average weekly maintenance cost was reduced by 62% (16,000 yuan vs. 42,000 yuan).

[0117] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A hotel guest demand task management method, characterized in that: include: Collect guest room sensor data, employee location data, and external environment data, extract spatiotemporal correlation features, and construct a multi-source heterogeneous data set; Use long short-term memory networks to predict the distribution of task requirements in future periods and generate virtual task sequences with probability weights; Based on the mixed integer programming algorithm, real-time tasks and virtual sequences are matched in time and space to generate a dynamic scheduling solution with rolling updates; Track task execution trajectories and calculate prediction deviations, dynamically correct model parameters and cross-departmental resource coordination rules.

2. The hotel guest demand task management method according to claim 1, characterized in that: Collecting guest room sensor data includes: deploying temperature and humidity sensors in guest rooms, periodically collecting environmental data and transmitting them via wireless communication protocols; Install door magnetic sensors to monitor the open and closed status of guest room doors and the corresponding timestamps; Setting a weight sensor to monitor weight changes of items in the minibar and generating a raw weight data stream; Collecting employee location data includes: configuring ultra-wideband location tags for employees and obtaining real-time three-dimensional coordinate data through deployed positioning base stations; Linked with the elevator control system to record the target information of employees moving across floors; The collection of external environmental data includes: accessing the meteorological data interface to obtain the rainfall and temperature forecast values ​​for the future period; Obtain real-time traffic congestion level data around the hotel; Parse the local event calendar file and extract the number of participants and time information.

3. The hotel guest demand task management method as claimed in claim 2, characterized in that: The extracting of spatiotemporal correlation features to construct a multi-source heterogeneous data set includes converting the timestamps of the guest room sensor data, employee location data, and external environment data to a time axis in the same time zone, and using an interpolation method to complete the missing data; Map employee location coordinates and guest room locations to a unified floor plan coordinate system; Aggregate sensor data by room number to form a time series dataset; Group employee location data into movement trajectory data sets, recording time, coordinates, and floor information; Parse the unstructured text of event calendars to extract event types, number of participants, and time ranges; Calculate sliding window statistics in the time dimension for room data, including operation frequency and consumption rate; Generate a spatial service density heat map based on employee trajectories to quantify the service coverage intensity of each area; Convert time series data into image features and input them into the pre-trained model to extract spatiotemporal correlation feature vectors; Detect logical conflicts between multi-source data and trigger manual review processes; Abnormal data beyond a reasonable range are eliminated, and valid data are retained to construct the final data set.

4. The hotel guest demand task management method as claimed in claim 3, characterized in that: The method of using the long short-term memory network to predict the distribution of task requirements in the future period includes constructing a long short-term memory network model, inputting a spatiotemporal correlation feature vector, and outputting a probability distribution of each task type in the future period; Regularly train the model incrementally and monitor prediction deviations, triggering emergency retraining to update model parameters; Identify the peak period of the task volume curve, and output the time window and task type distribution.

5. The hotel guest demand task management method as claimed in claim 4, characterized in that: The generating of a virtual task sequence with probability weights includes: generating a virtual task timestamp sequence based on a predicted task arrival rate, binding task types and confidence weights; Dynamically insert real-time tasks and remove low-confidence conflicting tasks; Pre-allocate resources and simulate to verify the feasibility of task sequences, and adjust weights to optimize execution efficiency.

6. The hotel guest demand task management method according to claim 5, characterized in that: The said performing spatiotemporal matching of real-time tasks and virtual sequences based on a mixed integer programming algorithm includes establishing an optimization model including task allocation variables, time window constraints and resource conflict detection; Set trigger conditions to dynamically solve the model and prioritize high-confidence task allocation; The branch and bound method is combined with heuristic rules to achieve efficient solution.

7. The hotel guest demand task management method according to claim 6, characterized in that: The generating of the dynamic scheduling scheme for rolling updates includes dividing the scheduling time window and defining the task adjustment strategy; Dynamically balance employee load and task priorities to optimize resource allocation; Generate visual navigation instructions and push them to the execution terminal in real time.

8. The hotel guest demand task management method according to claim 7, characterized in that: Tracking the task execution trajectory and calculating the prediction deviation includes collecting employee location data and task execution records to construct a spatiotemporal trajectory data set; Calculate the deviation between the actual task execution time and the predicted time and the path matching degree; Set deviation thresholds to trigger alerts and initiate the root cause analysis process.

9. The hotel guest demand task management method according to claim 8, characterized in that: The dynamic correction of model parameters and cross-departmental resource coordination rules include training the prediction model according to the deviation data increment and adjusting the network parameters in layers; Optimize inter-departmental task coordination agreements and use reinforcement learning to update coordination rules; The stability of the update solution is ensured through simulation verification and grayscale release.

10. A hotel guest demand task management system, used to implement the hotel guest demand task management method as claimed in any one of claims 1 to 9, characterized in that: include: Multi-source data fusion acquisition module: collects guest room sensor data, employee location data and external environment data, extracts spatiotemporal correlation features and constructs multi-source heterogeneous data sets; Intelligent demand forecasting module: uses long short-term memory network to predict the distribution of task demand in future periods and generate virtual task sequences with probability weights; Dynamic scheduling optimization module: Based on the mixed integer programming algorithm, real-time tasks and virtual sequences are matched in time and space to generate a dynamic scheduling solution with rolling updates; Execution feedback self-optimization module: tracks task execution trajectory and calculates prediction deviation, dynamically corrects model parameters and cross-departmental resource coordination rules.

Images