A pm2.5 prediction method and system of an adaptive geographic weighted artificial neural network model

Abstract

The application discloses a PM2.5 prediction method and system of an adaptive geographical weighted artificial neural network model, and relates to the technical field of spatial data analysis. Specifically, the method comprises the following steps: a data acquisition step: acquiring historical PM2.5 data sets of a to-be-measured region, and dividing the historical PM2.5 data sets into a training data set and a test data set; a step of constructing an AS-GWNN model: constructing a network architecture of the AS-GWNN model, including an input layer, a hidden layer and an output layer; a training step: inputting the training data set into the network architecture of the AS-GWNN model to obtain a trained AS-GWNN model; and a test step: inputting the test data set into the trained AS-GWNN model to obtain a predicted PM2.5 value. The application solves the problems that machine learning technology ignores the inherent spatial non-stationarity of spatial data, and that regression models considering spatial non-stationarity lack nonlinear fitting capability.

Description

Technical Field

[0001] This invention belongs to the field of spatial data analysis technology, and in particular relates to a PM2.5 prediction method and system based on an adaptive geographic weighted artificial neural network model. Background Technology

[0002] Machine learning excels at accurately modeling nonlinear relationships. Many machine learning techniques prioritize training global parameters, often ignoring the inherent spatial nonstationarity of spatial data. This limitation poses a significant obstacle to generating accurate estimates for specific geographic regions. Furthermore, traditional regression models that consider spatial nonstationarity, such as Drift Analysis of Regression Parameters, intentionally incorporate location information as an auxiliary parameter into ordinary linear regression. While this extended parameter estimation can initially simulate spatial variability, it remains a simple trend-fitting method and cannot represent complex spatial parameter variations. Spatially Adaptive Filtering emerged to compensate for spatial drift in regression coefficients, but it lacks the ability to statistically test its parameter estimates. Random Coefficient Models and Multilevel Models estimate randomly distributed regression coefficients based on Baves theory, but the assumption of parameter independence does not fully hold in spatial data.

[0003] Therefore, proposing an adaptive geographic weighted artificial neural network model for PM2.5 prediction and a system to address the difficulties of existing technologies is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] This invention addresses the problems of machine learning techniques neglecting the inherent spatial non-stationarity of spatial data and regression models that consider spatial non-stationarity lacking nonlinear fitting capabilities by proposing an adaptive geographic weighted artificial neural network model for PM2.5 prediction.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A PM2.5 prediction method using an adaptive geographic weighted artificial neural network model includes the following steps:

[0007] Data acquisition steps: Obtain the historical PM2.5 dataset of the area to be tested, and divide the historical PM2.5 dataset into a training dataset and a test dataset;

[0008] Steps to build an AS-GWNN model: Construct the network architecture of the AS-GWNN model, including the input layer, hidden layer, and output layer;

[0009] Input layer: used for assembling the independent variables;

[0010] Hidden layer: Stores geographic neurons, and the geographic weighted activation function is synchronously set on the corresponding geographic neurons;

[0011] Output layer: Corresponds to the output of the dependent variable, and contains only one neuron;

[0012] AS-GWNN model training steps: Input the training dataset into the network architecture of the AS-GWNN model to obtain the trained AS-GWNN model;

[0013] AS-GWNN model testing steps: Input the test dataset into the trained AS-GWNN model to obtain the predicted PM2.5 value.

[0014] Optionally, the PM2.5 dataset can be divided into training data and prediction data at a ratio of 1:20 using the Mini-Batch SGD algorithm.

[0015] Optionally, in the above method, the geographic neurons in the AS-GWNN model construction step include GWAF, GWA, and geographic bias units.

[0016] The above method, optionally, involves the following steps for training the AS-GWNN model: setting different values ​​for the number of geographic neurons, inputting the training dataset into the AS-GWNN model, selecting the optimal value for the number of geographic neurons, and finally evaluating the AS-GWNN model using the AICc metric to obtain the trained AS-GWNN model.

[0017] Optionally, the formula for collecting the optimal number of geographic neurons in the AS-GWNN model training step is as follows:

[0018]

[0019]

[0020] Where GN is the number of geographic neurons in the neural network, TN is the number of cells in the target remote sensing image, and scale is the scaling factor; TR is the resolution of the target remote sensing image, bw is the spatial bandwidth, exp is the exponential function, and gw is the vector number. nearest For nearest neighbor spatial weights, ON is the number of cells in the original remote sensing image, ln is the logarithmic function, and GN is the nearest neighbor spatial weight. optimal The optimal number of geographic neurons.

[0021] The above method may optionally include the following independent variables in the input layer: AOD, NDVI, TMP, SPFH, WIND, and HGT.

[0022] The geographic-weighted activation function in the hidden layer is expressed as:

[0023]

[0024]

[0025] Among them, net ij This represents the input signal quantity of training sample i on the j-th geographic neuron in the hidden layer; gw ij The spatial weights between training sample i and hidden layer neuron j; dist ij denoted by , represents the Euclidean distance between training sample i and hidden layer neuron j; bw is the spatial bandwidth, a positive decay parameter that controls the functional relationship between spatial weights and distance.

[0026] An adaptive geographic weighted artificial neural network model PM2.5 prediction system, implementing the aforementioned adaptive geographic weighted artificial neural network model PM2.5 prediction method, includes a data acquisition module, a data partitioning module, a model building module, a training module, and an output value prediction module connected in sequence;

[0027] Data acquisition module: used to acquire historical PM2.5 datasets for the area to be tested;

[0028] Data partitioning module: used to partition the historical PM2.5 dataset into training data set and test data set;

[0029] Model building module: used to build AS-GWNN models;

[0030] Training module: Used to train the AS-GWNN model using the training dataset;

[0031] Output value prediction module: Used to input the test dataset into the AS-GWNN model to obtain PM2.5 prediction values.

[0032] As can be seen from the above technical solution, compared with the prior art, the present invention provides a PM2.5 prediction method and system based on an adaptive geographic weighted artificial neural network model, which has the following beneficial effects:

[0033] This invention addresses the problems of machine learning techniques neglecting the inherent spatial nonstationarity of spatial data and regression models that consider spatial nonstationarity lacking nonlinear fitting capabilities. Furthermore, the training process of the neural network in this invention is dominated by spatial weights, enabling the new model to effectively explore complex geographical processes. The new network architecture significantly saves computer storage space and computing time. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0035] Figure 1 This invention discloses a PM2.5 prediction method and system flowchart using an adaptive geographic weighted artificial neural network model.

[0036] Figure 2 This is a block diagram of an adaptive geographic weighted artificial neural network model for PM2.5 prediction disclosed in this invention.

[0037] Figure 3 This is a diagram of the AS-GWNN network architecture disclosed in this invention;

[0038] Figure 4 The training curves of the AS-AGWNN model under different geographical neuron counts disclosed in this invention are shown.

[0039] Figure 5 This is a scatter plot of training results under different geographical neuron numbers disclosed in this invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] In this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0042] Reference Figure 1 As shown, an adaptive geographic weighted artificial neural network model for PM2.5 prediction includes the following steps:

[0043] Data acquisition steps: Obtain the historical PM2.5 dataset of the area to be tested, and divide the historical PM2.5 dataset into a training dataset and a test dataset;

[0044] Steps to build an AS-GWNN model: Construct the network architecture of the AS-GWNN model, including the input layer, hidden layer, and output layer;

[0045] Input layer: used for assembling the independent variables;

[0046] Hidden layer: Stores geographic neurons, and the geographic weighted activation function is synchronously set on the corresponding geographic neurons;

[0047] Output layer: Corresponds to the output of the dependent variable, and contains only one neuron;

[0048] AS-GWNN model training steps: Input the training dataset into the network architecture of the AS-GWNN model to obtain the trained AS-GWNN model;

[0049] AS-GWNN model testing steps: Input the test dataset into the trained AS-GWNN model to obtain the predicted PM2.5 value.

[0050] Furthermore, the PM2.5 dataset was divided into training data and prediction data at a ratio of 1:20 using the Mini-Batch SGD algorithm.

[0051] Furthermore, in the steps of constructing the AS-GWNN model, the geographic neurons include GWAF, GWA, and geographic bias units.

[0052] Furthermore, the specific steps for training the AS-GWNN model are as follows: set different values ​​for the number of geographic neurons, input the training dataset into the AS-GWNN model, select the optimal value for the number of geographic neurons, and finally evaluate the AS-GWNN model using the AICc index to obtain the trained AS-GWNN model.

[0053] Furthermore, it is characterized by,

[0054] The formula for collecting the optimal number of geographic neurons in the training steps of the AS-GWNN model is as follows:

[0055]

[0056]

[0057] Where GN is the number of geographic neurons in the neural network, TN is the number of cells in the target remote sensing image, and scale is the scaling factor; TR is the resolution of the target remote sensing image, bw is the spatial bandwidth, exp is the exponential function, and gw is the vector number. nearest For nearest neighbor spatial weights, ON is the number of cells in the original remote sensing image, ln is the logarithmic function, and GN is the nearest neighbor spatial weight. optimal The optimal number of geographic neurons.

[0058] Furthermore, the independent variables in the input layer include: AOD, NDVI, TMP, SPFH, WIND, and HGT;

[0059] The geographic-weighted activation function in the hidden layer is expressed as:

[0060]

[0061]

[0062] Among them, net ij This represents the input signal quantity of training sample i on the j-th geographic neuron in the hidden layer; gw ij The spatial weights between training sample i and hidden layer neuron j; dist ij denoted by , represents the Euclidean distance between training sample i and hidden layer neuron j; bw is the spatial bandwidth, a positive decay parameter that controls the functional relationship between spatial weights and distance.

[0063] Specifically, the regression process used in the AS-GWNN model is represented by a matrix equation:

[0064]

[0065] Among them, GW i The matrix represents the spatial weights between training sample i and all hidden layer geographic neurons, T is the matrix transpose, b is the bias coefficient, and X is the weights of the training sample i and all hidden layer geographic neurons. i For, W [1] for, for;

[0066] X i The expression is: X i =[1 x i1 … x ip ];

[0067] W [1] The expression is:

[0068] The expression is:

[0069] like Figure 2As shown, an adaptive geographic weighted artificial neural network model PM2.5 prediction system implements the PM2.5 prediction method of the adaptive geographic weighted artificial neural network model described above, including a data acquisition module, a data partitioning module, a model building module, a training module, and an output value prediction module connected in sequence.

[0070] Data acquisition module: used to acquire historical PM2.5 datasets for the area to be tested;

[0071] Data partitioning module: used to partition the historical PM2.5 dataset into training data set and test data set;

[0072] Model building module: used to build AS-GWNN models;

[0073] Training module: Used to train the AS-GWNN model using the training dataset;

[0074] Output value prediction module: Used to input the test dataset into the AS-GWNN model to obtain PM2.5 prediction values.

[0075] Specifically: The AS-GWNN model has a single-neuron output layer, uses the L2 norm loss function, and selects the NAdam algorithm as the learning rate optimizer. When machine learning deals with regression problems, common evaluation metrics include MAE, MSE, RMSE, R2, and adjusted R2.

[0076] Example 1:

[0077] To find the optimal number of geographic neurons for the AS-GWNN model, we used the eight-year average from 2015 to 2022 as training and prediction data. The training sample consisted of 2137 samples, and the prediction sample consisted of 407660 samples. Eight control groups were set for the number of geographic neurons: 10, 100, 1000, 2000, 3000, 4000, 5000, and 0000. First, the optimal value (bw) was found using GWR, and then incorporated into the hyperparameters of the AS-GWNN model to calculate the spatial weights between sample points and geographic neurons. Under a uniform background condition of lr = 0.0003, the AS-GWNN model was trained with eight different values ​​for the number of geographic neurons, yielding the results. Figure 4 .

[0078] Here, AICc and adjR2 are chosen as evaluation metrics for the model, and they begin to converge after a certain number of epochs. As clearly shown in Figure 4, the larger the number of geographic neurons, the later the convergence, and the more drastic the fluctuations in the learning curve. When the number of geographic neurons is less than 1000, the model's adjR2 does not reach 0.500, indicating insufficient fitting accuracy. While when the number of geographic neurons exceeds 4000, although the model's adjR2 can still remain above 0.900, the fluctuations in the learning curve are too large. The optimal number of geographic neurons for the model should be selected between 1000 and 4000.

[0079] Figure 5 The presentation shows scatter plots of predicted and sampled PM2.5 values ​​from different models, along with model evaluation scores. GWR shows good data distribution, with an adjR² of 0.948. The performance of the AS-GWNN model varies significantly depending on the number of geographic neurons. When GN=10, the adjR² is only 0.083; when GN is magnified 10 times, the adjR² rapidly increases to 0.479; when GN=1000, the adjR² is comparable to GWR; when 2000<=GN<=5000, the adjR² surpasses GWR; and when GN=10000, the adjR² slightly decreases. Increasing the training epochs slightly improves the fitting accuracy of AS-GWNN.

[0080] As shown in Table 1, the performance comparison data of the PM2.5 inference inversion model under different geographical neuron numbers can be obtained, thus demonstrating the advantages of the present invention.

[0081] Table 1: Performance Comparison of PM2.5 Inference and Inversion Models under Different Geographic Numbers of Neurons

[0082]

[0083] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to the method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0084] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A PM2.5 prediction method using an adaptive geographic weighted artificial neural network model, characterized in that, Includes the following steps: Data acquisition steps: Obtain the historical PM2.5 dataset of the area to be tested, and divide the historical PM2.5 dataset into a training dataset and a test dataset; Steps to build an AS-GWNN model: Construct the network architecture of the AS-GWNN model, including the input layer, hidden layer, and output layer; Input layer: used for assembling independent variables; the independent variables in the input layer include: AOD, NDVI, TMP, SPFH, WIND, HGT; Hidden layer: Stores geographic neurons, and the geographic weighted activation function is synchronously set on the corresponding geographic neurons; the geographic weighted activation function in the hidden layer is expressed as: in, This represents the input signal quantity of training sample i on the j-th geographic neuron in the hidden layer; The spatial weights between training sample i and hidden layer neuron j; This represents the Euclidean distance between training sample i and hidden layer neuron j; Spatial bandwidth is a positive decay parameter that controls the functional relationship between spatial weights and distance; Output layer: Corresponds to the output of the dependent variable, and contains only one neuron; AS-GWNN model training steps: Input the training dataset into the network architecture of the AS-GWNN model to obtain the trained AS-GWNN model; the formula for obtaining the optimal number of geographic neurons in the AS-GWNN model training steps is as follows: Where GN is the number of geographic neurons in the neural network, TN is the number of cells in the target remote sensing image, and scale is the scaling factor; TR is the resolution of the target remote sensing image, bw is the spatial bandwidth, exp is the exponential function, and gw is the vector number. nearest For nearest neighbor spatial weights, ON is the number of cells in the original remote sensing image, ln is the logarithmic function, and GN is the nearest neighbor spatial weight. optimal The optimal number of geographic neurons; AS-GWNN model testing steps: Input the test dataset into the trained AS-GWNN model to obtain the predicted PM2.5 value.

2. The PM2.5 prediction method using an adaptive geographic weighted artificial neural network model according to claim 1, characterized in that, The PM2.5 dataset was divided into training data and prediction data at a ratio of 1:20 using the Mini-Batch SGD algorithm.

3. The PM2.5 prediction method based on an adaptive geographic weighted artificial neural network model according to claim 1, characterized in that, The geographic neurons in the AS-GWNN model construction process include GWAF, GWA, and geographic bias units.

4. The PM2.5 prediction method using an adaptive geographic weighted artificial neural network model according to claim 1, characterized in that, The specific steps for training the AS-GWNN model are as follows: set different values ​​for the number of geographic neurons, input the training dataset into the AS-GWNN model, select the optimal value for the number of geographic neurons, and finally evaluate the AS-GWNN model using the AICc metric to obtain the trained AS-GWNN model.

5. A PM2.5 prediction system based on an adaptive geographic weighted artificial neural network model, characterized in that, A PM2.5 prediction method based on an adaptive geographic weighted artificial neural network model according to any one of claims 1-4 includes a data acquisition module, a data partitioning module, a model building module, a training module, and an output value prediction module connected in sequence. Data acquisition module: used to acquire historical PM2.5 datasets for the area to be tested; Data partitioning module: used to partition the historical PM2.5 dataset into training data set and test data set; Model building module: used to build AS-GWNN models; Training module: Used to train the AS-GWNN model using the training dataset; Output value prediction module: Used to input the test dataset into the AS-GWNN model to obtain PM2.5 prediction values.

Images