Big data-based intelligent site selection method and device for household waste classification and disposal point

Abstract

The application discloses a kind of based on big data's intelligent site selection method and device of domestic waste classification and putting point, belong to wisdom city environmental sanitation facility planning field, method includes: fusion population statistics, mobile phone signaling heat, interest point, building contour, road network, put metering and track of collection and transport multi-source data and mapping to unified space-time grid;Based on graph convolution and time series attention network, the space-time distribution and prediction interval of each grid garbage production are classified and predicted;With weighted putting distance, construction operation cost, neighbor-avoiding negative effect as target, with coverage and capacity as constraint, the initial layout is obtained by solving multi-objective site selection model;Layout adjustment is modeled as Markov decision process, and the layout is dynamically optimized by rolling according to the feedback of putting using reinforcement learning.The application realizes the change of putting point site selection from static experience planning to dynamic closed-loop optimization, and in the embodiment, the average putting distance of residents is shortened, and the peak overflow event rate is reduced.

Description

Technical Field

[0001] This invention relates to the field of smart city environmental sanitation facility planning and optimization technology, specifically to a method and device for intelligent site selection of household waste sorting and disposal points based on big data. Background Technology

[0002] With the nationwide implementation of the household waste sorting system in prefecture-level cities and above, the rationality of the spatial layout of waste sorting collection points, as the "first point of contact" for residents' waste sorting behavior, directly determines the convenience of residents' waste disposal, the participation rate, and the accuracy of waste disposal, thus affecting the operational efficiency of the entire chain of subsequent collection, transportation, and treatment. Overly dense placement of collection points leads to idle facilities and increased collection and transportation costs, while overly sparse placement results in excessively long distances for residents to dispose of waste, frequent overflows at collection points, and induces mixed and indiscriminate waste disposal. Therefore, how to scientifically and efficiently determine the spatial layout of household waste sorting collection points has become an urgent technical problem to be solved in the refined governance of cities.

[0003] Existing technologies mainly develop along three technical routes, but all have significant drawbacks: The first technical route is static site selection based on Geographic Information Systems (GIS). Existing technologies acquire geographic data to establish a spatial database, create thematic maps, and conduct spatial analysis. Then, a hierarchical comprehensive evaluation system is constructed to analyze the suitability of candidate areas, thereby determining the site selection result. This type of scheme relies on manually set hierarchical analysis weights to overlay and score static geographic elements. Its shortcomings are: the site selection basis only includes static geographic elements such as land use and hydrogeology, without integrating dynamic big data such as population spatiotemporal distribution and waste disposal behavior, failing to characterize the spatiotemporal heterogeneity of residents' waste disposal needs; it lacks a prediction step for future waste generation, and the site selection result only reflects the static situation at the evaluation point in time; once the site selection scheme is determined, it becomes fixed and cannot be adjusted based on actual operational feedback.

[0004] The second technical approach is a one-time solution based on intelligent optimization algorithms. Existing technologies divide villages into road grid zones and determine a set of optional site locations. Using the dual objective function of garbage disposal points being close to buildings and buildings being far from garbage disposal points, simulated annealing algorithms are used to iteratively solve for the optimal solution of the optional site location distribution. Similarly, the paper "Optimization of Rural Garbage Disposal Site Selection Based on Improved Simulated Annealing" published by Xu Jinglin et al. in *Information Technology for Civil Engineering*, Vol. 15, No. 2, 2023, and the hierarchical clustering garbage collection point layout method published by Viktorin et al. in *Computers & Industrial Engineering* (2023), also employ a similar "one-time static optimization" paradigm. The drawback of this type of scheme is that its optimization model takes static values ​​estimated by a fixed coefficient based on per capita output as input, and does not distinguish the significant differences in the generation and spatiotemporal distribution of different types of waste such as recyclables, kitchen waste, hazardous waste and other waste. The optimal solution obtained corresponds to a fixed demand scenario. When population influx, seasonal fluctuations, and changes in population flow during holidays cause demand distribution to drift, the original site selection scheme quickly becomes inefficient. This type of scheme does not have a closed-loop mechanism for re-optimization based on operational feedback and can only rely on manual restarting of the site selection calculation.

[0005] The third technical approach involves the single-point application of machine learning in waste production prediction or general facility site selection. For example, existing technologies predict monthly sorted waste production at the administrative region scale based on macro-statistical data, but the prediction results remain at the macro-level of administrative regions and cannot be scaled down to the spatial grid scale of hundreds of meters required for site selection. Furthermore, this approach does not involve the site selection optimization process.

[0006] In summary, no existing technology has yet found a comprehensive site selection solution for household waste sorting collection points that integrates multi-source big data fusion, demand prediction for household waste sorting, multi-objective intelligent site selection, and reinforcement learning dynamic optimization. Existing technologies suffer from at least the following technical deficiencies: a single data foundation, lacking unified spatiotemporal fusion of heterogeneous multi-source data such as population heat maps, POIs, building outlines, disposal metering, and collection trajectories, resulting in a coarse demand characterization; static demand input, lacking spatiotemporal prediction of generation volume based on a 100-meter grid and differentiating waste categories, leading to deviations between the site selection solution and actual demand from the outset; static site selection solutions, outputting a fixed layout in a single step, unable to respond to dynamic shifts in demand distribution; and a lack of operational feedback, with actual operational data such as overflow, complaints, and utilization rates not fed back into the site selection decision-making process, failing to form a closed loop for continuous optimization. These deficiencies result in existing site selection solutions generally exhibiting problems such as excessively long distances for residents to dispose of waste, high overflow rates during peak hours, and polarized facility utilization rates in actual operation. Summary of the Invention

[0007] (a) Technical problems to be solved To address the aforementioned shortcomings of existing technologies, the present invention aims to provide a method and apparatus for intelligent site selection of household waste sorting and disposal points based on big data. This method solves the technical problems of existing waste sorting and disposal point site selection technologies, such as a single data foundation, static demand input, fixed layout schemes, and lack of operational feedback loops. It aims to transform the site selection of disposal points from static experience-based planning to data-driven dynamic closed-loop optimization.

[0008] (II) Technical Solution To address the aforementioned technical problems, this invention proposes a method for intelligent site selection of household waste sorting and disposal points based on big data, comprising the following steps: Step S1, Multi-source Big Data Fusion: Acquire data from multiple heterogeneous data sources within the target area. The heterogeneous data sources include at least historical household waste sorting and disposal measurement data and road network data, and include at least one of the following: population statistics, mobile phone signal population heat map data, point of interest (POI) data, building outline data, and waste collection vehicle trajectory data. Perform preprocessing on the data from each data source, including coordinate unification, time alignment, missing value completion, and outlier removal. Map the preprocessed data to spatial grid cells of a preset scale to construct a multi-source fused spatiotemporal dataset with spatial grid cells as the basic unit and time slices as the sequence. Step S2, Prediction of Household Waste Sorting Demand: Construct a spatiotemporal graph structure using the spatial grid units as nodes and the road network reachability relationships between the spatial grid units as edges. Input the historical waste generation sequence and corresponding covariates of each spatial grid unit in the multi-source fusion spatiotemporal dataset into the trained spatiotemporal prediction model, and output the predicted generation values ​​and prediction intervals of each waste category for each spatial grid unit in the future preset time period, thereby obtaining the spatiotemporal distribution data of waste sorting demand; wherein, the spatiotemporal prediction model includes a graph convolutional layer for extracting spatial dependencies and a temporal attention layer for extracting temporal dependencies; Step S3, Multi-objective Intelligent Site Selection: Determine the demand weight of each spatial grid unit based on the spatiotemporal distribution data of waste sorting demand. Within the target area, candidate disposal points are screened according to land suitability rules. The objective functions are minimizing the weighted disposal distance of residents, minimizing the construction and operation and maintenance costs of facilities, and minimizing the NIMBY (Not In My Backyard) negative effect. The constraints are service coverage, disposal point capacity, and number of disposal points. The predicted value of the amount of various types of waste generated at any disposal point is determined according to the upper quantile value of the prediction interval. The upper quantile value is used to verify the upper limit of the capacity of the disposal point. A multi-objective site selection optimization model is constructed. The Pareto solution set is obtained by solving the multi-objective site selection optimization model. A compromise site selection scheme is determined from the Pareto solution set as the initial layout scheme of the household waste sorting disposal points. Step S4, reinforcement learning dynamic optimization: The layout adjustment process of household waste sorting and disposal points is modeled as a Markov decision process. The state space includes at least the predicted generation value of each spatial grid cell, the utilization rate of each disposal point, the number of overflow events, and the number of resident complaints. The action space includes at least the activation, deactivation, relocation, and capacity adjustment of disposal points. The reward function includes at least the weighted disposal distance term, the overflow penalty term, the layout adjustment cost term, and the complaint penalty term. A reinforcement learning agent is trained based on the Markov decision process. After the initial layout scheme is put into operation, the actual disposal feedback data is collected and the state is updated according to the preset optimization cycle. The trained reinforcement learning agent outputs layout adjustment actions to perform rolling dynamic optimization of the layout of household waste sorting and disposal points.

[0009] Preferably, in step S1, mapping the preprocessed data to spatial grid cells of a preset scale specifically includes: dividing the target area into square grids of preset side lengths or irregular grids divided according to community boundaries and road-enclosed areas; aggregating the mobile phone signal population heat map data into each spatial grid cell by time slice to obtain the heat map values ​​of the resident population and the floating population of each spatial grid cell; calculating the number and mixing index of each type of POI data in each spatial grid cell according to residential, commercial, school, and office categories; using the building outline data to calculate the total building area, plot ratio, and residential building area ratio of each spatial grid cell; mapping the historical household waste classification and disposal measurement data to the corresponding spatial grid cell according to the disposal point affiliation, and aggregating them according to four categories: recyclables, kitchen waste, hazardous waste, and other waste to form a time series of classified waste generation; and using Kriging interpolation based on spatially adjacent grids or collaborative filtering based on matrix decomposition to complete missing values.

[0010] Preferably, in step S2, the spatiotemporal prediction model includes a graph convolutional layer, a temporal attention layer, and a multi-task output layer connected in sequence; the graph convolutional layer performs spatial aggregation of node features using a normalized road network reachable adjacency matrix, and the edge weights of the adjacency matrix are determined based on the reciprocal of the shortest path distance between spatial grid cells; the temporal attention layer applies multi-head self-attention computation to the time series features of each node output by the graph convolutional layer to capture the daily, weekly, and holiday fluctuation patterns of waste generation; the multi-task output layer includes multiple output branches corresponding one-to-one with waste categories, each output branch outputting at least two quantile values ​​of the waste generation corresponding to the category, with the quantile values ​​representing the prediction interval of the generation prediction value; the covariates include at least population heat value, POI mixing index, holiday indicator variable, and meteorological data.

[0011] Preferably, in step S3, the demand weight is determined based on the weighted sum of the predicted values ​​of various types of waste generation in each spatial grid unit within a future preset time period, and the weight coefficient of each type of waste is set according to the collection frequency and unit collection cost of the corresponding category; the weighted disposal distance for residents is the sum of the products of the demand weight of each spatial grid unit and the shortest walking distance along the road network from the spatial grid unit to its assigned disposal point; the NIMBY (Not In My Backyard) negative effect is determined based on the proximity of the candidate disposal point to residential buildings, schools, and hospitals, as well as the service scale of the candidate disposal point; the constraints include: the walking distance from any spatial grid unit to its assigned disposal point does not exceed the preset service radius, the predicted value of waste generation assigned to any disposal point does not exceed the capacity limit of the disposal point, and the total number of disposal points does not exceed the maximum number corresponding to the budget constraint.

[0012] Preferably, in step S3, the multi-objective location optimization model is solved using a non-dominated sorting genetic algorithm with an elite retention strategy. The start / stop status of each candidate placement point is encoded as a chromosome, and the objective function is used as the fitness evaluation criterion. Iterative evolution is achieved through non-dominated sorting and crowding calculation, and the mutation probability is adaptively adjusted based on the demand density of the candidate placement points during crossover and mutation operations to obtain a Pareto solution set. A multi-attribute decision-making method combining entropy weighting and approximation of ideal solution sorting is used to determine the compromise location scheme from the Pareto solution set. This involves normalizing the objective function values ​​of each solution in the Pareto solution set, determining the objective weights of each objective using entropy weighting, calculating the relative proximity of each solution to the positive and negative ideal solutions, and selecting the solution with the highest relative proximity as the compromise location scheme.

[0013] Preferably, in step S4, the reinforcement learning agent is trained using a deep Q-network or a proximal policy optimization algorithm. The training process includes an offline training phase and an online optimization phase. The offline training phase is conducted in a simulation environment constructed based on the multi-source fusion spatiotemporal dataset. The simulation environment simulates demand fluctuations, overflow events, and disposal behaviors based on historical waste generation sequences. The weight coefficients of the weighted disposal distance term, overflow penalty term, layout adjustment cost term, and complaint penalty term in the reward function are set according to the management preferences of the target area. The preset optimization period is from one week to one quarter. At the end of each optimization period, the status is updated based on the collection of disposal point measurement data, overflow sensor data, and complaint records within that period. When the reinforcement learning agent uses a deep Q-network, it calculates the action value of each candidate action in the current state and executes the layout adjustment action with the highest action value. When the reinforcement learning agent uses a proximal policy optimization algorithm, the policy network outputs the probability distribution of each candidate action in the current state and executes the layout adjustment action with the highest probability. The actual feedback of the execution result is stored in the experience replay pool to incrementally update the network parameters of the reinforcement learning agent.

[0014] Preferably, when the target area is a newly established area without historical household waste sorting and disposal measurement data, step S2 further includes: selecting an existing area that meets a preset threshold in terms of similarity to the target area in terms of population structure, land use composition, and climate conditions as the source area; transferring the graph convolutional layer parameters of the spatiotemporal prediction model trained in the source area to the spatiotemporal prediction model of the target area; and calibrating the transferred model using the planned population import data, planned land use data, and prior demand estimation of the analog grid of the target area; as actual disposal measurement data accumulates in the target area, gradually increasing the proportion of actual data in the model retraining samples according to a preset ratio to achieve rolling calibration of demand prediction.

[0015] To address the aforementioned technical problems, this invention also proposes an intelligent site selection device for household waste sorting and disposal points based on big data, comprising: The data fusion module is used to acquire data from multiple heterogeneous data sources within the target area. These heterogeneous data sources include at least historical household waste sorting and disposal measurement data and road network data, and at least one of the following: population statistics, mobile phone signal population heat map data, point of interest (POI) data, building outline data, and waste collection vehicle trajectory data. The module performs preprocessing on the data from each data source, including coordinate unification, time alignment, missing value completion, and outlier removal. The preprocessed data is then mapped to spatial grid cells of a preset scale to construct a multi-source fused spatiotemporal dataset with spatial grid cells as the basic unit and time slices as the sequence. The demand prediction module is used to construct a spatiotemporal graph structure with the spatial grid units as nodes and the road network reachability relationships between the spatial grid units as edges. It inputs the historical waste generation sequence of each spatial grid unit in the multi-source fusion spatiotemporal dataset and the corresponding covariates into a trained spatiotemporal prediction model, and outputs the predicted generation values ​​and prediction intervals for each waste category in each spatial grid unit within a preset future time period, thus obtaining the spatiotemporal distribution data of waste classification demand. The spatiotemporal prediction model includes a graph convolutional layer for extracting spatial dependencies and a temporal attention layer for extracting temporal dependencies. The site selection calculation module is used to determine the demand weight of each spatial grid unit based on the spatiotemporal distribution data of waste sorting demand, screen candidate disposal points within the target area according to land suitability rules, and construct a multi-objective site selection optimization model with the objectives of minimizing the weighted disposal distance of residents, minimizing the construction and operation and maintenance costs of facilities, and minimizing the NIMBY (Not In My Backyard) negative effect, and with service coverage, disposal point capacity, and number of disposal points as constraints. The multi-objective site selection optimization model is solved to obtain the Pareto solution set, and a compromise site selection scheme is determined from the Pareto solution set as the initial layout scheme of household waste sorting disposal points. The dynamic optimization module is used to model the layout adjustment process of household waste sorting and disposal points as a Markov decision process. The state space includes at least the predicted generation value of each spatial grid cell, the utilization rate of each disposal point, the number of overflow events, and the number of resident complaints. The action space includes at least the activation, deactivation, relocation, and capacity adjustment of disposal points. The reward function includes at least a weighted disposal distance term, an overflow penalty term, a layout adjustment cost term, and a complaint penalty term. A reinforcement learning agent is trained based on the Markov decision process. After the initial layout scheme is put into operation, the module collects actual disposal feedback data to update the state according to a preset optimization cycle. The trained reinforcement learning agent outputs layout adjustment actions to perform rolling dynamic optimization of the layout of household waste sorting and disposal points. The output module is used to output the initial layout scheme and the location scheme corresponding to the layout adjustment action.

[0016] Preferably, the spatiotemporal prediction model in the demand prediction module includes a graph convolutional layer, a temporal attention layer, and a multi-task output layer connected in sequence. The multi-task output layer includes multiple output branches corresponding one-to-one with waste categories, and each output branch outputs at least two quantile values ​​of the waste generation amount for the corresponding category. The dynamic optimization module also includes an experience replay pool and a parameter update unit. The experience replay pool is used to store layout adjustment actions and their actual feedback. The parameter update unit is used to incrementally update the network parameters of the reinforcement learning agent based on the data in the experience replay pool at the end of each optimization cycle. The output module is also used to output the site selection scheme in at least one of the following forms: electronic map layer overlay, layout comparison chart, and adjustment suggestion list.

[0017] The present invention also proposes an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the computer program, when executed by the processor, implements the steps of the intelligent site selection method for household waste sorting and disposal points based on big data as described in any one of claims 1 to 7.

[0018] (iii) Beneficial technical effects Compared with existing technologies, the beneficial effects of this invention are: 1. Demand forecasting is scaled down to a grid of 100 meters and distinguishes four waste categories. Verified by examples, the mean absolute percentage error (MAPE) of the predicted waste generation in the next week is no higher than 12%, which is significantly better than existing macro-scale forecasting methods. This improves the matching degree between site selection schemes and actual demand from the source. 2. The weighted delivery distance is optimized by using predicted demand as the weight. In the example, the average walking distance for residents is reduced by more than 20% compared with the manual experience-based delivery method, and more than 95% of residents are within the coverage area of ​​the preset service radius. 3. Enhanced learning and dynamic optimization enable the layout of delivery points to be continuously adjusted according to demand. In the example, the overflow rate of delivery points during peak hours is reduced by more than 35% compared with a fixed layout, and the regional difference coefficient of facility utilization is reduced by more than 40%. 4. The fully automated calculation process replaces manual site surveys and multiple rounds of review. In the example, the time for generating site selection schemes with a scale of thousands of grids is shortened from several weeks by manual methods to less than 2 hours, and the layout adjustment decision for each optimization cycle can be completed within minutes. Attached Figure Description

[0019] Figure 1 This is a diagram illustrating the architecture of a big data-based intelligent site selection system for household waste sorting and disposal points, as provided in an embodiment of the present invention. Figure 2 A flowchart of multi-source big data fusion provided for embodiments of the present invention; Figure 3This is a schematic diagram of the structure of the household waste sorting demand prediction model provided in an embodiment of the present invention; Figure 4 A flowchart of a multi-objective intelligent location selection and reinforcement learning dynamic optimization algorithm provided in an embodiment of the present invention; Figure 5 A block diagram of a smart site selection device for household waste sorting and disposal points based on big data, provided in an embodiment of the present invention.

[0020] The diagram shows: 101 Data Layer, 102 Fusion Layer, 103 Prediction Layer, 104 Decision Layer, 105 Optimization Layer, 106 Application Layer, 201 Data Source Access, 202 Coordinate Unification, 203 Time Alignment, 204 Data Quality Processing, 205 Spatial Grid Mapping, 206 Spatiotemporal Dataset Construction, 301 Graph Structure Construction Unit, 302 Graph Convolutional Layer, 303 Temporal Attention Layer, 304 Multi-Task Output Layer, 401 Candidate Point Screening, 402 Demand Weight Determination, 403 Multi-Objective Model Construction, 404 Pareto Solution, 405 Compromise Decision, 406 Layout Implementation, 407 Status Acquisition, 408 Action Decision, 409 Parameter Update, 50 Overall Device, 501 Data Fusion Module, 502 Demand Prediction Module, 503 Site Selection Calculation Module, 504 Dynamic Optimization Module, 5041 Experience Replay Pool, 5042 Parameter Update Unit, 505 Output Module. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention. Where there is no conflict, the following embodiments and features can be combined with each other.

[0022] Some of the key terms of this invention are defined as follows: The reward function consists of a weighted delivery distance item, an overflow penalty item, a layout adjustment cost item, and a complaint penalty item. After normalization, each item is summed according to the weight coefficients set by management preferences.

[0023] Prediction interval / upper quantile value: The prediction interval refers to the range of values ​​under a given confidence level when making a probabilistic prediction of waste generation; the upper quantile value refers to the upper boundary value of the prediction interval, which is used as the basis for determining the amount of waste to be generated at the disposal point during capacity verification.

[0024] Mixing index: The mixing index is the Shannon diversity index of the number of each type of POI, which is the negative of the sum of the proportions of each type of POI multiplied by their logarithms; NIMBY Negative Effect: The NIMBY negative effect of a candidate delivery point is equal to the sum of its negative effect components on each sensitive target. Each component is positively correlated with the service scale of the candidate delivery point and decreases as the square of the distance from the candidate delivery point to the sensitive target increases. Sensitive targets include residential buildings, schools, and hospitals, with the weights decreasing in that order. Layout adjustment cost item: A fixed cost equivalent is assigned according to the type of adjustment action. The cost equivalent of migration action is higher than that of capacity adjustment action, and the cost equivalent of activation and deactivation action is in between. The cost equivalent is calculated by the unit cost of construction and operation and maintenance rate of sanitation facilities in the target area.

[0025] First, combined Figure 1 The system architecture common to various embodiments of the present invention will be described. For example... Figure 1 As shown, the system implementing the method of this invention comprises, from bottom to top, a data layer 101, a fusion layer 102, a prediction layer 103, a decision layer 104, an optimization layer 105, and an application layer 106. The data layer 101 accesses seven types of heterogeneous data sources: demographic data, mobile phone signal population heatmap data, point of interest (POI) data, building outline data, road network data, historical household waste sorting and disposal measurement data, and waste collection vehicle trajectory data. The fusion layer 102 performs coordinate unification, time alignment, missing value completion, and outlier removal on the data, and maps the data to spatial grid cells to form a multi-source fused spatiotemporal dataset. The prediction layer 103 runs a spatiotemporal prediction model, outputting grid-level and category-specific predicted values ​​and prediction intervals for waste generation. The decision layer 104 constructs and solves a multi-objective site selection optimization model, outputting an initial layout scheme. The optimization layer 105 runs a reinforcement learning agent, continuously outputting layout adjustment actions based on operational feedback. Application layer 106 provides sanitation management departments with interactive functions such as electronic map layer overlay, layout comparison charts and adjustment suggestion lists, and feeds back actual operation data to data layer 101 to form a closed loop.

[0026] Combination Figure 2 The multi-source big data fusion process in step S1 is explained. For example... Figure 2As shown, data fusion sequentially includes data source access 201, coordinate unification 202, time alignment 203, data quality processing 204, spatial gridding mapping 205, and spatiotemporal dataset construction 206. Coordinate unification 202 transforms various data sources from coordinate systems such as WGS-84 and GCJ-02 to the projected coordinate system of the target area; time alignment 203 aggregates data from different sampling frequencies into a preset time slice, such as 1 hour or 1 day; in data quality processing 204, missing values ​​are filled using kriging interpolation based on spatially adjacent grids or collaborative filtering based on matrix factorization, and outliers exceeding physically reasonable ranges or contradicting multi-source cross-validation results are removed; spatial gridding mapping 205 divides the target area into square grids of preset side lengths or irregular grids based on community boundaries and road-enclosed areas, with the preferred grid side length for urban built-up areas being 100 to 300 meters. In rural and town areas, the distance can be widened to 300 to 500 meters. During mapping, mobile phone signal population heat map data is aggregated to each grid according to time slices to obtain the heat map values ​​of permanent residents and migrant populations. POI data is categorized into residential, commercial, school, and office types, and the number of each type of POI within each grid is counted, and the mixing index is calculated. Building outline data is used to calculate the total building area, plot ratio, and residential building area ratio of each grid. Historical waste sorting and disposal measurement data is mapped to the corresponding grid according to the disposal point and aggregated into a time series of waste generation by four categories: recyclables, kitchen waste, hazardous waste, and other waste. Collection vehicle trajectory data is used to extract the actual collection frequency and collection time of each disposal point. The spatiotemporal dataset is constructed by organizing the above gridding results into a tensor structure of "grid number × time slice × feature dimension" for storage.

[0027] Combination Figure 3 The demand prediction model for household waste sorting in step S2 is explained. For example... Figure 3As shown, the spatiotemporal prediction model includes a graph structure building unit 301, a graph convolutional layer 302, a temporal attention layer 303, and a multi-task output layer 304. The graph structure building unit 301 uses spatial grid cells as nodes and road network reachability relationships between nodes as edges. The edge weights of the adjacency matrix are determined by the reciprocal of the shortest path distance in the road network and then symmetrically normalized. The graph convolutional layer 302 performs spatial aggregation of node features based on the normalized adjacency matrix, so that the feature expression of each grid integrates the demand information of its reachable neighborhood. A two-layer graph convolutional structure can be used, with the ReLU activation function used between layers. The temporal attention layer 303 applies multi-head self-attention calculation to the time series features of each node output by the graph convolutional layer. The number of attention heads is preferably 4 to 8 to simultaneously capture... The model analyzes the daily, weekly, and holiday fluctuation patterns of waste generation. The multi-task output layer 304 includes four output branches corresponding to recyclables, kitchen waste, hazardous waste, and other waste. Each branch outputs the 0.1, 0.5, and 0.9 quantile values ​​for the corresponding category's generation using quantile regression. The 0.5 quantile is used as the predicted generation value, the interval between the 0.1 and 0.9 quantiles is used as the prediction interval, and the 0.9 quantile is used as the capacity verification input. This allows for reserving facility capacity under demand bias, which is one of the direct causes of the decrease in peak overflow rates. Model training uses a multi-source fusion spatiotemporal dataset from historical periods, with the sum of quantile losses from each branch as the optimization objective, and employs the Adam optimizer iteratively until convergence. The covariates input to the model include at least population heat values, POI mixing index, holiday indicator variables, and meteorological data.

[0028] Combination Figure 4 The algorithm flow for steps S3 and S4 is explained. For example... Figure 4As shown, the process includes: Candidate point screening 401, removing locations within the preset safety distances of green spaces, water areas, fire lanes, building entrances, and unusable land parcels from each grid based on land suitability rules, thus obtaining a set of candidate disposal points; Demand weight determination 402, weighting and summing the predicted amounts of various types of waste generated by each grid in a preset future time period according to weight coefficients set for the collection frequency and unit collection cost of each category, thus obtaining the demand weight of each grid; Multi-objective model construction 403, with the objective functions being minimizing the weighted disposal distance for residents, minimizing facility construction and operation costs, and minimizing the NIMBY (Not In My Backyard) negative effects, where residents... The weighted delivery distance is the sum of the products of the demand weight of each grid and the shortest walking distance along the road network from that grid to its assigned delivery point. The NIMBY (Not In My Backyard) negative effect is determined based on the proximity of candidate delivery points to residential buildings, schools, and hospitals, as well as the service scale of the candidate delivery points. Constraints include: the walking distance from any grid to its assigned delivery point not exceeding the preset service radius; the predicted output value assigned to any delivery point not exceeding the capacity limit; and the total number of delivery points not exceeding the maximum number corresponding to the budget constraint. The Pareto solution is 404. A non-dominated sorting genetic algorithm with an elite retention strategy is used, with the start / stop state of candidate delivery points encoded as chromosomes. The dominance ranking and congestion calculation iteratively evolve, and the mutation probability is adaptively adjusted according to the demand density of candidate placement points during the crossover and mutation operation to obtain the Pareto solution set; compromise decision 405, the objective function values ​​of each solution in the Pareto solution set are normalized, the objective weight of each objective is determined using the entropy weight method, the relative closeness of each solution to the positive and negative ideal solutions is calculated, and the solution with the largest relative closeness is selected as the compromise site selection scheme, i.e., the initial layout scheme; after layout implementation 406, the dynamic optimization stage begins: status acquisition 407 collects the metering data, overflow sensor data and complaint records of each placement point according to the preset optimization cycle, and combines them with the predicted generation value to form the current status. State; Action Decision 408: The trained reinforcement learning agent outputs a layout adjustment action based on the current state. When the reinforcement learning agent uses a deep Q-network, it calculates the action value of each candidate action in the current state and executes the layout adjustment action with the highest action value. When the reinforcement learning agent uses a proximal policy optimization algorithm, the policy network outputs the probability distribution of each candidate action in the current state and executes the layout adjustment action with the highest probability. Parameter Update 409: The actual feedback of the execution result is stored in the experience replay pool, and the agent network parameters are incrementally updated. Then, the state acquisition 407 is returned to start the next optimization cycle.

[0029] In step S4, the reinforcement learning agent can be trained using a deep Q-network or a near-end policy optimization algorithm. Training is divided into an offline training phase and an online optimization phase. The offline training phase is conducted in a simulation environment constructed based on a multi-source fusion spatiotemporal dataset. This simulation environment models demand fluctuations, overflow events, and disposal behavior based on historical waste generation sequences, allowing the agent to learn layout adjustment strategies under low-cost trial-and-error conditions. In the online optimization phase, the agent makes rolling decisions directly based on real-world operational feedback. The weight coefficients in the reward function are set according to the management preferences of the target area. For example, the weight of complaint penalties can be increased for NIMBY-sensitive areas, and the weight of overflow penalties can be increased for areas with strained collection resources. The preset optimization period is one week to one quarter, which can be set by those skilled in the art based on the implementation cost of facility adjustments and the frequency of demand fluctuations.

[0030] Example 1: Urban Community Scenario This embodiment takes the main urban area of ​​a prefecture-level city as the application scenario. The area covers 38.6 square kilometers, includes 42 communities, and has a resident population of about 510,000. Household waste sorting has been implemented for more than two years, and relatively complete historical waste disposal measurement data has been accumulated.

[0031] In step S1, the area is divided into 986 square grids of 200m x 200m each. The data sources include: community-specific population statistics from the district statistics department; mobile phone signaling data from telecom operators, which, after anonymization, is aggregated into population heatmaps for each grid at 15-minute granularity; 47,000 POI data from map service providers; building outline vector data from the natural resources department; open road network data; weighing and measurement data recorded over the past two years by 218 smart waste sorting and disposal devices within the area, categorized into four types of waste, totaling approximately 38 million disposal records; and GPS trajectory data from 12 collection vehicles of sanitation companies. After coordinate unification, time alignment using 1-hour time slices, and Kriging interpolation to complete approximately 2.3% of the missing grid data, a multi-source fusion spatiotemporal dataset of "986 grids × 24 months × 38 feature dimensions" is constructed.

[0032] In step S2, a spatiotemporal graph is constructed using 986 grid nodes, and the edge weights of the adjacency matrix are normalized by the reciprocal of the shortest path distance in the road network. The spatiotemporal prediction model employs a two-layer graph convolution, an 8-head temporal attention mechanism, and four quantile output branches, trained on data from the previous 18 months and validated on data from the following 6 months. Validation results show that the MAPE (Modular Action Per Expectation) for the predicted generation of the four types of waste in each grid for the coming week is 10.7%, with the highest accuracy for food waste (MAPE 8.9%), while the accuracy for hazardous waste is relatively lower (MAPE 15.2%) due to its smaller generation volume and greater dispersion. This meets the accuracy requirements for demand input in site selection applications. The prediction output reveals a bimodal distribution pattern in the area during weekdays from 7:00 AM to 9:00 AM and from 6:00 PM to 8:00 PM, as well as a fluctuation pattern where grids with high commercial mix show a 23% increase in generation on weekends.

[0033] In step S3, after eliminating unsuitable locations based on land suitability rules, 412 candidate placement points were obtained. The preset service radius was 300 meters (approximately 5 minutes' walk), the maximum capacity of a single point was set at serving no more than 1200 people, and the total number of placement points was budgeted to be no more than 260. A non-dominated sorting genetic algorithm with an elite retention strategy was used to solve the problem. The population size was 200, and after 500 iterations, a Pareto solution set containing 47 non-dominated solutions was obtained. A compromise solution obtained through entropy weight-TOPSIS decision-making was used to place 236 placement points. The average weighted walking distance for residents was 187 meters, which is 24.6% shorter than the original solution in the area based on the empirical rule of "one point for every 150 households". 97.3% of the households are within a 300-meter service radius, and there are no placement points within 50 meters of NIMBY sensitive targets (schools, hospitals, kindergartens).

[0034] In step S4, during the offline training phase, the deep Q-network agent is trained for 3000 rounds in a simulation environment. The simulation environment overlays measured fluctuation noise according to the predicted demand distribution to simulate overflow and delivery behavior. After implementation, the scheme is run on a rolling basis with a two-week optimization cycle: In the third optimization cycle, the agent identifies that the utilization rate of three delivery points around a commercial street exceeds 92% continuously during weekday evening peak hours, with an average of 4.2 overflow events per week. It outputs the adjustment action of "activating one nearby candidate point, expanding the capacity of that point from 4 bins to 8 bins, and simultaneously deactivating a delivery point 300 meters away with a utilization rate of less than 25%"; after execution, the overflow events in the area drop to zero in the next cycle, and the maintenance resources saved from the deactivated point are used to add new points. After six months of operation, statistics show that the overflow event rate of delivery points in the entire area has decreased by 38.5% compared to the fixed layout before optimization, the difference coefficient of facility utilization rate in various communities has decreased from 0.61 to 0.34, the monthly average number of resident complaints has decreased by 42%, and the average daily collection mileage per collection vehicle has decreased by 11.8%. The calculation of the entire initial site selection scheme took 1.7 hours, and the action decision time for each optimization cycle was no more than 3 minutes.

[0035] Example 2: Rural market town scene This embodiment takes a township in the eastern coastal area of ​​a province as the application scenario. The township has one market town and 17 administrative villages, with a registered population of about 43,000. The houses are scattered and the road network is sparse. There is a custom of going to the market on the third and eighth days of the lunar calendar. The population fluctuates significantly in time and space. Moreover, the garbage disposal metering data only covers 12 disposal points in the market town area.

[0036] To address the sparse data characteristics of rural scenarios, this embodiment adapts the method as follows: Spatial grid units are divided using a combination of 500m x 500m square grids and village boundaries, resulting in 214 grid units. To address insufficient coverage of waste collection and metering data, the amount of waste generated in each village is inverted using loading sensor records from collection vehicle trajectory data and collection point information, and spatial calibration is performed by combining this with the per capita waste generation rate obtained from household sampling surveys. Missing value completion employs matrix decomposition and collaborative completion, using population heatmaps and homestead building outlines as collaborative features. Market day fluctuations are characterized by adding a market day indicator variable to the covariates, allowing the time-series attention layer to learn a cyclical pattern where the waste generation in the town grid on market days is approximately 65% ​​higher than on weekdays.

[0037] In the multi-objective site selection phase, considering the travel habits of rural residents, the service radius was widened to 500 meters. Simultaneously, the "avoidance of the orientation range of residential entrances and exits" and "avoidance of farmland, water areas, and village road access sections" were incorporated into the land use suitability rules. In the calculation of the NIMBY (Not In My Backyard) negative effect, a residential orientation factor was introduced, imposing additional penalties on candidate points located within the orientation sector of the main residential entrances and exits. The resulting compromise solution deployed 89 placement points. In the town area, 31 points were densely deployed with a service radius of 200 meters, and in each village, 58 points were deployed according to the principle of "multiple points per village, along main village roads." The average placement distance for residents was shortened from 412 meters in the original plan to 296 meters, and all points avoided farmland and drinking water source protection areas.

[0038] During the dynamic optimization phase, the intelligent agent identified two types of drift: First, during the busy farming season, the peak collection times in some villages occurred earlier in the morning and evening. The agent adjusted the collection frequency of the relevant collection points from once every two days to once a day. Second, after the occupancy rate of a newly built centralized farmer residence increased, the utilization rate of the collection points originally attached to the old village decreased, and littering appeared around the centralized residence. The agent output the action of "relocating 2 collection points to within 150 meters of the entrance and exit of the centralized residence, facing away from them." After the relocation, complaints about littering in the area dropped from an average of 11 per month to 1 within a month. After one year of operation, the accuracy rate of waste sorting in the whole town increased from 71% to 86%, and the empty collection rate decreased by 14%.

[0039] Example 3: New Urban Area Scene This embodiment takes a newly built area of ​​a national-level new district as the application scenario. The area has a planned area of ​​24 square kilometers and a planned population of 300,000. The land parcels that have been transferred are in the phased construction stage and there is no historical garbage disposal measurement data. It belongs to the cold start scenario.

[0040] To address the lack of historical data, this embodiment employs the transfer learning and rolling calibration mechanism described in claim 7: A mature urban area already developed in the city is selected as the source region. The similarity between this source region and the current area in terms of population age structure, residential and commercial land ratio, and climate conditions is assessed by cosine similarity and reaches 0.87, meeting the preset threshold of 0.8. The graph convolutional layer parameters of the spatiotemporal prediction model trained in the source region are transferred to the model in the current area. The transferred model is calibrated using the planned population import plan (inhabitation population by year and by plot), planned land use data (plot ratio and business type of each plot), and prior demand estimates from analogous grids. The analogous grids refer to grids in the source region with similar land use to the target grid, whose historical demand is converted to a population size ratio and used as priors. The calibrated model outputs grid-level demand predictions for the first batch of delivered plots in the current area, thereby completing the first round of site selection: 47 deployment points are set up around the first batch of 8 residential communities, 2 schools, and 1 commercial complex, with an average deployment distance of 205 meters.

[0041] As the delivery of the area progressed, actual delivery measurement data accumulated monthly, and the proportion of actual data in the model retraining samples was gradually increased according to a preset ratio: 20% in the first quarter, increasing by 20% each subsequent quarter, reaching 100% in the fifth quarter, thus completing a smooth transition from "transferring priors" to "local realities" for the model. During this process, the reinforcement learning agent continuously adjusted its deployment on a monthly optimization cycle. One month before the delivery of the second batch of residences, based on updated demand forecasts, it pre-activated six delivery points from the preset candidate points around the second batch of delivered communities, ensuring that new residents were within a 250-meter service radius in the first week of occupancy, avoiding the common problem of "missing delivery points in the early stages of occupancy" in newly built communities. After one year of operation, the average delivery distance for residents in this area stabilized within 210 meters, and the demand forecast MAPE converged from 19.4% at the initial cold start to 11.3%, verifying the effectiveness of the transfer learning calibration mechanism and rolling closed-loop optimization. Example

[0042] Combination Figure 5 The apparatus of the present invention will be described below. Figure 5As shown, the intelligent site selection device 50 for household waste sorting and disposal points based on big data includes a data fusion module 501, a demand prediction module 502, a site selection calculation module 503, a dynamic optimization module 504, and an output module 505. The data fusion module 501 is connected to an external data source and the internal bus of the device to execute step S1; the demand forecasting module 502 is connected to the data fusion module 501 to execute step S2; the site selection calculation module 503 is connected to the demand forecasting module 502 to execute step S3; the dynamic optimization module 504 is connected to the site selection calculation module 503 and the data fusion module 501 to execute step S4, and its internal components further include an experience playback pool 5041 and a parameter update unit 5042. The experience playback pool 5041 is used to store layout adjustment actions and their actual feedback, and the parameter update unit 5042 is used to incrementally update the network parameters of the reinforcement learning agent based on the data in the experience playback pool at the end of each optimization cycle; the output module 505 is connected to the site selection calculation module 503 and the dynamic optimization module 504 to output the initial layout scheme and the site selection scheme corresponding to the layout adjustment actions to a display device or sanitation management business system in at least one of the following forms: electronic map layer overlay, layout comparison chart, and adjustment suggestion list.

[0043] The above modules can be implemented through software, hardware, or a combination of both. In one implementation, device 50 is a server or edge computing device, including a memory and a processor. When a computer program stored in the memory is executed by the processor, it implements the steps of the above method. The computer program can also be stored in a computer-readable storage medium for distribution and deployment.

[0044] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent site selection of household waste sorting and disposal points based on big data, characterized in that, Includes the following steps: Step S1, Multi-source Big Data Fusion: Acquire data from multiple heterogeneous data sources within the target area. The heterogeneous data sources include at least historical household waste sorting and disposal measurement data and road network data, and include at least one of the following: population statistics, mobile phone signal population heat map data, point of interest (POI) data, building outline data, and waste collection vehicle trajectory data. Perform preprocessing on the data from each data source, including coordinate unification, time alignment, missing value completion, and outlier removal. Map the preprocessed data to spatial grid cells of a preset scale to construct a multi-source fused spatiotemporal dataset with spatial grid cells as the basic unit and time slices as the sequence. Step S2, Prediction of Household Waste Sorting Demand: Construct a spatiotemporal graph structure using the spatial grid units as nodes and the road network reachability relationships between the spatial grid units as edges. Input the historical waste generation sequence and corresponding covariates of each spatial grid unit in the multi-source fusion spatiotemporal dataset into the trained spatiotemporal prediction model, and output the predicted generation values ​​and prediction intervals of each waste category for each spatial grid unit in the future preset time period, thereby obtaining the spatiotemporal distribution data of waste sorting demand; wherein, the spatiotemporal prediction model includes a graph convolutional layer for extracting spatial dependencies and a temporal attention layer for extracting temporal dependencies; Step S3, Multi-objective Intelligent Site Selection: Determine the demand weight of each spatial grid unit based on the spatiotemporal distribution data of waste sorting demand. Within the target area, candidate disposal points are screened according to land suitability rules. The objective functions are minimizing the weighted disposal distance of residents, minimizing the construction and operation and maintenance costs of facilities, and minimizing the NIMBY (Not In My Backyard) negative effect. The constraints are service coverage, disposal point capacity, and number of disposal points. The predicted value of the amount of various types of waste generated at any disposal point is determined according to the upper quantile value of the prediction interval. The upper quantile value is used to verify the upper limit of the capacity of the disposal point. A multi-objective site selection optimization model is constructed. The Pareto solution set is obtained by solving the multi-objective site selection optimization model. A compromise site selection scheme is determined from the Pareto solution set as the initial layout scheme of the household waste sorting disposal points. Step S4, reinforcement learning dynamic optimization: The layout adjustment process of household waste sorting and disposal points is modeled as a Markov decision process. The state space includes at least the predicted generation value of each spatial grid cell, the utilization rate of each disposal point, the number of overflow events, and the number of resident complaints. The action space includes at least the activation, deactivation, relocation, and capacity adjustment of disposal points. The reward function includes at least the weighted disposal distance term, the overflow penalty term, the layout adjustment cost term, and the complaint penalty term. A reinforcement learning agent is trained based on the Markov decision process. After the initial layout scheme is put into operation, the actual disposal feedback data is collected and the state is updated according to the preset optimization cycle. The trained reinforcement learning agent outputs layout adjustment actions to perform rolling dynamic optimization of the layout of household waste sorting and disposal points.

2. The intelligent site selection method for household waste sorting and disposal points based on big data according to claim 1, characterized in that, In step S1, the preprocessed data is mapped to spatial grid cells of a preset scale. Specifically, this includes: dividing the target area into square grids of preset side lengths or irregular grids based on community boundaries and road enclosures; aggregating the mobile phone signal population heat map data into each spatial grid cell by time slice to obtain the heat map values ​​of the resident population and the transient population of each spatial grid cell; calculating the number and mixing index of each type of POI data in each spatial grid cell according to residential, commercial, school, and office categories; using the building outline data to calculate the total building area, plot ratio, and residential building area ratio of each spatial grid cell; mapping the historical household waste sorting and disposal measurement data to the corresponding spatial grid cells according to the disposal point affiliation, and aggregating them according to four categories: recyclables, kitchen waste, hazardous waste, and other waste to form a time series of waste generation; and using Kriging interpolation based on spatially adjacent grids or collaborative filtering based on matrix decomposition to complete missing values.

3. The intelligent site selection method for household waste sorting and disposal points based on big data according to claim 1, characterized in that, In step S2, the spatiotemporal prediction model includes a graph convolutional layer, a temporal attention layer, and a multi-task output layer connected in sequence. The graph convolutional layer spatially aggregates node features using a normalized road network reachability adjacency matrix, and the edge weights of the adjacency matrix are determined based on the reciprocal of the shortest path distance between spatial grid cells. The temporal attention layer applies multi-head self-attention computation to the time series features of each node output by the graph convolutional layer to capture the daily, weekly, and holiday fluctuation patterns of waste generation. The multi-task output layer includes multiple output branches that correspond one-to-one with waste categories. Each output branch outputs at least two quantile values ​​of the waste generation for the corresponding category, and the quantile values ​​represent the prediction range of the generation prediction value. The covariates include at least population heat value, POI mixing index, holiday indicator variable and meteorological data.

4. The intelligent site selection method for household waste sorting and disposal points based on big data according to claim 1, characterized in that, In step S3, the demand weight is determined by the weighted sum of the predicted values ​​of the amount of various types of waste generated in each spatial grid unit within a future preset time period. The weight coefficient of each type of waste is set according to the collection frequency and unit collection cost of the corresponding category. The weighted disposal distance of residents is the sum of the products of the demand weight of each spatial grid unit and the shortest walking distance along the road network from the spatial grid unit to its assigned disposal point. The NIMBY (Not In My Backyard) negative effect is determined based on the proximity of candidate disposal points to residential buildings, schools, and hospitals, as well as the service scale of the candidate disposal points. The constraints include: the walking distance from any spatial grid cell to its assigned disposal point does not exceed the preset service radius; the predicted waste generation of any disposal point does not exceed the capacity limit of that disposal point; and the total number of disposal points does not exceed the maximum number corresponding to the budget constraint.

5. The intelligent site selection method for household waste sorting and disposal points based on big data according to claim 1, characterized in that, In step S3, the multi-objective location optimization model is solved using a non-dominated sorting genetic algorithm with an elite retention strategy. The start / stop status of each candidate placement point is encoded as a chromosome, and the objective function is used as the fitness evaluation criterion. Iterative evolution is achieved through non-dominated sorting and crowding calculation, and the mutation probability is adaptively adjusted based on the demand density of the candidate placement points during crossover and mutation operations to obtain a Pareto solution set. A multi-attribute decision-making method combining entropy weighting and approximation of ideal solution sorting is used to determine the compromise location scheme from the Pareto solution set. This involves normalizing the objective function values ​​of each solution in the Pareto solution set, determining the objective weights of each objective using entropy weighting, calculating the relative proximity of each solution to the positive and negative ideal solutions, and selecting the solution with the highest relative proximity as the compromise location scheme.

6. The intelligent site selection method for household waste sorting and disposal points based on big data according to claim 1, characterized in that, In step S4, the reinforcement learning agent is trained using a deep Q-network or a proximal policy optimization algorithm. The training process includes an offline training phase and an online optimization phase. The offline training phase is conducted in a simulation environment constructed based on the multi-source fusion spatiotemporal dataset. The simulation environment simulates demand fluctuations, overflow events, and disposal behaviors based on historical waste generation sequences. The weight coefficients of the weighted disposal distance term, overflow penalty term, layout adjustment cost term, and complaint penalty term in the reward function are set according to the management preferences of the target area. The preset optimization cycle is from one week to one quarter. At the end of each optimization cycle, the status is updated based on the collection of disposal point measurement data, overflow sensor data, and complaint records during that cycle. When the reinforcement learning agent uses a deep Q-network, it calculates the action value of each candidate action in the current state and executes the layout adjustment action with the highest action value. When the reinforcement learning agent uses a proximal policy optimization algorithm, the policy network outputs the probability distribution of each candidate action in the current state and executes the layout adjustment action with the highest probability. The actual feedback of the execution results is stored in the experience replay pool to incrementally update the network parameters of the reinforcement learning agent.

7. The intelligent site selection method for household waste sorting and disposal points based on big data according to any one of claims 1 to 6, characterized in that, When the target area is a newly established area with no historical household waste sorting and disposal measurement data, step S2 further includes: selecting an existing area that meets a preset threshold in terms of similarity to the target area in terms of population structure, land use composition, and climate conditions as the source area; transferring the graph convolutional layer parameters of the spatiotemporal prediction model trained in the source area to the spatiotemporal prediction model of the target area; and calibrating the transferred model using the planned population import data, planned land use data, and prior demand estimation of the analog grid of the target area; as actual disposal measurement data accumulates in the target area, gradually increasing the proportion of actual data in the model retraining samples according to a preset ratio to achieve rolling calibration of demand prediction.

8. A smart site selection device for household waste sorting and disposal points based on big data, characterized in that, include: The data fusion module is used to acquire data from multiple heterogeneous data sources within the target area. These heterogeneous data sources include at least historical household waste sorting and disposal measurement data and road network data, and at least one of the following: population statistics, mobile phone signal population heat map data, point of interest (POI) data, building outline data, and waste collection vehicle trajectory data. The module performs preprocessing on the data from each data source, including coordinate unification, time alignment, missing value completion, and outlier removal. The preprocessed data is then mapped to spatial grid cells of a preset scale to construct a multi-source fused spatiotemporal dataset with spatial grid cells as the basic unit and time slices as the sequence. The demand prediction module is used to construct a spatiotemporal graph structure with the spatial grid units as nodes and the road network reachability relationships between the spatial grid units as edges. It inputs the historical waste generation sequence of each spatial grid unit in the multi-source fusion spatiotemporal dataset and the corresponding covariates into a trained spatiotemporal prediction model, and outputs the predicted generation values ​​and prediction intervals for each waste category in each spatial grid unit within a preset future time period, thus obtaining the spatiotemporal distribution data of waste classification demand. The spatiotemporal prediction model includes a graph convolutional layer for extracting spatial dependencies and a temporal attention layer for extracting temporal dependencies. The site selection calculation module is used to determine the demand weight of each spatial grid unit based on the spatiotemporal distribution data of waste sorting demand, screen candidate disposal points within the target area according to land suitability rules, and construct a multi-objective site selection optimization model with the objectives of minimizing the weighted disposal distance of residents, minimizing the construction and operation and maintenance costs of facilities, and minimizing the NIMBY (Not In My Backyard) negative effect, and with service coverage, disposal point capacity, and number of disposal points as constraints. The multi-objective site selection optimization model is solved to obtain the Pareto solution set, and a compromise site selection scheme is determined from the Pareto solution set as the initial layout scheme of household waste sorting disposal points. The dynamic optimization module is used to model the layout adjustment process of household waste sorting and disposal points as a Markov decision process. The state space includes at least the predicted generation value of each spatial grid cell, the utilization rate of each disposal point, the number of overflow events, and the number of resident complaints. The action space includes at least the activation, deactivation, relocation, and capacity adjustment of disposal points. The reward function includes at least a weighted disposal distance term, an overflow penalty term, a layout adjustment cost term, and a complaint penalty term. A reinforcement learning agent is trained based on the Markov decision process. After the initial layout scheme is put into operation, the module collects actual disposal feedback data to update the state according to a preset optimization cycle. The trained reinforcement learning agent outputs layout adjustment actions to perform rolling dynamic optimization of the layout of household waste sorting and disposal points. The output module is used to output the initial layout scheme and the location scheme corresponding to the layout adjustment action.

9. The intelligent site selection device for household waste sorting and disposal points based on big data as described in claim 8, characterized in that, The spatiotemporal prediction model in the demand prediction module includes a graph convolutional layer, a temporal attention layer, and a multi-task output layer connected in sequence. The multi-task output layer includes multiple output branches corresponding one-to-one with waste categories, and each output branch outputs at least two quantile values ​​of the waste generation amount for the corresponding category. The dynamic optimization module also includes an experience replay pool and a parameter update unit. The experience replay pool is used to store layout adjustment actions and their actual feedback. The parameter update unit is used to incrementally update the network parameters of the reinforcement learning agent based on the data in the experience replay pool at the end of each optimization cycle. The output module is also used to output the site selection scheme in at least one of the following forms: electronic map layer overlay, layout comparison chart, and adjustment suggestion list.

10. An electronic device, characterized in that, The device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, it implements the steps of the intelligent site selection method for household waste sorting and disposal points based on big data as described in any one of claims 1 to 7.

Images