Dynamic scheduling method and system for urban garbage classification collection and transportation vehicles

Abstract

The invention relates to the technical field of urban environmental sanitation intelligent scheduling, and discloses an urban garbage classification collection and transportation vehicle dynamic scheduling method and system, and the method comprises the steps: obtaining historical daily output data from a weighing sensor of an intelligent garbage can, and constructing a multi-dimensional feature matrix in combination with calendar features and meteorological data; predicting the next-day four-class garbage yield of each collection point through a gradient lifting regression model; path planning is carried out by adopting an improved saving algorithm introducing a load utilization rate reward item, and MTZ constraint and load capacity double constraint are applied; an alarm collection point is dynamically inserted into a nearest in-transit vehicle path through GPS real-time monitoring when an overflow alarm is given; counting effect indexes every day and feeding back and updating prediction model parameters.

Description

Technical Field

[0001] This invention relates to the field of intelligent scheduling technology for urban sanitation, and in particular to a method and system for dynamic scheduling of urban waste sorting and collection vehicles based on waste production prediction and dynamic optimization of vehicle routes. Background Technology

[0002] With the accelerating pace of urbanization, the amount of urban household waste generated is increasing year by year. According to data released by the National Bureau of Statistics, the amount of urban household waste collected in my country exceeded 240 million tons in 2019, and has maintained an annual growth rate of about 10% in recent years. Under the policy background of my country's mandatory implementation of the garbage classification system, household waste is classified and collected in four categories: kitchen waste, recyclables, hazardous waste, and other waste. This places more stringent requirements on the vehicle dispatching capacity in the garbage collection and transportation process. Each type of waste must be transported independently by corresponding dedicated vehicles. Different types of vehicles differ in terms of load limits, volume limits, and operating times, constituting a typical heterogeneous vehicle fleet dispatching problem.

[0003] Most existing urban waste collection and transportation scheduling schemes are based on a static scheduling model with fixed routes and fixed timetables. This means that the daily routes and arrival times are planned once and remain unchanged over a long period, based on the historical average waste volume at each collection point within the region. For example, Chinese invention application CN114386792A discloses a method for scheduling urban household waste collection vehicles based on a genetic algorithm. This method collects a basic dataset within the service area, calculates the average daily volume of different types of waste using historical waste volume data to predict future output, then uses a K-means clustering algorithm to cluster and partition the collection points, and finally establishes a waste reverse logistics network planning model with the goal of minimizing costs and solves it using a genetic algorithm. This scheme achieves some degree of optimized calculation of collection routes, but it has the following shortcomings.

[0004] First, in the waste production prediction stage, this scheme only estimates future daily waste production by calculating the simple arithmetic mean of historical data, failing to consider the significant impact of external factors such as seasonal changes, weather variations, and holiday effects on waste generation behavior. In practice, waste production during hot summer days and long holidays such as the Spring Festival is often several times higher than normal levels. The simple averaging method is unable to capture such fluctuations, resulting in low prediction accuracy. Second, in the route planning stage, this scheme uses a genetic algorithm to solve the reverse logistics network planning model. Although it can find routes close to the global optimum, it does not provide targeted optimization guidance for vehicle load utilization. The scheduling results often show an imbalance where some vehicles are almost fully loaded while others are severely empty. The empty load rate can exceed 40% on days with low waste production, causing significant waste of transport capacity and fuel consumption. Third, this scheme is a typical day-ahead one-time planning mode, lacking the ability to dynamically adjust routes based on unforeseen events during daily operation. When a collection point generates far more waste than predicted due to unforeseen circumstances such as community events or merchant promotions, it can only wait until the next operating cycle to collect the waste, during which time overflowing bins, environmental pollution, and resident complaints are highly likely to occur. Fourth, the scheme lacks a quantitative evaluation and feedback mechanism for scheduling effectiveness, and cannot form a continuous optimization closed loop of prediction-planning-execution-evaluation-improvement.

[0005] Therefore, there is an urgent need for a dynamic scheduling method for urban waste sorting and collection vehicles that can comprehensively consider external influencing factors to accurately predict waste production, proactively improve load utilization in route planning, dynamically respond to sudden demands during operation, and conduct effect evaluation and model updates after daily scheduling. From an operations research perspective, this problem is essentially a complex optimization problem that deeply couples demand forecasting, heterogeneous fleet route optimization, and online dynamic replanning. Existing literature has not yet seen a complete solution that integrates the above three levels of methods into a unified framework and forms a closed-loop feedback mechanism. Summary of the Invention

[0006] To address the technical problem of mismatch between static scheduling schemes and dynamic demands in existing technologies, this invention provides a method and system for dynamic scheduling of urban waste sorting and collection vehicles.

[0007] Specifically, this invention discloses a method for dynamic scheduling of urban waste sorting and collection vehicles, comprising the following steps: Step S1, Multi-source data acquisition and feature construction steps: Obtain historical daily output data from the weighing sensors of the smart trash cans deployed at each collection point, and simultaneously collect calendar feature data and meteorological data for the corresponding time period. After aligning the historical daily output data, calendar feature data, and meteorological data according to the timestamp, construct a multi-dimensional feature matrix. The calendar feature data includes the week number, whether it is a statutory holiday, and the month. The meteorological data includes the average daily temperature, daily precipitation, and average daily wind speed.

[0008] Step S2, waste production prediction step: Using a multidimensional feature matrix as input, establish gradient boosting regression models for four types of waste: kitchen waste, recyclables, hazardous waste, and other waste. Four independent prediction sub-models are obtained by training on historical data. The trained gradient boosting regression models are then used to input the calendar features and weather forecast data for the next day, and output the predicted production of the four types of waste at each collection point for the next day.

[0009] Step S3, route planning step with enhanced load utilization: The predicted output of the four types of waste at each collection point is input into the improved saving algorithm for vehicle route planning of heterogeneous fleets. The improved saving algorithm introduces a load utilization bonus item on the basis of the classic saving value calculation. When the load utilization of the vehicle after merging two candidate paths exceeds the preset load utilization threshold, the bonus value is added to the merged saving value to improve the merging priority. At the same time, MTZ constraints are applied to eliminate sub-loops and the load limit constraints and volume limit constraints of each type of vehicle, and the initial scheduling route scheme of each type of vehicle is output.

[0010] Step S4, Real-time Monitoring and Dynamic Route Re-optimization: Each vehicle performs its collection task according to the initial scheduling route plan output in Step S3. During the day's operation, the real-time location and task completion progress of each vehicle are tracked through the vehicle's GPS. When a collection point generates more garbage than predicted due to a sudden event and triggers an overflow alarm signal, the alarm collection point is automatically inserted into the route of the nearest on-the-way vehicle with sufficient remaining load capacity. The re-optimization process only adjusts the access order of the on-the-way vehicle to subsequent unvisited nodes without affecting the completed route segments.

[0011] Step S5, scheduling effect evaluation and parameter feedback: After the daily scheduling is completed, the system automatically calculates the actual load utilization rate, total travel distance and total transportation cost. The statistical results are compared with the corresponding indicators of the static fixed route plan to generate a daily report on optimization effect. The actual output data of each collection point on the day is used as a new training sample and fed back to the gradient boosting regression model in step S2 to update the model parameters.

[0012] This invention also provides a dynamic scheduling system for urban waste sorting and collection vehicles, comprising: a multi-source data acquisition and feature construction module for acquiring historical daily output data, calendar feature data, and meteorological data and constructing a multi-dimensional feature matrix; a waste sorting output prediction module for outputting the predicted output of four types of waste at each collection point on the following day based on a gradient boosting regression model; a route planning module with enhanced load utilization for vehicle route planning using an improved saving algorithm; a real-time monitoring and dynamic route re-optimization module for tracking vehicles via onboard GPS and dynamically re-optimizing routes when an overflow alarm is triggered; and a scheduling effect evaluation and parameter feedback module for generating a daily report of optimization effects and feeding back actual output data to the prediction module.

[0013] The beneficial effects of adopting the above technical solution are as follows: By combining the gradient boosting regression model with calendar features and meteorological data, it is possible to capture the nonlinear impact of factors such as seasons, weather, and holidays on waste production, significantly improving prediction accuracy; By introducing a load utilization rate reward term into the saving algorithm, the path planning algorithm tends to generate routes with high load utilization rates, thereby effectively reducing the total number of vehicles dispatched and the empty driving distance; Through the dynamic path re-optimization mechanism triggered by the overflow alarm, a rapid response to sudden demands during operation is achieved without affecting already executed path segments; Through daily effect evaluation and model parameter feedback, a closed-loop continuous optimization is formed, enabling the system to gradually improve prediction accuracy and scheduling efficiency in long-term operation. Attached Figure Description

[0014] Figure 1 This is a flowchart of the dynamic scheduling method for urban waste sorting and collection vehicles according to the present invention.

[0015] Figure 2 This is an architecture diagram of the urban waste sorting and collection vehicle dynamic scheduling system of the present invention. Detailed Implementation

[0016] The technical solution of the present invention will now be clearly and completely described in conjunction with the accompanying drawings. It should be noted that the specific parameters and numerical ranges given in the following embodiments are merely illustrative examples to facilitate understanding of the technical solution of the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0017] like Figure 1As shown, the dynamic scheduling method for urban waste sorting and collection vehicles provided by this invention includes steps S1 to S5, forming a closed-loop collaborative architecture of data collection → output prediction → route planning → dynamic re-optimization → effect evaluation feedback. The actual output data output in step S5 is fed back to the prediction model in step S2 for parameter updates, and the load utilization rate deviation output in step S5 is fed back to the route planning algorithm in step S3 for reward parameter calibration, thus forming a closed loop of dynamic adjustment of parameters of the preceding steps by the subsequent steps. The specific implementation methods of each step are described in detail below.

[0018] Step S1: Multi-source data acquisition and feature construction. The data acquisition process of this invention involves the collaborative access and fusion processing of three types of heterogeneous data sources, as detailed below.

[0019] The first type of data source is the real-time weighing data reported by the built-in weighing sensors of the smart trash cans deployed at each collection point. In one embodiment of the present invention, each collection point is equipped with at least four smart trash cans corresponding to four categories: kitchen waste, recyclables, hazardous waste, and other waste. Each smart trash can is equipped with a pressure-type weighing sensor at the bottom with a range of 0 to 200 kg and an accuracy of not less than 0.5 kg. The weighing sensor uploads the current trash mass data in the can to the data server via NB-IoT narrowband Internet of Things at a sampling period of 15 minutes. The data server takes the maximum value of all sampled data before 24:00 each day and subtracts the final value of the previous day to obtain the net output of that type of waste at the collection point for that day. Preferably, to eliminate the interference of sudden drops in the weight of the trash cans caused by trash collection operations on the calculation of net output, the present invention first performs collection event detection on the original time series data when calculating the daily output—when the absolute value of the difference between two adjacent sampling values ​​exceeds 30 kg, the moment is marked as the collection moment, and the increments before and after collection are accumulated in segments when calculating the daily output. After the above processing, a historical daily output data sequence of four types of waste for each collection point for no less than 90 consecutive days was obtained.

[0020] The second type of data source is calendar feature data. This invention obtains the calendar attributes of each day within the time period corresponding to historical daily production data from a public calendar interface, including three dimensions of features: the weekday number uses an integer encoding from 1 to 7, where 1 represents Monday and 7 represents Sunday; whether it is a statutory holiday uses a binary encoding of 0 and 1, where 0 represents a weekday and 1 represents a statutory holiday and its adjusted workday; and the month uses an integer encoding from 1 to 12. In a preferred embodiment of this invention, whether it is the last 3 days of the month (binary encoding) and whether it is a day of a large public event (binary encoding) are further introduced as auxiliary calendar features to capture the impact of the end-of-month settlement period and large-scale events on waste production.

[0021] The third type of data source is meteorological data. This invention obtains daily meteorological observation data for each time period corresponding to historical daily production data from the meteorological department's open data interface. This data includes three dimensions: daily average temperature (°C) with an accuracy of 0.1°C; daily precipitation (mm) with an accuracy of 0.1mm; and daily average wind speed (m / s) with an accuracy of 0.1m / s. The selection of meteorological data is based on the following: daily average temperature directly affects the rate of food spoilage and thus the amount of kitchen waste; studies show that when the daily average temperature exceeds 30°C in summer, the amount of kitchen waste increases by 25% to 40% compared to the spring and autumn averages; daily precipitation affects residents' shopping and dining behaviors, thus affecting the amount of recyclables and kitchen waste; and daily average wind speed affects the storage and collection efficiency of lightweight recyclables.

[0022] After acquiring the above three types of data, this invention aligns the data sources according to timestamps to construct a multi-dimensional feature matrix. Let the total number of collection points be... The number of historical days is Then for the first Collection points ( ) and the sky( ), feature vector Defined as: , in: For the first The day of the week number, which is an integer from 1 to 7; For the first The day is marked as a holiday, with a value of 0 or 1. For the first The month in which the day falls takes the value of an integer from 1 to 12; For the first The average daily temperature of the day, expressed in °C; For the first Daily rainfall, in mm; For the first The average daily wind speed, in m / s. The corresponding label indicates the actual daily output of a certain type of waste at that collection point on that day. ,in These represent kitchen waste, recyclables, hazardous waste, and other waste, respectively, with the unit being kg.

[0023] Before constructing the feature matrix, this invention also performs outlier detection and repair on the historical daily output data. Specifically, the upper quartiles are calculated for the daily output sequence of each type of waste at each collection point. and lower quartiles Interquartile range Daily production will exceed or below Data is marked as outliers and replaced with the average daily output of one normal day before and after the outlier date. The technical advantage of outlier detection is that it avoids polluting the model training process with abnormal data caused by sensor failures, data transmission errors, or extreme emergencies, thereby improving the model's robustness and generalization ability. In one embodiment of the invention, after outlier detection, the proportion of outliers in 90 days of historical data from 120 collection points in a certain urban area was approximately 2.3%, and the replaced data sequence was smoother and retained normal fluctuation characteristics.

[0024] Furthermore, after outlier repair, this invention performs standardization preprocessing on the feature matrix. For continuous features (daily average temperature, daily precipitation, and daily average wind speed), z-score standardization is used, subtracting the mean from each feature variable and dividing by its standard deviation to ensure the mean of the feature value distribution is 0 and the standard deviation is 1, thus eliminating the impact of dimensional differences between different features on subsequent model training. For discrete calendar features (weekday number and month), since they inherently have ordinal meaning, they are directly encoded as raw integers and input into the model without standardization. The above data preprocessing operations are automatically executed every morning, with the entire processing delay of step S1 not exceeding 5 minutes, providing timely and reliable feature input for yield prediction in the subsequent step S2.

[0025] Step S2: Waste Generation Prediction. The core of this step lies in using a Gradient Boosting Regression Tree (GBRT) model to accurately predict the generation of the four types of waste at each collection point the following day. The technical basis for choosing the Gradient Boosting Regression Tree model is as follows: Compared with traditional arithmetic mean or linear regression methods, the Gradient Boosting Regression Tree model can automatically capture the nonlinear mapping relationship between feature variables and target variables, and has significant advantages in dealing with the interaction of multiple factors such as seasonal changes, holiday effects, and weather fluctuations; Compared with deep neural networks, the Gradient Boosting Regression Tree model has a faster training speed and lower overfitting risk in scenarios with a moderate sample size (hundreds to thousands of records), and has a good ability to interpret feature importance, making it easier for maintenance personnel to understand the basis of the model's predictions.

[0026] This invention establishes four independent gradient boosting regression prediction sub-models for four types of waste, respectively. The predictive sub-model for garbage is denoted as , Kitchen waste ( For example, the training process of the model is as follows: (The rest of the text appears to be a list of steps or instructions, and doesn't translate directly.) Collection points Feature-label data pairs for each day The dataset is aggregated into a training set, which is then randomly divided into a training subset and a validation subset in an 8:2 ratio. Training employs a forward stepwise addition model framework, with the number of iterations set to [value missing]. In each round, a regression decision tree is fitted in the current residual direction. The output of the wheel is: , in: For the first The cumulative prediction function after round; For the first The cumulative prediction function after round; The learning rate controls the proportion of each tree's contribution to the final prediction, and its value ranges from 0.01 to 0.1. In one embodiment of the present invention, it is set to... ; For the first The regression decision tree fitted in round-robin fashion. The maximum depth of each regression decision tree. Set to an integer value between 4 and 8; in one embodiment of the present invention, the value is taken as... This depth achieves a good balance between model complexity and generalization ability. (Number of iterations) The number of rounds is set to 100 to 500, and in one embodiment of the present invention, it is taken as... During training, Mean Squared Error (MSE) is used as the loss function. , in: The value is the loss function value, which is dimensionless. For training subsets; The total number of samples in the training subset; This represents the actual daily output, expressed in kg. The value is the model's predicted value, in kg. The negative gradient of the loss function with respect to the model's predicted value yields the pseudo-residual. Each new decision tree in each round fits this pseudo-residual, thus achieving gradient descent optimization.

[0027] To prevent overfitting, this invention introduces the following regularization strategies: the minimum number of samples per leaf node in each tree is no less than 10; only 80% of the features in each tree are randomly selected for node splitting (column sampling); and only 80% of the samples are randomly selected in each training round to participate in tree construction (row sampling). The Mean Absolute Percentage Error (MAPE) is used to evaluate prediction accuracy on the validation subset. , in: To verify the subset; To verify the total number of samples in the subset, in one embodiment of the present invention, after training with 90 days of data from 120 collection points in a certain urban area, the MAPE of the four types of waste prediction sub-models on the validation set were as follows: kitchen waste 8.7%, recyclables 11.2%, hazardous waste 15.6%, and other waste 9.3%, which are far better than the MAPE values ​​corresponding to the traditional arithmetic mean method (22.4%, 28.6%, 35.1%, and 21.8%, respectively).

[0028] After the model is trained, when making predictions for the next day, the calendar features (week number, holiday markers, month) and weather forecast data (forecast daily average temperature, forecast daily precipitation, forecast daily average wind speed) of the next day are combined to form the feature vector for the next day. Input four prediction sub-models respectively, and output the predicted output of the four types of waste at each collection point the next day. , In a preferred embodiment of the present invention, when the MAPE is lower than a preset accuracy threshold (e.g., 15%), the model prediction value is automatically adopted; when the MAPE of a certain type of waste is higher than the accuracy threshold, the prediction value of that type of waste is automatically multiplied by a safety factor of 1.15 to increase the prediction margin and reduce the risk of waste overflowing due to underestimation.

[0029] Step S3: Route planning with enhanced load utilization. The core of this step is to transform the predicted output of the four types of waste at each collection point output in Step S2 into optimal collection and transportation route schemes for each type of vehicle. This invention uses an improved Clarke-Wright saving algorithm as the route planning engine. Compared with the genetic algorithm, the saving algorithm has the advantages of fast computation speed and stable solution quality, making it suitable for application scenarios that require rapid daily generation of scheduling schemes. The core improvement of this invention over the classic saving algorithm lies in the introduction of a load utilization reward term, which makes the algorithm tend to generate routes with high load utilization when merging routes, thereby effectively reducing the total number of vehicles dispatched.

[0030] First, let's explain the problem modeling for route planning. Let the transfer center be numbered 0, and the collection point numbered... The transfer center to the first The distance between the collection points is , No. The collection point to the first The distance between the collection points is The unit is km. For the... Waste ( Let the corresponding special vehicle type be... The maximum load capacity of the vehicle is (Unit: kg), maximum vehicle volume is (Unit is) ), No. The predicted output of each collection point is (Unit: kg), the The predicted volume for each collection point is: (Unit is) The predicted volume is based on the average density of various types of waste. (Unit is) The conversion yields: In one embodiment of the present invention, the average density of the four types of waste is as follows: kitchen waste recyclable materials Hazardous waste Other waste .

[0031] It should be noted that the route planning is independent for different types of waste. Kitchen waste, recyclables, hazardous waste, and other waste are collected by their respective dedicated fleets. This step executes the route planning process once for each of the four types of waste, outputting four independent scheduling route schemes. The route planning for each type of waste can be performed in parallel computationally without interference.

[0032] The following is the first The execution process of the improved saving algorithm is detailed using path planning for waste-like objects as an example.

[0033] Sub-step S3.1: Construct the initial solution. Starting from the transfer center, construct the initial solution for each collection point that needs to receive the goods. ( Construct one single-point round-trip route for each location. The total number of vehicles in the initial plan is equal to the number of collection points that need to be collected. In one embodiment of the present invention, for the food waste category, the predicted output of 108 out of 120 collection points is greater than 0, therefore the initial plan includes 108 single-point round-trip routes.

[0034] Sub-step S3.2: Calculate the enhanced savings value. For any two collection points... and ( Classic Savings Value The calculation formula is: , in: To collect points and The distance saved by merging two separate routes into one route, expressed in km; From the transfer center to the collection point The distance, in km; From the transfer center to the collection point The distance, in km; Collection point to collection point The distance is measured in kilometers. A higher savings value indicates a higher combined benefit.

[0035] Based on the classic savings value, this invention introduces a load utilization bonus. Let the total load capacity of the merged route be... This value equals the sum of the predicted outputs at all collection points along the route; assuming the maximum vehicle load capacity for this route is... The load utilization rate of the merged route Defined as: , in: The load utilization rate of the merged route is dimensionless and ranges from 0 to 1. The total load capacity of the merged route is expressed in kg. For the first The maximum load capacity for this type of vehicle, expressed in kg. Exceeding the preset load utilization threshold At the same time, enhance the value of saving. Add a bonus value to the classic savings value: , in: To enhance the savings value, the unit is km; The reward coefficient is dimensionless and ranges from 0.1 to 0.3. In one embodiment of the present invention, it is taken as... ; The load utilization rate threshold is dimensionless and ranges from 0.75 to 0.85. In one embodiment of the present invention, it is taken as... ; This is an indicator function; it takes the value 1 when the condition within the parentheses is true, and 0 otherwise. The technical effect of introducing a load utilization bonus is that when the vehicle load utilization rate increases from 50% to over 80% after merging two paths, the merging priority increases by an additional 20%, causing the algorithm to prioritize merging operations that fully utilize vehicle capacity, thereby effectively reducing empty runs and the total number of trips.

[0036] Sub-step S3.3: Sorting and Greedy Merging. Combine all candidate collection point pairs. According to the enhanced savings value Sort the candidates from largest to smallest to form a candidate merge list. Then, check each pair of candidate merges sequentially from the beginning of the sorting list to see if it satisfies the following constraints: Constraint 1 (Load Capacity Limit Constraint): The sum of the predicted output of all collection points on the merged route shall not exceed the vehicle load capacity limit. ,Right now .

[0037] Constraint 2 (Volume Limit Constraint): The sum of the predicted volumes of all collection points on the merged route shall not exceed the vehicle volume limit. ,Right now ,in This is the set of collection points along the merged route.

[0038] Constraint 3 (Endpoint Condition): Collection Point It must be the last collection point on its route (i.e., the node immediately following the return to the transfer center). It must be the first collection point on its route (i.e., the node immediately following the transfer center) to ensure that a valid single link is formed after merging.

[0039] For candidate merges that simultaneously meet the above three constraints, the merge will be executed, and the collection points will be... All nodes along the route are added to the collection point in sequence. At the end of the original route, a new route is formed, and the original collection points are deleted. The independent route in which it is located. Candidate merges that do not meet any of the constraints are skipped. Repeat the above checking process until the entire candidate merge list has been checked.

[0040] Sub-step S3.4: MTZ constraint eliminates sub-loops. In the path scheme after greedy merging, this invention further applies Miller-Tucker-Zemlin (MTZ) constraints to mathematically and rigorously eliminate possible sub-loops. An auxiliary variable is introduced. Indicates collection point The access order position in the path, for any two non-transfer center collection points in the path. and If the vehicle comes from Drive directly to (Right now , 0-1 decision variables indicate whether there exists a... arrive If the direct travel arc is given, then the following constraints are established: , in: Collection point Access order auxiliary variable, dimensionless, with a value range of 1 to 1. ( (Total number of collection points) Collection point Access order auxiliary variable; For 0-1 decision variables, when the vehicle departs from... Drive directly to hour ,otherwise The technical advantage of MTZ constraints is that if there are sub-loops in the path that do not pass through the transit center (e.g., ... Applying the above constraints to each arc on the sub-loop will produce... The contradiction of value -- It is impossible for both to be true simultaneously, thus mathematically eliminating the existence of sub-loops and ensuring that each generated path is a valid Hamiltonian path with the transfer center as both the starting and ending point.

[0041] In one embodiment of the present invention, the route planning results of 120 collection points in a certain urban area show that after adopting the improved saving algorithm, the number of vehicles dispatched for kitchen waste disposal decreased from 18 trips in the static plan to 14 trips, the average load utilization rate increased from 62.3% to 83.7%, and the total driving distance was reduced from 487km to 412km.

[0042] Step S4: Real-time monitoring and dynamic route re-optimization. The initial scheduling route plan output in step S3 is generated and distributed to the vehicle terminals of each vehicle in the early morning of the same day. During the operation of the day, this invention tracks the location and task completion progress of each vehicle in real time through the vehicle GPS positioning module, builds operational situation awareness capability, and provides decision support for dynamic route re-optimization.

[0043] In one embodiment of the present invention, the positioning accuracy of the vehicle-mounted GPS positioning module is ±5m, and the data reporting cycle is 30s. The dispatch server maintains a vehicle status table based on GPS data, recording the current geographical coordinates (longitude and latitude, accurate to 6 decimal places), the list of completed collection points, the next target collection point, the total mass of loaded garbage, and the remaining load capacity for each vehicle. When a vehicle arrives at a collection point and completes garbage loading, the vehicle-mounted terminal automatically reports the loading completion signal and the actual loaded mass to the dispatch server, which then updates the list of completed collection points and the total mass of loaded garbage in the vehicle status table accordingly.

[0044] The triggering mechanism for dynamic route re-optimization is as follows. Each smart trash can at a collection point is equipped with an overflow alarm threshold; in one embodiment of this invention, this threshold is set to 90% of the can's capacity. When a smart trash can at a collection point detects that the amount of trash inside has reached or exceeded 90% of its capacity, it automatically sends an overflow alarm signal to the scheduling server. The overflow alarm signal carries the alarm collection point number, the type of trash being alarmed, and the estimated mass of trash currently in the can. Preferably, the alarm signal also carries the geographical coordinates of the collection point to accelerate the subsequent distance calculation process.

[0045] After receiving the overflow alarm signal, the scheduling server executes the following dynamic insertion process.

[0046] Step 1: Filter candidate vehicles en route. From the vehicle status table, filter all vehicles en route that meet the following conditions: the vehicle type matches the dedicated vehicle type corresponding to the alarm garbage type; the vehicle is currently en route and has unvisited collection points; the vehicle has remaining load capacity. Estimated garbage mass greater than or equal to the alarm collection point Remaining load capacity The calculation formula is: , in: For the first The remaining load capacity of a vehicle en route, in kg; This is the maximum load capacity for the vehicle type, in kg. This represents the total mass of garbage currently loaded onto the vehicle, expressed in kg.

[0047] Step 2: Calculate distances and sort them. For all candidate vehicles on the road, calculate alarm collection points. Straight-line distance between the current positions of each candidate vehicle : , in: Alarm collection point With the The spherical distance between candidate vehicles, in km; The average radius of the Earth is taken as 6371 km; and Alarm collection points Latitude and longitude, in radians; and The first The latitude and longitude of each candidate vehicle's current location, in radians. All candidate vehicles are then sorted according to... Sort the vehicles from smallest to largest, and select the first-ranked candidate vehicle as the target vehicle, denoted as the [number]. A vehicle.

[0048] Step 3, minimum cost insertion. Insert alarm collection points. Insert target vehicle The position with the minimum cost in the subsequent unvisited node sequence. Let the target vehicle... The subsequent unvisited node sequence is ( (Number of unvisited nodes), in the node and Insert between Insertion cost for: , in: For in position Insert alarm collection point at location The increased driving distance is measured in km. For nodes to alarm collection point The distance, in km; Alarm collection point To the node The distance, in km; For nodes To the node The original distance, in km. Traverse all Select after choosing one insertable position. Minimum position alarm collection point Insert at this position to form a new sequence of subsequent nodes. .

[0049] It should be noted that dynamic route re-optimization only adjusts the access order of unvisited nodes for the target vehicle, without affecting completed route segments. The technical advantage of this design is that it avoids backtracking interference with already executed tasks, ensuring the monotonic progression and execution stability of the scheduling scheme. In one embodiment of this invention, if the remaining load capacity of all candidate vehicles en route does not meet the requirements of the alarm collection point, the system automatically dispatches one vehicle of the corresponding type from the standby vehicle pool directly to the alarm collection point to ensure that the overflow event is responded to within 2 hours.

[0050] Preferably, when multiple overflow alarm signals occur within a short period of time, the present invention processes them sequentially according to the order of alarm time. After processing each alarm, the remaining load capacity of the target vehicle and the subsequent node sequence in the vehicle status table are updated before processing the next alarm. This serial processing strategy ensures that each insertion decision is based on the latest vehicle status information, avoiding load overload conflicts that may occur due to parallel processing. In one embodiment of the present invention, when more than three overflow alarms occur within the same 5-minute window, the system automatically triggers a batch processing mode—first, all alarm collection points to be processed are sorted according to their urgency (from high to low bucket fill rate), and then dynamic insertion operations are performed sequentially according to the sorting results, prioritizing the response to the most urgent overflow points.

[0051] In a preferred embodiment of the present invention, the dynamic route re-optimization module also includes a built-in route quality assessment sub-function. After dynamic insertion is completed, the system automatically assesses the impact of the insertion operation on the target vehicle's route and calculates the ratio of the total distance increment of the inserted route to the original total route distance. When this ratio exceeds 25%, the system prompts the dispatching management personnel for manual confirmation to ensure that the dynamically adjusted route remains reasonable and efficient. In actual operation, because the algorithm prioritizes selecting the on-the-way vehicle closest to the alarm collection point and uses the minimum cost insertion method, more than 90% of the route increments caused by dynamic insertion operations do not exceed 10% of the original route distance, thus having a very limited impact on overall transportation efficiency.

[0052] Step S5: Scheduling effect evaluation and parameter feedback. This step is executed automatically after all daily scheduling tasks are completed. Its core functions include three aspects: scheduling indicator statistics, static scheme comparison, and prediction model parameter feedback.

[0053] Dispatch metric statistics. The system extracts the operation records of all train trips for the day from the vehicle status table and GPS trajectory database, and automatically calculates the following three core dispatch metrics.

[0054] The first indicator is the actual load utilization rate. : , in: The actual load utilization rate of all train services on that day is dimensionless and ranges from 0 to 1. This represents the total number of vehicles dispatched that day. For the first The actual total loaded mass of the vehicle when it finally returns to the transfer center, in kg; For the first The maximum rated load capacity of a vehicle, expressed in kg.

[0055] The second indicator is the total driving distance. : , in: The total distance traveled by all vehicles on that day, in km; According to the first The actual distance traveled by the vehicle as calculated by GPS trajectory, in km.

[0056] The third indicator is total transportation cost. : , in: This represents the total transportation cost for the day, expressed in yuan. This represents the total number of vehicles dispatched that day. The fixed scheduling cost per trip (including labor costs and vehicle depreciation) is expressed in yuan per trip. In one embodiment of the present invention, it is taken as... Yuan / train trip; Total distance traveled, in km; The fuel cost per unit distance is expressed in yuan / km. In one embodiment of the present invention, it is taken as... Yuan / km.

[0057] Static scheme comparison. The system maintains a static scheduling scheme based on fixed routes in the database as a comparison benchmark. This static scheme is planned once based on historical average output, and the route and vehicle configuration remain unchanged over a long period. The actual load utilization rate of the dynamic scheduling scheme on the day is compared. Total driving distance and total transportation costs Corresponding indicators to the static scheme respectively , and Compare the results and calculate the improvement rate for each indicator: , , , in: The load utilization improvement rate is represented by a positive value, indicating that the dynamic solution is better than the static solution. For the distance saving rate, a positive value indicates that the dynamic solution covers a shorter distance; The transportation cost saving rate is represented by a positive value, indicating that the dynamic solution has a lower cost. The system automatically generates a daily optimization performance report after summarizing the above indicators, presenting the daily and historical cumulative optimization effects in the form of tables and trend charts. In one embodiment of the invention, performance evaluation data from 30 consecutive days of operation shows that, compared to the static solution, the dynamic scheduling solution increases load utilization by an average of 18.5 percentage points (from 61.2% to 79.7%), saves an average of 15.3% in total travel distance, and saves an average of 21.7% in total transportation costs.

[0058] Predictive model parameter feedback. After the daily scheduling is completed, the system adds the actual output data of the four types of waste at each collection point (derived from the weighing data of the vehicle weighing system and the weighing data of the smart trash cans before the last collection) as new training samples to the training dataset in step S2. An incremental learning strategy is used to update the parameters of the gradient boosting regression model—preserving the previous parameters of the model. The decision trees remain the same; only the last one is retrained on the new data. Decision trees, among which The value ranges from 10 to 30 trees. The technical advantage of the incremental learning strategy is that it avoids the high computational cost of daily full retraining while enabling the model to promptly absorb the distribution trends in the latest data, achieving a balance between computational efficiency and prediction accuracy. Preferably, this invention performs a full model retraining every 30 days to completely eliminate prediction biases that may accumulate during incremental learning.

[0059] Furthermore, step S5 also feeds back the deviation between the actual load utilization rate of the day and the predicted load utilization rate in step S3 to the path planning algorithm in step S3. Specifically, if the actual load utilization rate of the day is consistently lower than the predicted load utilization rate by more than 5 percentage points (for 3 consecutive days), the system automatically adjusts the load utilization rate threshold. The rate will be lowered by 0.05 to relax the reward trigger conditions and encourage more route merging; conversely, if the actual load utilization rate on a given day is consistently higher than the predicted load utilization rate by more than 5 percentage points (for 3 consecutive days), then... The reward was increased by 0.02 to tighten the incentive conditions and avoid the risk of vehicle overloading. This closed-loop feedback mechanism enables the path planning parameters to adaptively adjust with changes in the actual operating environment, ensuring that the system maintains optimal scheduling performance in long-term operation.

[0060] like Figure 2 As shown, the dynamic scheduling system for urban waste sorting and collection vehicles provided by the present invention includes 5 functional modules, each of which corresponds to one of the 5 steps in the method embodiment, forming a complete hardware and software collaborative architecture.

[0061] The multi-source data acquisition and feature construction module 1 corresponds to step S1. This module includes four sub-modules: a sensor data access sub-module, a calendar data access sub-module, a meteorological data access sub-module, and a feature matrix construction sub-module. The sensor data access sub-module establishes a data channel with the smart trash cans at each collection point via an NB-IoT gateway, receiving weighing data at 15-minute intervals and storing it in a time-series database. In a preferred embodiment of this invention, InfluxDB is used as the time-series database engine, capable of supporting a write throughput of over 10,000 records per second. The calendar data access sub-module obtains calendar features by calling a public calendar API and writes them to a relational database at a daily granularity. The meteorological data access sub-module obtains historical observations and 7-day forecast meteorological data by calling the meteorological bureau's open data interface. The forecast data is updated twice daily (at 08:00 and 20:00 respectively) to ensure that the meteorological input used for the next day's forecast is the latest available version. The feature matrix construction submodule automatically performs data alignment, outlier detection and repair, z-score normalization, and feature matrix assembly tasks at 01:00 every day. The specific methods for outlier detection and normalization are consistent with those described in step S1. The assembled feature matrix is ​​stored in JSON format and pushed to the downstream prediction module via a message queue.

[0062] The waste sorting output prediction module 2 corresponds to step S2. This module internally maintains four independent gradient boosting regression prediction sub-models, each corresponding to one of the four waste categories. After receiving the latest feature matrix output from the feature matrix construction sub-module at midnight each day, this module inputs the calendar features and weather forecast data for the next day into each prediction sub-model, outputs the predicted output of the four waste categories for each collection point for the next day, and pushes it to the downstream path planning module. The specific implementation of model training, hyperparameter configuration, and regularization strategies is consistent with the description in step S2. In a preferred embodiment of the invention, this module also includes a built-in model performance monitoring sub-module, which continuously tracks the MAPE change trend of each prediction sub-model within a sliding window. When the MAPE exceeds a preset warning threshold, the model retraining process is automatically triggered. Preferably, this module employs a model version management mechanism, automatically saving a snapshot of the current model (including all decision tree parameters and feature importance ranking) after each model update, and retaining the most recent 30 versions in the version repository. When the model's performance degrades on the validation set after an incremental update (MAPE increases by more than 2 percentage points compared to the previous version), the system automatically rolls back to the previous stable version and records the degradation log for operations and maintenance personnel to investigate the cause. This version management mechanism ensures the stability and traceability of the prediction module in long-term operation.

[0063] The enhanced load utilization route planning module 3 corresponds to step S3. After receiving the predicted output data, this module calls the improved saving algorithm engine to execute route planning. The algorithm engine internally implements all sub-steps, including enhanced saving value calculation, greedy merging, MTZ constraint verification, and feasibility verification of dual constraints (load limit and volume limit), with technical details consistent with those described in step S3. The output of the route planning module is the initial scheduling route plan for each type of vehicle, including the collection point access sequence, estimated travel distance, and estimated load capacity for each vehicle. This module also includes a route visualization sub-module, which can intuitively display the planned routes of each vehicle on a GIS map, facilitating manual review and fine-tuning by dispatching managers. In a preferred embodiment of the present invention, the route planning module also has a parameter configuration interface, allowing dispatching managers to dynamically adjust the load utilization threshold according to actual operational needs. and reward coefficient The value of can be adjusted. For example, during peak waste production periods (such as long holidays like Spring Festival and National Day), managers can appropriately reduce . This encourages more proactive route merging to improve capacity utilization efficiency; default parameter settings are restored on regular workdays. Furthermore, the module includes a built-in scheme quality assessment sub-function. After generating a scheduling route scheme, it automatically calculates indicators such as expected load utilization rate, expected total travel distance, and expected total number of trips, and quickly compares them with historical data from the same period. If any indicator deviates significantly from the historical average by more than two standard deviations, an early warning is issued to dispatch management personnel to prevent a decline in route scheme quality due to abnormal predicted data.

[0064] The real-time monitoring and dynamic route re-optimization module 4 corresponds to step S4, and this module operates continuously throughout the day. The GPS data receiving submodule receives GPS positioning data from each vehicle every 30 seconds and updates the vehicle status table. After receiving the overflow alarm signal from the smart trash can, the alarm processing submodule calls the dynamic insertion engine to execute all dynamic re-optimization sub-steps, including candidate vehicle screening, distance calculation and sorting, and minimum cost insertion. The technical details are consistent with those described in step S4. The re-optimization results are sent to the on-board terminal of the target vehicle in real time to update the navigation route. In a preferred embodiment of the present invention, this module also includes a backup vehicle scheduling submodule—when all on-the-road vehicles cannot meet the alarm requirements, vehicles are automatically dispatched from the backup vehicle pool.

[0065] The scheduling effect evaluation and parameter feedback module 5 corresponds to step S5, and this module is automatically executed after the daily scheduling ends. The indicator statistics submodule extracts the daily operation data from the vehicle status table and GPS trajectory database, and calculates the actual load utilization rate, total travel distance, and total transportation cost. The comparative analysis submodule compares the dynamic scheme indicators with the static benchmark scheme and generates a daily optimization effect report. The parameter feedback submodule appends the daily actual output data to the training set and triggers the incremental update of the prediction model, while feeding back the load utilization rate deviation to the path planning module for adaptive adjustment of the threshold parameters. Its technical details are consistent with those described in step S5. The daily report is automatically sent to the work email of the scheduling management personnel in PDF format. In a preferred embodiment of the present invention, this module also has a trend analysis subfunction, which can perform sliding window statistical analysis on the scheduling indicators of the most recent 7 days, 30 days, and 90 days, automatically identify trend anomalies such as continuous decline in load utilization rate, abnormal increase in travel distance, or sudden increase in overflow frequency, and present them with a red warning mark in the daily report. The technical benefits of trend analysis are: it helps scheduling managers discover the slow degradation trend of system performance from a global perspective and take timely intervention measures, such as recalibrating sensor accuracy, updating meteorological data sources, or performing full model retraining.

[0066] The data flow relationship between the above 5 functional modules is as follows: the multi-source data acquisition and feature construction module 1 outputs the feature matrix to the waste sorting production prediction module 2; the waste sorting production prediction module 2 outputs the predicted production to the route planning module 3 for enhanced load utilization; the route planning module 3 outputs the initial scheduling route plan to the real-time monitoring and dynamic route re-optimization module 4; the dynamic route re-optimization module 4 outputs the daily operation data to the scheduling effect evaluation and parameter feedback module 5; the scheduling effect evaluation and parameter feedback module 5 feeds back the model update data to the waste sorting production prediction module 2, and at the same time feeds back the threshold adjustment parameters to the route planning module 3, thus forming a closed-loop system architecture that combines forward transmission and backward feedback.

[0067] To verify the effectiveness of the technical solution of this invention, a test environment was used for the actual waste sorting and collection scenario in District A of a certain city. The test area included 120 collection points, distributed across approximately 50... Within the urban area, there is one waste transfer center. Available vehicles include: 12 dedicated food waste trucks (maximum load capacity...). kg, maximum volume 8 special vehicles for recyclable materials (with maximum load capacity) kg, maximum volume 3 special vehicles for hazardous waste (with maximum load capacity) kg, maximum volume ) and 10 other garbage trucks (load limit) kg, maximum volume The testing period spans 90 consecutive days (including the Spring Festival holiday, the summer heat period, and regular working days).

[0068] Regarding prediction accuracy, the gradient boosting regression model used in this invention is compared with the traditional arithmetic mean method. Using the last 30 days as the test set and the first 60 days as the training set, the prediction accuracy comparison results for the four types of waste are as follows: the MAPE of kitchen waste decreased from 22.4% of the arithmetic mean method to 8.7% of the gradient boosting regression model, an improvement rate of 61.2%; the MAPE of recyclables decreased from 28.6% to 11.2%, an improvement rate of 60.8%; the MAPE of hazardous waste decreased from 35.1% to 15.6%, an improvement rate of 55.6%; and the MAPE of other waste decreased from 21.8% to 9.3%, an improvement rate of 57.3%. These results indicate that the gradient boosting regression model can effectively capture the nonlinear influence of calendar features and meteorological factors on the production of various types of waste, and significantly outperforms traditional methods across all waste categories.

[0069] In terms of route planning, this invention compares the enhanced load utilization efficiency saving algorithm with the classic saving algorithm without load utilization efficiency rewards. Average data over 90 days shows that the average number of trips for kitchen waste disposal decreased from 16.8 trips in the classic saving algorithm to 13.5 trips in the enhanced saving algorithm, a reduction of 19.6%; the average load utilization efficiency increased from 64.1% to 82.3%, an increase of 18.2 percentage points; and the average total travel distance decreased from 476km to 398km, a saving of 16.4%. Recyclables, hazardous waste, and other waste categories showed similar improvement trends. The total daily transportation cost for the four categories decreased from 28,500 yuan in the static plan to 22,200 yuan in the dynamic plan, a saving rate of 22.1%.

[0070] Regarding dynamic response, a total of 87 overflow alarm events occurred during the 90-day test period (approximately one per day). Of these, 83 (95.4%) were successfully handled by dynamically inserting on-the-go vehicles, with an average response time of 38 minutes. The remaining 4 events required the dispatch of backup vehicles due to insufficient remaining load capacity of on-the-go vehicles, with an average response time of 67 minutes. In contrast, under the static solution, overflow events could only wait until the next day's collection, with an average dwell time exceeding 12 hours. The dynamic dispatch solution reduced the average response time for overflow events from over 12 hours to less than 40 minutes, effectively preventing environmental pollution and resident complaints caused by garbage overflow.

[0071] Regarding closed-loop feedback, as the system runs for more days, the predictive model continuously absorbs the latest data through daily incremental learning and monthly full retraining. The average MAPE (Marginal Performance Efficiency) of the four types of waste shows a stable downward trend over 90 days: the average MAPE is 12.4% from day 1 to 30, decreasing to 10.1% from day 31 to 60, and further decreasing to 8.9% from day 61 to 90. Meanwhile, the load utilization threshold... Under the closed-loop feedback mechanism, the value stabilized between 0.78 and 0.82 after multiple adaptive adjustments from the initial value of 0.80, indicating that the system can dynamically optimize according to the actual operating environment.

[0072] It is particularly noteworthy that the gradient boosting regression model demonstrates a significant predictive advantage during periods of dramatic fluctuations in waste production before and after the Spring Festival holiday. Taking the Spring Festival holiday as an example, the production of kitchen waste from the first to the third day of the Lunar New Year surged by approximately 180% compared to normal levels. While the traditional arithmetic mean method yielded a MAPE (Mean Average Error Per Second) of 45.2% during this period, the gradient boosting regression model of this invention, by utilizing holiday characteristics to capture such regular fluctuations, achieved a MAPE of only 13.8%, improving prediction accuracy by nearly 70%. During periods of high summer temperatures (consecutive days with average daily temperatures exceeding 35°C), kitchen waste production increases by approximately 30% to 40% compared to days with normal temperatures. The gradient boosting regression model effectively tracks this increase through nonlinear mapping of temperature characteristics, while the arithmetic mean method fails to respond at all. These test results fully validate the technical rationality and effectiveness of this invention in incorporating calendar features and meteorological data into waste production prediction.

[0073] Furthermore, from a computational efficiency perspective, the improved resource-saving algorithm of this invention, with a scale of 120 collection points, takes approximately 2.1 seconds for path planning of a single type of waste (running on a server environment with an Intel Core i7-12700 processor and 32GB of memory), and the path planning of the four types of waste can be executed in parallel, with a total time not exceeding 3 seconds. In contrast, the genetic algorithm requires 500 to 1000 iterations to converge at the same scale, and the path planning of a single type of waste takes approximately 45 to 90 seconds. Therefore, the improved resource-saving algorithm adopted in this invention has an order-of-magnitude advantage in computational speed, fully meeting the actual need for rapid generation of scheduling schemes every morning. During the dynamic path re-optimization process, the dynamic insertion calculation for a single overflow alarm takes no more than 0.5 seconds, enabling near real-time response to sudden events.

[0074] In summary, this invention achieves synergistic effects of accurate waste production prediction, full utilization of vehicle load capacity, timely response to sudden demands, and continuous improvement of system performance through deep coupling of gradient boosting regression prediction and improved path planning of the waste-saving algorithm, rapid response of real-time monitoring and dynamic re-optimization, and closed-loop adaptive effect of effect evaluation and model parameter feedback. All four core indicators (prediction accuracy, load capacity utilization, total transportation cost, and overflow response time) have been significantly improved compared with the existing static scheduling scheme.

[0075] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A method for dynamic scheduling of urban waste sorting and collection vehicles, characterized in that, Includes the following steps: Step S1, Multi-source data acquisition and feature construction steps: Obtain historical daily output data from the weighing sensors of smart trash cans at each collection point, and simultaneously collect calendar feature data and meteorological data, and construct a multi-dimensional feature matrix after aligning by timestamp; The calendar features include the week number, whether it is a statutory holiday, and the month; the meteorological data includes the average daily temperature, daily precipitation, and average daily wind speed. Step S2, waste production prediction step: Using the multidimensional feature matrix as input, establish gradient boosting regression models for the four types of waste (kitchen waste, recyclables, hazardous waste, and other waste) and train them to obtain four prediction sub-models. After inputting the calendar features and weather forecast data for the next day, output the predicted production of the four types of waste at each collection point for the next day. Step S3, route planning step with enhanced load utilization: Input the predicted output into the improved saving algorithm for heterogeneous fleet route planning. This algorithm introduces a load utilization reward item on the basis of the classic saving value. When the load utilization of the candidate merging path exceeds the preset threshold, the reward value is increased to improve the merging priority. At the same time, MTZ constraints are applied to eliminate sub-loops and apply double constraints of load limit and volume limit to output the initial scheduling route scheme. Step S4, Real-time monitoring and dynamic route re-optimization: Each vehicle performs the collection task according to the initial scheduling route scheme output in step S3. The real-time location and completion progress of each vehicle are tracked through the vehicle GPS. When an overflow alarm is triggered at a collection point, the alarm collection point is inserted into the route of the nearest vehicle with sufficient remaining load. Only the order of subsequent unvisited nodes is adjusted without affecting the completed route segment. Step S5, scheduling effect evaluation and parameter feedback: After the daily scheduling is completed, the actual load utilization rate, total travel distance and total transportation cost are calculated and compared with the static fixed route plan to generate a daily report on the optimization effect. The actual output data of the day is fed back to the gradient boosting regression model in step S2 to update the model parameters.

2. The method for dynamic scheduling of urban waste sorting and collection vehicles according to claim 1, characterized in that, In step S1, the historical daily output data collection period is no less than 90 days, the weighing sensor accuracy at each collection point is no less than 0.5 kg, the weekday number in the calendar feature data uses integer encoding from 1 to 7, whether it is a statutory holiday uses binary encoding from 0 and 1, and the month uses integer encoding from 1 to 12.

3. The method for dynamic scheduling of urban waste sorting and collection vehicles according to claim 1, characterized in that, In step S2, the maximum depth of the decision tree of the gradient boosting regression model is set to an integer value between 4 and 8, the learning rate is set to a range of 0.01 to 0.1, the number of iteration rounds is set to 100 to 500 rounds, and the mean squared error is used as the loss function during training.

4. The method for dynamic scheduling of urban waste sorting and collection vehicles according to claim 1, characterized in that, In step S3, the load utilization threshold is set to a value between 0.75 and 0.85, and the bonus value is set to 0.1 to 0.3 times the original savings value before merging.

5. The method for dynamic scheduling of urban waste sorting and collection vehicles according to claim 1, characterized in that, In step S2, before constructing the multidimensional feature matrix, an outlier detection step is also included for historical daily production data. The production data that exceeds the upper quartile plus 1.5 times the quartile or is lower than the lower quartile minus 1.5 times the quartile is marked as outliers using a method based on the interquartile range, and the average of adjacent normal days is used to replace them.

6. The method for dynamic scheduling of urban waste sorting and collection vehicles according to claim 1, characterized in that, In step S3, the improved saving algorithm includes the following sub-steps: First, construct the initial solution for a single-point round trip of all collection points starting from the transfer center, and then calculate the classic saving value between any two collection points; for candidate merging paths where the load utilization rate exceeds the load utilization rate threshold, add a reward value to the classic saving value to obtain the enhanced saving value; sort all candidate merging paths according to the enhanced saving value from largest to smallest and attempt to merge them in sequence, checking whether the load limit constraint and volume limit constraint are satisfied during merging, executing the merging that satisfies the constraints, and skipping the merging that does not satisfy the constraints, until all candidate merging checks are completed.

7. The method for dynamic scheduling of urban waste sorting and collection vehicles according to claim 1, characterized in that, In step S4, the method for inserting the alarm collection point into the path of on-the-road vehicles is as follows: calculate the straight-line distance between the alarm collection point and the current position of all on-the-road vehicles, sort them in ascending order of distance, and then check whether the remaining load capacity of each on-the-road vehicle meets the garbage volume requirement of the alarm collection point. Select the nearest on-the-road vehicle that meets the requirement as the target vehicle, and use the minimum cost insertion method to insert the alarm collection point into the position with the minimum cost in the subsequent unvisited node sequence of the target vehicle.

8. The method for dynamic scheduling of urban waste sorting and collection vehicles according to claim 1, characterized in that, In step S5, the actual load utilization rate is calculated as the ratio of the total actual transported mass of all trips on that day to the sum of the rated load capacity of all trips. The total transportation cost is calculated as two parts: fixed vehicle scheduling cost and fuel cost proportional to the travel distance.

9. The method for dynamic scheduling of urban waste sorting and collection vehicles according to claim 1, characterized in that, In step S3, the MTZ constraint takes the form of: for any two non-transfer center collection points in the path, an auxiliary variable representing the access order is introduced and an increasing order constraint is established to eliminate sub-loops. At the same time, the value range of the auxiliary variable is from 1 to the total number of collection points.

10. A dynamic scheduling system for urban waste sorting and collection vehicles, used to implement the dynamic scheduling method for urban waste sorting and collection vehicles as described in any one of claims 1-9, characterized in that, include: The multi-source data acquisition and feature construction module is used to acquire historical daily output data from the weighing sensors of smart trash cans deployed at various collection points, and at the same time collect calendar feature data and meteorological data for the corresponding time period. After aligning the historical daily output data, calendar feature data and meteorological data according to the timestamp, a multi-dimensional feature matrix is ​​constructed. The waste sorting and production prediction module receives the multi-dimensional feature matrix output by the multi-source data acquisition and feature construction module, and establishes gradient boosting regression models for the four types of waste: kitchen waste, recyclables, hazardous waste, and other waste, and outputs the predicted production of the four types of waste at each collection point on the next day. The route planning module with enhanced load utilization is used to receive the predicted output from the waste sorting output prediction module. It introduces a load utilization reward term on the basis of the classic saving value calculation through an improved saving algorithm to plan vehicle routes, and applies MTZ constraints and load and volume dual constraints to output the initial scheduling route scheme for each type of vehicle. The real-time monitoring and dynamic route re-optimization module is used to track the real-time location and task completion progress of each vehicle through the vehicle GPS. When an overflow alarm signal is received, the alarm collection point is automatically inserted into the route of the nearest vehicle with sufficient remaining load capacity. The scheduling effect evaluation and parameter feedback module is used to calculate the actual load utilization rate, total travel distance and total transportation cost after the daily scheduling is completed, and compare them with the static fixed route plan to generate a daily optimization effect report. It also feeds back the actual output data of each collection point on the same day to the classified waste output prediction module to update the gradient boosting regression model parameters.

Images