An ANN-based prediction method for the delivery volume of low-value recyclables in communities

Abstract

The present invention provides a method for predicting the delivery volume of low-value recyclables in communities based on ANN, aiming to optimize the resource allocation and layout strategies of intelligent recycling systems. By collecting the historical delivery data and its external feature data of low-value recyclable delivery stations covering multiple communities, data preprocessing is carried out, and an ANN model is trained. This model can capture the complex pattern relationships of residents' delivery behaviors and provide short-term accurate predictions for the delivery volume of each recyclable delivery station in the next 3 to 7 days. In addition, the comprehensive quality score and convenience features of the community are selected as the two major bases for community classification, and based on the K-means clustering method, the formulation of differentiated waste collection and resource scheduling strategies for communities is realized, thereby improving the operational efficiency and sustainable development ability of intelligent recycling systems.

Description

Technical Field

[0001] The present invention relates to the technical field of solid waste management and intelligent recycling systems, and particularly to a method for predicting the delivery volume of low-value recyclables in communities based on ANN. Background Art

[0002] In recent years, the rapid development of cities and the improvement of residents' living standards have led to a sharp increase in garbage production, and garbage disposal and recycling are facing great pressure. Low-value recyclables refer to solid waste with low economic value and easy to be discarded, such as waste paper boxes, waste plastics, glass bottles, foams, etc. Such solid waste is everywhere in daily life, but its economic value is low and the recycling cost is relatively high. At the same time, the traditional recycling method of low-value recyclables has low efficiency and serious resource waste, posing a non-negligible pressure on the environment and resources. To address the above problems, intelligent recycling systems based on the Internet of Things have been widely used in recent years. The system promotes residents' active participation in recycling by providing an incentive mechanism. By setting up intelligent garbage recycling bins in the community, residents can deliver low-value recyclables by themselves, realizing the automation of the recycling process. However, intelligent recycling systems usually optimize the recycling strategy by adopting a fixed time and space layout, and it is difficult to achieve the optimal configuration in terms of equipment layout and waste collection and transportation scheduling. Therefore, how to use advanced data prediction algorithms to accurately predict the delivery volume of low-value recyclables has become a key link in the optimization management of intelligent recycling systems.

[0003] In the field of intelligent recycling systems, there are still many limitations in the prediction of the delivery volume of low-value recyclables in the existing technology. Traditional prediction methods, such as the ARIMA (Autoregressive Integrated Moving Average) model and simple linear regression model, although widely used in some fields, essentially assume a linear relationship between data. However, the delivery volume of low-value recyclables is affected by various complex factors, such as weather, holidays, etc., and the relationship between these factors and the delivery volume is often non-linear. Therefore, linear models show obvious deficiencies in dealing with such complex relationships, and it is difficult to capture the internal pattern of the data, resulting in low prediction accuracy. At the same time, with the popularization of intelligent recycling systems, the amount and dimension of data are increasing continuously. Traditional methods are prone to the "curse of dimensionality" when dealing with high-dimensional data, the model complexity and calculation cost increase significantly, and the prediction accuracy will also decline due to data sparsity.

[0004] In summary, there are obvious deficiencies in the existing technology in predicting the delivery volume of low-value recyclables, mainly reflected in the insufficient prediction accuracy caused by the difficulty of traditional methods in capturing complex relationships and dealing with high-dimensional data. These deficiencies limit the optimization and popularization of intelligent recycling systems, and there is an urgent need for a new prediction method that can effectively handle complex non-linear relationships. Summary of the Invention

[0005] The objective of the present invention is to provide a prediction method for the delivery volume of low-value recyclables in communities based on ANN, which uses ANN (Artificial Neural Network) to process and analyze the delivery data of multiple sites in the community, realizes the prediction of the daily delivery volume, optimizes the resource allocation and layout strategy of the intelligent recycling system, improves the operation efficiency and promotes the sustainable development of the community, and solves the problem of insufficient prediction accuracy in the layout and waste collection strategy optimization of the existing intelligent recycling system.

[0006] On the one hand, the present invention provides a prediction method for the delivery volume of low-value recyclables in communities based on ANN, including the following steps:

[0007] Data collection: Obtain the historical delivery data, external feature data, and community feature data of the delivery sites of low-value recyclables covering multiple communities; wherein, the external feature data includes weather information, weekday status, and festival promotion information, and the community feature data includes community economic features and convenience features;

[0008] Data preprocessing: For the historical delivery data and external feature data, use the Lagrange interpolation method to fill in the missing values, use the box plot method to detect and remove outliers, and use the Max-Min method for normalization processing to ensure data quality and consistency;

[0009] Model establishment: Design a multi-layer ANN structure, which includes an input layer, two hidden layers, and an output layer; wherein, the input layer receives multi-dimensional feature vectors, the hidden layer uses the non-linear activation function ReLU to capture the complex pattern relationships of residents' delivery behaviors, and the output layer provides the prediction results for the daily delivery volume of specific low-value recyclable delivery sites;

[0010] Model training: During the training process, use the Adam optimizer to adjust the parameters of the ANN model, and optimize the hyperparameter configuration through cross-validation; at the same time, use a variety of prediction evaluation metrics to comprehensively evaluate the prediction accuracy of the ANN model;

[0011] Multi-site prediction: Use the sliding window method to generate time series samples, and use the trained ANN model to independently predict the delivery volume of each low-value recyclable delivery site within the next 3 to 7 days to obtain short-term prediction data;

[0012] Community difference analysis: First, based on the economic characteristics of the community, calculate the comprehensive quality scores of each community through the entropy method. Secondly, combine the convenience characteristics and use the K-means clustering algorithm to statistically analyze the comprehensive quality scores of the community. Finally, classify the community types according to the scores; based on the classification results, formulate targeted resource scheduling strategies according to historical delivery data and short-term prediction data, and optimize the collection frequency and facility layout.

[0013] Optionally, the method for obtaining external feature data and community feature data in data collection is specifically to obtain detailed weather conditions, temperature, wind force level, community housing price, community housing age, community greening rate, and community plot ratio from open Internet data sources, as well as custom working days, festivals, or promotional days.

[0014] Optionally, the data preprocessing also includes encoding conversion of text data or categorical data for easy model input.

[0015] Optionally, the first hidden layer in model establishment contains 256 neurons, the second hidden layer contains 128 neurons, and the initial learning rate is set to 0.001.

[0016] Optionally, in model training, the batch size of the training process is set to 32, and the maximum number of iterations is 500 times to ensure full training of the model while avoiding overfitting.

[0017] Optionally, the convenience feature is measured by the number of people served by a single device.

[0018] In a second aspect, the present invention also provides a computer device, including a memory, a processor, and a computer program stored on the memory. The processor executes the computer program to implement the steps of the foregoing method for predicting the delivery volume of low-value recyclables in communities based on ANN.

[0019] In a third aspect, the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the foregoing method for predicting the delivery volume of low-value recyclables in communities based on ANN.

[0020] One or more technical solutions provided by the present invention have at least the following technical effects or advantages:

[0021] First, the present invention covers the historical delivery data and external feature data of low-value recyclable delivery stations in multiple communities, forming a complete multi-dimensional feature matrix, including external feature variables such as weather information, working day status, festival promotion information, etc. In the model construction stage, a multi-level ANN structure is adopted, and the non-linear activation function ReLU is used to capture the complex patterns of residents' delivery behaviors, reducing the prediction error.

[0022] In addition, the ANN model developed in the present invention can effectively process high-dimensional data through feature input and parameter optimization, avoid the "curse of dimensionality", and ensure that the model can still maintain a high prediction accuracy under the condition of data sparsity. During the training process of the model, the cross-validation method is adopted to optimize the parameter relationship of the model, and a variety of prediction evaluation indicators are used to comprehensively evaluate the prediction accuracy of the model. Finally, by predicting the delivery volume of each low-value recyclable delivery site in the next 3 to 7 days, it assists in adjusting the resource scheduling strategy, optimizing the cleaning frequency and facility layout, so as to improve the operation efficiency and reduce resource waste.

[0023] Therefore, the present invention adopts a prediction model based on ANN, which significantly improves the prediction accuracy of the delivery volume of low-value recyclables. Compared with other models in the prior art, the ANN model shows obvious advantages in terms of prediction error. The MAPE value of the ANN model developed in the present invention is only 0.38, which is significantly lower than that of the LSTM (Long Short-Term Memory) model, random forest model, XGBoost (eXtreme Gradient Boosting) model, and SVR (Support Vector Regression) model. Description of the Drawings

[0024] Figure 1 It is a flowchart of the ANN-based prediction method for the delivery volume of community low-value recyclables in the embodiment of the present invention;

[0025] Figure 2 It is a network structure diagram of the ANN model in the embodiment of the present invention;

[0026] Figure 3 It is a comparison chart of the predicted delivery volume and the actual delivery volume of a specific low-value recyclable delivery site in the embodiment of the present invention;

[0027] Figure 4 It is a bubble chart of the five-category community division situation in the embodiment of the present invention. Detailed Embodiments

[0028] In order to overcome the limitations of traditional prediction methods, the present invention introduces ANN on the basis of traditional methods. By integrating historical delivery volume, external feature data (weather information, weekday status, festival promotion information, etc.), a multi-site-level ANN model capable of predicting delivery volume is constructed. The ANN model captures the complex pattern relationships of residents' delivery behaviors through a multi-layer network, can provide high-precision prediction results, support dynamic optimization of resource allocation, and thus achieve efficient and sustainable operation management.

[0029] A prediction method for the delivery volume of low-value recyclables in communities based on ANN provided by the present invention includes six steps: data collection, data preprocessing, model establishment, model training, multi-site prediction, and community difference analysis.

[0030] Next, please refer to Figure 1 , and describe Embodiment 1 of the present invention in detail.

[0031] 1. Data collection;

[0032] In the present invention, historical delivery data of low-value recyclables from 436 low-value recyclable delivery sites covering 173 communities in a certain city from January 1, 2023 to March 31, 2024 are collected through Internet of Things intelligent devices, mainly including the following core data:

[0033] 1) Site number: used to identify each intelligent recycling device;

[0034] 2) Daily delivery volume: the total weight of low-value recyclables received by each device per day, in kilograms;

[0035] 3) Other residents' delivery behavior characteristics: such as user number, street where located, community where located, delivery frequency, remuneration obtained, etc.

[0036] To capture external factors affecting delivery behavior, the present invention obtains external feature data by combining open data sources. Among them, weather information is obtained from "China Weather Network" through Internet crawlers, including descriptive information of weather conditions (such as sunny, cloudy, overcast, light rain, heavy rain, etc.), temperature (average value, in °C), and wind force level (1-6 levels). In addition, date information is also obtained through Internet crawlers, including whether it is a working day, festival or promotion day (such as "618", "Double 11", "Double 12"), as well as periodic characteristics of year, month, and season.

[0037] Finally, to apply the ANN model of the present invention to each community, it is necessary to achieve community clustering based on community economic characteristics and convenience characteristics, so as to realize the difference analysis of the historical values and predicted values of the delivery volume of low-value recyclables in different categories of communities. Community economic characteristics and convenience characteristics are collected with 173 communities as the smallest analysis unit. Community economic characteristics are obtained from Anjuke website through Internet crawlers, including community housing price, community housing age, community greening rate, and community plot ratio. The convenience characteristics are calculated from the data provided by the community property, including the number of people in the community and the number of recycling sites laid in the community, and then the number of people served by a single device is obtained, and the number of people served by a single device is used to measure the convenience characteristics. Finally, based on the comprehensive quality scores and convenience characteristics of each community, 173 communities are divided into 5 categories based on the K-means clustering analysis method.

[0038] After collection and processing, individual external feature data and historical delivery volume data are integrated into a complete set of feature vectors, denoted as X = {x1, x2, …, x d}, where d is the number of features. Combining with data entries, a multi-dimensional feature matrix D ∈ X n×d is finally formed, where n is the total number of records. Each record contains the daily delivery volume of low-value recyclable delivery stations and the input external feature variable data for that day.

[0039] 2. Data preprocessing;

[0040] To ensure the quality of the model input data, a series of preprocessing is required for the collected data.

[0041] First, for the descriptive information of weather conditions collected (such as sunny, cloudy, overcast, light rain, heavy rain, etc.), it is converted into numerical form during the data preprocessing stage. The specific coding rules are as follows: "1" represents sunny, cloudy, overcast or changing from cloudy to overcast; "2" represents fog + cloudy or fog + overcast; "3" represents light snow, light rain or shower; "4" represents heavy rain, rainstorm, moderate snow, heavy snow or blizzard. For the date feature, whether it is a working day, a festival or a promotion day is also numerically processed, where "yes" is marked as 0 and "no" is marked as 1. In the year feature, "2023" is marked as 0 and "2024" is marked as 1; the month is marked as 1 - 12; the season is marked as 1 - 4 according to the international astronomical division standard.

[0042] Secondly, it is to fill in the missing values. For the occasionally missing numerical values in the weather information (such as precipitation or temperature), the Lagrange interpolation method is used to complete the time series.

[0043] For outliers (such as extreme values in the daily delivery volume), the present invention detects abnormal data through the IQR (Interquartile Range) box plot method and eliminates or smooths the abnormal data. The interquartile range is set as ΔQ = Q3 - Q1, where Q3 is the third quartile and 75% of the data in the dataset is less than or equal to this value; Q1 is the first quartile and 25% of the data in the dataset is less than or equal to this value. When a data point x satisfies x > Q3 + 1.5×ΔQ or x < Q1 - 1.5×ΔQ, that is, when the daily delivery volume exceeds 1.5 times the interquartile range, it is regarded as an outlier and eliminated.

[0044] Finally, since the data value ranges of different features vary greatly, before inputting into the ANN model, it is also necessary to normalize the data to eliminate the dimensional differences between different features. The normalization process uses the Max - Min method to normalize the data into the interval [0, 1]. The specific formula is as follows:

[0045] ,

[0046] Among them, is the normalized feature data, is the original feature data, and are the minimum and maximum values of this feature, respectively.

[0047] 3. Model establishment and training;

[0048] The present invention uses an ANN for multi-site delivery volume prediction. As Figure 2 shown, the model consists of an input layer, two hidden layers, and an output layer. The main purpose of its network structure design is to capture the complex non-linear relationship of the delivery volume of low-value recyclables.

[0049] The input of the input layer is a multi-dimensional feature vector D ∈ X n×d , where n is the total number of records and d is the number of features. Let the weight matrix of the hidden layer be , the bias term be , and the mapping from input to output is realized through a non-linear activation function. The calculation formula of the hidden layer is:

[0050] ,

[0051] ,

[0052] Among them, represents the layer index of the neural network; is the result of the linear transformation of the th layer (the value before passing through the activation function); is the weight matrix of the th layer, and its size depends on the number of input neurons and output neurons of this layer; is the output of the th layer; is the output of the th layer (the value after passing through the activation function); is the bias term of the th layer, which is used to adjust the activation value of the neuron; is the activation function, which is used to introduce non-linear characteristics.

[0053] The hidden layer uses the ReLU (Rectified Linear Unit) activation function. ReLU can effectively avoid the problem of gradient disappearance and improve the training efficiency. Its form is:

[0054] ,

[0055] Among them, is the input value (i.e., )。ReLU ensures that negative values become zero, avoiding the vanishing gradient problem and improving training efficiency.

[0056] The output layer of the model is a one-dimensional predicted value, representing the delivery volume of a specific low-value recyclable delivery site on a certain day. The output layer uses a linear activation function:

[0057] ,

[0058] where, is the final output of the model, that is, the predicted value of the delivery volume of a specific low-value recyclable delivery site on a certain day, represents the last layer (output layer) of the neural network.

[0059] The model is optimized using the Adam optimizer, and its update rule is:

[0060] ,

[0061] ,

[0062] ,

[0063] where, is the model parameter at the th iteration; and are the first-order and second-order momenta respectively; represents the gradient of the loss function with respect to the parameter ; is the learning rate; , are momentum hyperparameters, usually set to = 0.9, = 0.999; is a small number to prevent the denominator from being zero, usually taking the value of 10 -8 .

[0064] 4. Model performance evaluation;

[0065] During the training process, the preprocessed data is divided into a training set and a test set, where 70% of the data is used for training and 30% of the data is used for testing. The training objective is to enable the model to accurately predict the delivery volume on the test set. During training, the batch size is set to 32 and the maximum number of iterations is 500 to ensure sufficient training of the model. The standardized data is used as the input, and the prediction results are restored to the original scale through inverse transformation.

[0066] The five-fold cross-validation method is adopted to evaluate the model performance and ensure the generalization ability of the prediction results. The grid search method is used to optimize hyperparameters such as the number of neurons in the hidden layer and the learning rate. Finally, the number of neurons in the hidden layer is determined to be {256, 128} (the first hidden layer contains 256 neurons, and the second hidden layer contains 128 neurons), and the initial learning rate is 0.001.

[0067] To more comprehensively evaluate the performance of the model for predicting the delivery volume of low-value recyclables, four widely used prediction evaluation metrics are adopted: the coefficient of determination R 2 , the mean absolute error MAE, the mean absolute percentage error MAPE, and the root mean square error RMSE. The closer the value of R 2 is to 1, the closer the predicted value in the test set is to the true value, indicating better fitting performance of the prediction model; the lower the values of the other three prediction evaluation metrics MAE, MAPE, and RMSE, the higher the prediction accuracy of the model. The calculation formulas for the prediction evaluation metrics are as follows:

[0068] ,

[0069] ,

[0070] ,

[0071] ,

[0072] Among them, is the actual delivery volume, is the predicted delivery volume, is the mean of all actual delivery volumes, is the number of samples.

[0073] Through training, the ANN model of the present invention has a high coefficient of determination on the test set and excellent error metrics, where MAPE is 0.38 and RMSE is 13.14, significantly superior to traditional time series models and other machine learning algorithms.

[0074] Such as Figure 3The comparison chart of the predicted delivery volume and the actual delivery volume of specific low-value recyclables shown. The horizontal axis represents time, with the unit of days, and the vertical axis represents the delivery volume, with the unit of kg. The solid line with dots represents the actual delivery volume (true value), and the dashed line with crosses represents the predicted delivery volume (predicted value). From the overall trend, the change trends of the predicted value and the actual value are basically the same, indicating that the prediction model performs well in capturing the overall trend. At some obvious peaks and valleys, the fitting of the predicted value and the actual value is good, indicating that the model can also predict well in these extreme cases. Generally speaking, the ANN prediction model can largely reflect the actual change of the delivery volume, reasonably grasps the data trend, and the actual value and the predicted value are relatively close, indicating that the prediction accuracy of the ANN prediction model is relatively high.

[0075] 5. Multi-site prediction;

[0076] In the multi-site prediction of the present invention, for 436 low-value recyclable delivery sites covering multiple communities, by constructing an ANN model, the delivery volume of each site in the next 3 to 7 days is predicted. The present invention takes the historical delivery data and external characteristic data of the site as input. After the model is trained, it can capture the complex dependence relationship between the delivery behavior and the external characteristic data. After generating time series samples by the sliding window method, the ANN model makes independent predictions based on the site-specific input features and generates the predicted daily delivery volume values for the next few days.

[0077] 6. Community difference analysis;

[0078] In the present invention, in order to comprehensively analyze the community differences, it is necessary to effectively classify the communities. The present invention selects two main classification bases: the comprehensive quality score of each community and the convenience characteristics. These two bases respectively reflect the economic level of the community and the delivery convenience of participating in intelligent recycling. The community economic quality score is obtained by comprehensively scoring four community economic characteristics, namely the housing price, the housing age, the greening rate, and the plot ratio of the community, by the entropy method. The convenience characteristic, that is, the number of people served by a single device, is directly related to the convenience of residents using intelligent recycling facilities and has an important impact on the recycling efficiency of the community.

[0079] The entropy method is a method that can objectively assign weights and is applicable to measuring the economic characteristics of different communities. Defining the comprehensive score of the four community economic characteristics as the comprehensive quality score of each community to reflect the economic level of the community. The entropy method assigns weights by calculating the dispersion degree of the community economic characteristics, ensuring the relative objectivity of each community economic characteristic in the comprehensive score.

[0080] First, standardize each community economic characteristic to eliminate the influence of dimension. For each community economic characteristic, calculate its entropy value, and the formula is as follows:

[0081] ,

[0082] wherein, is the entropy value of the th community economic feature; is the standardized value of the th community under the th community economic feature; is the original value of the th community under the th community economic feature; is the total number of communities.

[0083] Finally, calculate the weight of each community economic feature according to the entropy value. The specific calculation formula is as follows:

[0084] ,

[0085] wherein, is the weight of the th community economic feature, is the total number of community economic features.

[0086] Thus, the comprehensive quality scores of each community obtained by the entropy method can objectively reflect the economic level of each community and are then used for subsequent clustering analysis. Subsequently, the K-means clustering algorithm is used to classify the communities using the comprehensive quality scores and convenience features of each community. The K-means clustering algorithm performs clustering by minimizing the within-cluster sum of squares SSE. Its calculation formula is:

[0087] ,

[0088] wherein, is the number of clusters, is the th cluster, is the cluster center, is the data point.

[0089] In the clustered community groups, the collection situation of low-value recyclables within each type of community will show similar characteristics. As Figure 4 shown, this figure shows the relationship between the comprehensive quality scores of different community types and the number of people served by a single device. The abscissa represents the comprehensive quality score, reflecting the economic level of the community; the ordinate represents the number of people served by a single device, that is, the average number of residents served by each recycling device, reflecting the delivery convenience. The size of the bubble reflects the number of people served by a single device, and the color of the bubble reflects different community types. It can be seen from Figure 4 that the present invention obtains 5 types after clustering 173 communities. Combining the historical data situation and the prediction results, the following descriptions are obtained for these 5 types of communities:

[0090] The first type of community (low quality - low convenience), with a relatively low economic level and poor delivery convenience, and relatively scarce resource allocation;

[0091] The second type of community (low quality - high convenience), with a relatively low economic level but high delivery convenience and relatively good facility coverage;

[0092] The third type of community (medium quality - medium convenience), with a medium level in both economic level and delivery convenience, showing a relatively balanced performance;

[0093] The fourth type of community (high quality - low convenience), with a relatively high economic level but insufficient delivery convenience and short - board in facility coverage;

[0094] The fifth type of community (high quality - high convenience), with excellent performance in both economy and delivery convenience, perfect recycling facilities and good resource allocation.

[0095] In summary, based on the accurate prediction of the delivery volume of each site by the high - precision ANN model developed in the present invention, the prediction results of the site delivery volume are further classified and summarized according to the community type, so as to provide targeted optimization suggestions for resource allocation in the middle - end waste collection and transportation link of the intelligent recycling system. Specifically, different types of communities can adopt different waste collection and transportation strategies and resource allocation plans according to their economic levels and delivery convenience to improve the overall operation efficiency and recycling effect.

[0096] Example 2, The embodiment of the present invention also provides a computer device, including a memory, a processor, and a computer program stored on the memory, and the processor executes the computer program to implement the steps of a method for predicting the delivery volume of low - value recyclables in communities based on ANN in the foregoing embodiment.

[0097] Example 3, The embodiment of the present invention also provides a computer - readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of a method for predicting the delivery volume of low - value recyclables in communities based on ANN in the foregoing embodiment are implemented.

[0098] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these changes and modifications.

Claims

1. A prediction method for the delivery volume of low-value recyclables in a community based on ANN, characterized in that, Processing and analyzing community multi-site delivery data using an ANN includes the following steps: Data collection: Obtain historical delivery data, external feature data, and community feature data of low-value recyclable delivery sites covering multiple communities; wherein, the external feature data includes weather information, weekday status, and festival promotion information, and the community feature data includes community economic features and convenience features; the methods for obtaining external feature data and community feature data in the data collection are specifically to obtain detailed weather conditions, temperature, wind force level, community housing price, community housing age, community greening rate, and community floor area ratio from open Internet data sources, as well as custom weekdays, festivals, or promotion days; the convenience features are calculated from data provided by the community property management, including the number of people in the community and the number of recycling sites laid out in the community, and then the number of people served per device is obtained; the convenience features are measured by the number of people served per device; the number of people served per device is the average number of residents served by each recycling device, reflecting the delivery convenience; Data preprocessing: For historical delivery data and external feature data, use the Lagrange interpolation method for missing value filling, use the box plot method for outlier detection and removal, and use the Max-Min method for normalization processing to ensure data quality and consistency; Model establishment: Design a multi-layer ANN structure, which includes an input layer, two hidden layers, and an output layer; wherein, the input layer receives multi-dimensional feature vectors, the hidden layers use the non-linear activation function ReLU to capture the complex pattern relationships of residents' delivery behaviors, and the output layer provides prediction results for the daily delivery volume of specific low-value recyclable delivery sites; in the model establishment, the first hidden layer contains 256 neurons, the second hidden layer contains 128 neurons, and the initial learning rate is set to 0.001; Model training: During the training process, use the Adam optimizer to adjust the parameters of the ANN model and optimize the hyperparameter configuration through cross-validation; at the same time, use a variety of prediction evaluation metrics to comprehensively evaluate the prediction accuracy of the ANN model; in the model training, the batch size of the training process is set to 32, and the maximum number of iterations is 500 times to ensure full training of the model while avoiding overfitting; Multi-site prediction: Use the sliding window method to generate time series samples, and use the trained ANN model to independently predict the delivery volume of each low-value recyclable delivery site within the next 3 to 7 days to obtain short-term prediction data; Community difference analysis: First, calculate the comprehensive quality scores of each community through the entropy method based on community economic features, then combine the convenience features and use the K-means clustering algorithm to statistically analyze the community comprehensive quality scores, and finally classify the community types according to the scores; based on the classification results, formulate targeted resource scheduling strategies according to historical delivery data and short-term prediction data to optimize the collection frequency and facility layout.

2. The method according to claim 1, characterized in that, The data preprocessing also includes encoding conversion of text data or categorical data for easy model input.

3. A computer device, comprising a memory, a processor, and a computer program stored on the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-2.

4. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-2.

Images