PM2.5 Concentration Prediction Method Based on Gaussian Process Regression and Deep Learning
By combining deep learning and Gaussian process regression to predict PM2.5 concentration, and using convolutional neural networks and long short-term memory networks to build a model, the accuracy and uncertainty problems of traditional PM2.5 concentration prediction are solved, and efficient point prediction and interval prediction are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-05-10
- Publication Date
- 2026-03-10
AI Technical Summary
Existing traditional methods are ineffective in predicting PM2.5 concentrations and cannot meet the requirements of intelligent decision-making systems for accurate and uncertainty-based predictions. Single-point predictions at a single time cannot meet the modeling needs of complex attribute systems.
By combining deep learning and Gaussian process regression, a PM2.5 concentration prediction model integrating convolutional neural networks and long short-term memory networks is constructed. Interval prediction is performed through Gaussian process regression, integrating spatial information and temporal features to provide point prediction and interval prediction results.
It achieves accurate prediction of PM2.5 concentration, provides point prediction results and reliable interval prediction results, and improves the robustness of the model and the reliability of the prediction.
Smart Images
Figure CN116796805B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of software application, in particular to a PM2.5 concentration prediction method based on Gaussian process regression and deep learning. BACKGROUND
[0002] PM2.5, also known as fine particulate matter, is a kind of respirable particulate matter with an aerodynamic equivalent diameter of ≤2.5 microns in the ambient air. It can be suspended in the air for a long time and is one of the key indicators of air quality monitoring and urban air pollution. With the continuous improvement of public environmental awareness, the regional composite air pollution represented by PM2.5 has attracted widespread attention from the government and all walks of life. Therefore, establishing a reasonable and accurate PM2.5 concentration prediction model can help to develop effective preventive measures and plan government social activities, so as to avoid the harm caused by serious air pollution.
[0003] PM2.5 concentration data, as a typical spatio-temporal sequence data, has the characteristics of spatial aggregation, temporal periodicity, spatio-temporal correlation and uncertainty mutation. Therefore, the effect of predicting PM2.5 concentration by traditional numerical method is not good.
[0004] In recent years, artificial intelligence has developed rapidly and is deeply integrated with many application scenarios. Its powerful learning and reasoning capabilities have attracted the attention of many scholars. Artificial neural network, as one of the widely used artificial intelligence technologies, is an important tool for simulating nonlinear phenomena, which can make up for the shortcomings of traditional recursive methods, and is therefore suitable for PM2.5 concentration prediction. With the continuous improvement of air quality, single-point PM2.5 concentration prediction value at a single time cannot meet the requirement of providing more comprehensive information for intelligent decision-making system. Compared with single-point prediction, interval prediction can successfully capture the uncertainty of the expected result, providing a new framework for quantifying the uncertainty of time series prediction and improving the robustness of the model. Gaussian process regression, as a non-parametric probabilistic technique for non-linear regression problems, provides prediction uncertainty explained by estimated variance. In addition, the covariance function of Gaussian process regression is composed of base kernels that capture many function characteristics, which is very suitable for a series of practical prediction applications, making the hybrid model developed by Gaussian process regression highly adaptable and flexible when modeling systems with complex attributes.
[0005] Therefore, how to combine deep learning and Gaussian process regression to produce a new hybrid model that performs well in point prediction and interval prediction is a theoretical and practical engineering problem that needs to be solved. SUMMARY
[0006] The PM2.5 concentration prediction method based on Gaussian process regression and deep learning provided by the embodiment of the application comprises:
[0007] Step 1: obtaining PM2.5 concentration historical data; the PM2.5 concentration historical data comprises input variables being PM2.5 concentration true values and output variables being PM2.5 concentration prediction values;
[0008] Step 2: dividing the PM2.5 concentration historical data into a training data set and a test data set according to a preset proportion, and preprocessing the training data set and the test data set;
[0009] Step 3: constructing a PM2.5 concentration point prediction model fusing a convolutional neural network and a long short-term memory network, inputting the preprocessed training data set into the PM2.5 concentration point prediction model for multiple times of training, obtaining ideal parameters, and configuring the PM2.5 concentration point prediction model based on the ideal parameters;
[0010] Step 4: inputting the preprocessed test data set into the configured PM2.5 concentration point prediction model to obtain a point prediction result of the output variable;
[0011] Step 5: constructing a PM2.5 concentration prediction hybrid model of Gaussian process regression and deep learning, inputting the point prediction result into the PM2.5 concentration prediction hybrid model to obtain a probability distribution function and a prediction interval corresponding to the point prediction result.
[0012] Preferably, in the step 2, the preprocessing of the training data set and the test data set comprises:
[0013] screening and eliminating abnormal values in the training data set and the test data set;
[0014] normalizing the screening and eliminating result, and a normalization formula is as follows:
[0015] y * =(y-min) / (max-min)
[0016] wherein y is any original data value in the screening and eliminating result, min is the minimum original data value in the screening and eliminating result, max is the maximum original data value in the screening and eliminating result, and y * is a normalized data value corresponding to any original data value in the screening and eliminating result.
[0017] Preferably, in the step 3, the preprocessed training data set is inputted into the PM2.5 concentration point prediction model for multiple times of training to obtain ideal parameters, and the step 3 comprises:
[0018] setting a neural network parameter of the PM2.5 concentration point prediction model;
[0019] inputting the training data set into the PM2.5 concentration point prediction model for multiple times of training, obtaining an accuracy and a corresponding optimization parameter of each time of training;
[0020] taking the optimization parameter corresponding to the maximum accuracy as an ideal parameter;
[0021] The setting of the neural network parameter of the PM2.5 concentration point prediction model comprises:
[0022] setting a CNN layer initialization function of the PM2.5 concentration point prediction model as Kaiming;
[0023] setting an LSTM, CLSTM, GRU layer initialization function of the PM2.5 concentration point prediction model as Orthogonal;
[0024] setting an optimizer of the PM2.5 concentration point prediction model as Adam;
[0025] setting a learning rate of the PM2.5 concentration point prediction model as 1e-3;
[0026] setting a loss function of the PM2.5 concentration point prediction model as MSE;
[0027] setting a batch size of the PM2.5 concentration point prediction model as 20;
[0028] setting a dropout rate of the PM2.5 concentration point prediction model as 0.2.
[0029] Preferably, the step 5 further comprises:
[0030] calculating each index of the point prediction and the interval prediction, and comparing the PM2.5 concentration prediction mixed model with a CNN-GPR model, an LSTM-GPR model and a GPR model to obtain a better model,
[0031] The index of the point prediction is calculated according to the following formula:
[0032]
[0033]
[0034]
[0035]
[0036] wherein y iis the i th observation value, Var is the mean value of the i th observation value, Y i is the i th prediction value of the PM2.5 concentration prediction mixed model, CNN-GPR model, LSTM-GPR model, and GPR model, is the variance, N is the number of prediction samples, E and σ are operators, MAE is the mean absolute error, RMSE is the mean square root error, and R is the correlation coefficient.
[0037] The index of the interval prediction is calculated according to the following formula:
[0038]
[0039]
[0040]
[0041]
[0042] wherein, is the upper limit of the prediction interval of the i th point prediction value, is the lower limit of the prediction interval of the i th point prediction value, and α is the confidence level.
[0043] Preferably, in step 1, the PM2.5 concentration historical data is obtained, comprising:
[0044] A preset PM2.5 monitoring station distribution map corresponding to the target area is obtained.
[0045] The PM2.5 monitoring station distribution map is preprocessed to obtain a credible station distribution map.
[0046] Based on the credible station distribution map, a monitoring missing point is determined.
[0047] Based on the monitoring missing point and the preset regional map corresponding to the target area, a mobile monitoring route is planned.
[0048] Based on the mobile monitoring route, a mobile monitoring trolley is controlled to monitor PM2.5 at the monitoring missing point to obtain first historical monitoring data.
[0049] Second historical monitoring data of PM2.5 monitoring of each credible station in the credible station distribution map is obtained.
[0050] The first historical monitoring data and the second historical monitoring data are integrated to obtain PM2.5 concentration historical data.
[0051] Preferably, the PM2.5 monitoring station distribution map is preprocessed, comprising:
[0052] sequentially traversing each PM2.5 monitoring station in the PM2.5 monitoring station distribution map;
[0053] In each traversal, based on a preset information acquisition template, the station information of the traversed PM2.5 monitoring station is acquired;
[0054] Based on a preset first feature extraction template, the station information is feature extracted to obtain a plurality of first features;
[0055] The first features are matched with second features in a preset indication feature library. If the match is consistent, the preset second feature extraction template and feature requirement corresponding to the matching second feature are acquired;
[0056] Based on the second feature extraction template, the station information is feature extracted to obtain a plurality of third features;
[0057] If the third features do not meet the feature requirement, the traversed PM2.5 monitoring station is removed from the PM2.5 monitoring station distribution map;
[0058] When all the PM2.5 monitoring stations that need to be removed are removed, the PM2.5 monitoring station distribution map is used as a trusted station distribution map.
[0059] Preferably, based on the trusted station distribution map, a monitoring missing point is determined, comprising:
[0060] In the trusted station distribution map, each trusted station is taken as the center of a circle, and a preset radius length is taken as the radius to form a circular range;
[0061] Based on a preset grid division rule, the trusted station distribution map is divided into a plurality of grid areas;
[0062] Each grid area is sequentially traversed;
[0063] In each traversal, the remaining area in the traversed grid area except the circular range is extracted;
[0064] Based on a preset third feature extraction template, the remaining area is feature extracted to obtain a plurality of fourth features;
[0065] The fourth features are summarized to obtain a fourth feature set;
[0066] The fourth feature set is matched with a preset index feature set to obtain a matching degree;
[0067] If the matching degree is greater than or equal to a preset matching degree threshold, a monitoring missing point setting requirement is based on the remaining area to set a monitoring missing point.
[0068] The set monitoring missing point position is taken as a new trusted site, and a new circular range is made with the new trusted site as the center and the radius length as the radius.
[0069] Continuing to traverse the grid area;
[0070] The monitoring missing point position setting requirements include:
[0071] The first straight line distance between the set monitoring missing point position and the center of any circular range around the set monitoring missing point position is greater than or equal to a preset first straight line distance threshold, the minimum straight line distance between the set monitoring missing point position and any circular range around the set monitoring missing point position is greater than or equal to a preset second straight line distance threshold, and the third straight line distance between any two monitoring missing point positions is greater than or equal to a preset third straight line distance threshold.
[0072] Preferably, a mobile monitoring route is planned based on the monitoring missing point position and a preset area map corresponding to the target area, and the planning includes:
[0073] Determining a map position corresponding to the monitoring missing point position in the area map;
[0074] Determining a target map position with the minimum fourth straight line distance between the map position and a preset monitoring trolley initial position;
[0075] Taking the target map position as a route starting point, a shortest route passing through the remaining map positions is planned and is taken as a mobile monitoring route.
[0076] The present application has the following beneficial effects.
[0077] 1. Convolutional neural networks and long short-term memory networks can be used to fully extract information from process data, improving the ability of the model to handle complex nonlinear data. Gaussian process regression is used to collect data and produce accurate interval predictions.
[0078] 2. The PM2.5 concentration prediction method based on Gaussian process regression and deep learning integrates spatial information extracted by convolutional neural networks and time features extracted by long short-term memory networks, and uses a Gaussian regression process model for interval prediction, which can accurately predict PM2.5 concentration. The above method predicts the PM2.5 concentration at future times according to the daily changes in PM2.5 concentration, while giving point prediction results and corresponding interval prediction results, and the output results have good credibility.
[0079] Other features and advantages of the present application will be set forth in the description that follows, and in part will be apparent from the description, or can be learned by practice of the application. The purposes and other advantages of the present application will be realized and attained by the structure particularly pointed out in the written description and claims hereof as well as the appended drawings.
[0080] The technical solutions of the present application are described in further detail below with the help of the accompanying drawings and examples. BRIEF DESCRIPTION OF DRAWINGS
[0081] The accompanying drawings are included to provide a further understanding of the present application, and are incorporated in and constitute a part of the specification, illustrate embodiments of the present application and explain the present application, but do not limit the present application. In the drawings:
[0082] Figure 1 is a general flowchart of the method according to the present application.
[0083] Figure 2 is a CLSTM model structure diagram according to an embodiment of the present application.
[0084] Figure 3 is an implementation diagram of data set division and prediction according to an embodiment of the present application.
[0085] Figure 4 is a CLSTM and GPR hybrid model flowchart according to an embodiment of the present application.
[0086] Figure 5 is a point prediction result diagram a according to an embodiment of the present application.
[0087] Figure 6 is a point prediction result diagram b according to an embodiment of the present application.
[0088] Figure 7 is an interval prediction result diagram a according to an embodiment of the present application.
[0089] Figure 8 is an interval prediction result diagram b according to an embodiment of the present application.
[0090] Figure 9 is a point prediction result comparison table of four models according to an embodiment of the present application.
[0091] Figure 10 is an interval prediction result comparison table of four models according to an embodiment of the present application. DETAILED DESCRIPTION
[0092] The preferred embodiments of the present application are described below with the help of the accompanying drawings, and it should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and do not limit the present application.
[0093] The embodiment of the present application provides a PM2.5 concentration prediction method based on Gaussian process regression and deep learning, as shown in the formula (1), comprising the following steps. Figure 1
[0094] Step 1: obtaining PM2.5 concentration historical data; the PM2.5 concentration historical data comprises input variables which are PM2.5 concentration true values and output variables which are PM2.5 concentration prediction values;
[0095] Step 2: dividing the PM2.5 concentration historical data into a training data set and a test data set according to a preset proportion, and preprocessing the training data set and the test data set;
[0096] Step 3: constructing a PM2.5 concentration point prediction model which is fused with a convolutional neural network and a long short-term memory network, inputting the preprocessed training data set into the PM2.5 concentration point prediction model for multiple times of training, obtaining ideal parameters, and configuring the PM2.5 concentration point prediction model based on the ideal parameters;
[0097] Step 4: inputting the preprocessed test data set into the configured PM2.5 concentration point prediction model to obtain a point prediction result of the output variable; the point prediction result is a simple number, for example, 63;
[0098] Step 5: constructing a PM2.5 concentration prediction hybrid model which is fused with Gaussian process regression and deep learning, inputting the point prediction result into the PM2.5 concentration prediction hybrid model to obtain a probability distribution function and a prediction interval corresponding to the point prediction result. For example, the point prediction result is A, and the prediction interval is [A-1.96σ, A+1.96σ], wherein σ is a parameter generated by the PM2.5 concentration prediction hybrid model, and the probability distribution function is a probability distribution of the point prediction result;
[0099] In the step 2, the preprocessing of the training data set and the test data set comprises the following steps.
[0100] Screening and removing abnormal values in the training data set and the test data set; when the abnormal value is greater than a threshold value, the abnormal value is removed;
[0101] Normalizing the screening and removing result, and the normalization formula is as follows:
[0102] y * =(y-min) / (max-min)
[0103] Wherein, y is any original data value in the screening and elimination result, min is the minimum original data value in the screening and elimination result, max is the maximum original data value in the screening and elimination result, y * is a normalized data value corresponding to any original data value in the screening and elimination result.
[0104] In step 3, the preprocessed training data set is input into the PM2.5 concentration point prediction model for multiple training to obtain ideal parameters, including:
[0105] Setting the neural network parameters of the PM2.5 concentration point prediction model;
[0106] The training data set is input into the PM2.5 concentration point prediction model for multiple training to obtain the accuracy of each training and the corresponding optimization parameter;
[0107] The optimization parameter corresponding to the maximum accuracy is taken as the ideal parameter;
[0108] Wherein, setting the neural network parameters of the PM2.5 concentration point prediction model comprises:
[0109] The CNN layer initialization function of the PM2.5 concentration point prediction model is set to Kaiming;
[0110] The LSTM, CLSTM, GRU layer initialization function of the PM2.5 concentration point prediction model is set to Orthogonal;
[0111] The optimizer of the PM2.5 concentration point prediction model is set to Adam;
[0112] The learning rate of the PM2.5 concentration point prediction model is set to 1e-3;
[0113] The loss function of the PM2.5 concentration point prediction model is set to MSE;
[0114] The batch size of the PM2.5 concentration point prediction model is set to 20;
[0115] The dropout rate of the PM2.5 concentration point prediction model is set to 0.2.
[0116] The step 5 further comprises:
[0117] The indicators of point prediction and interval prediction are calculated, and the PM2.5 concentration prediction mixed model is compared with the CNN-GPR model, the LSTM-GPR model and the GPR model to obtain a better model,
[0118] The indicators of point prediction are calculated according to the following formula:
[0119]
[0120]
[0121]
[0122]
[0123] wherein, y i is the i th observation value, Var is the mean value of the i th observation value, Y i is the i th prediction value output by the PM2.5 concentration prediction hybrid model, the CNN-GPR model, the LSTM-GPR model, and the GPR model for PM2.5 prediction, is the variance, N is the number of prediction samples, E and σ are operators, MAE is the mean absolute error, RMSE is the mean square root error, and R is the correlation coefficient;
[0124] The index of the interval prediction is calculated according to the following formula:
[0125]
[0126]
[0127]
[0128]
[0129] wherein, is the upper limit of the prediction interval of the i th point prediction value, is the lower limit of the prediction interval of the i th point prediction value, and α is the confidence level.
[0130] The working principle and beneficial effects of the above technical solution are:
[0131] S100, selecting daily PM2.5 concentration historical data of an environmental monitoring station, the historical data including input variables and output variables, and dividing the data set into a training data set and a test data set, and then performing data preprocessing on the training data set and the test data set;
[0132] S200, constructing a PM2.5 concentration point prediction model fusing a convolutional neural network and a long short-term memory network, inputting the training data set into the PM2.5 concentration point prediction model for training, performing multiple tests on the model to obtain ideal parameters, and then inputting the parameters into the PM2.5 concentration point prediction model to improve the performance of the model;
[0133] S300, input the test data set into the trained PM2.5 concentration point prediction model to obtain a point prediction result of the output variable;
[0134] S400, construct a PM2.5 concentration prediction hybrid model based on Gaussian process regression and deep learning, input the point prediction result of the output variable into the trained PM2.5 concentration prediction hybrid model to obtain a point prediction result of the output variable and a corresponding probability distribution function, and a prediction interval corresponding to the point prediction result of the output variable.
[0135] For step S100
[0136] S100, select daily PM2.5 concentration historical data of an environmental monitoring station, the historical data including input variables and output variables, divide the data set into a training data set and a test data set, and then perform data preprocessing on the training data set and the test data set.
[0137] The step S100 comprises the following steps:
[0138] S110, after dividing the data set into a training data set and a test data set according to a ratio of 8:2, constructing the training data set and the test data set,
[0139] S120, screening and removing outliers in the data set, and then performing normalization processing on the data set by the following formula:
[0140] y * =(y-min) / (max-min)
[0141] Wherein, y is the original data value of the input variable, min is the minimum value of each group of original data of the input variable, max is the maximum value of each group of original data of the input variable, y * is the normalized data value corresponding to each group of original data of the input variable.
[0142] The input variable includes daily PM2.5 concentration historical data of the monitoring station, and the output variable includes daily PM2.5 concentration prediction value.
[0143] For step S200
[0144] S200, construct a PM2.5 concentration point prediction model combining convolutional neural network and long short-term memory network, input the training data set into the PM2.5 concentration point prediction model for training, perform multiple tests on the model to obtain ideal parameters, and then input the parameters into the PM2.5 concentration point prediction model to improve the performance of the model.
[0145] The step S200 comprises the following steps:
[0146] S210, setting the neural network parameters of the PM2.5 concentration point prediction model, the model comprising an input layer, a hidden layer and an output;
[0147] S220, training the PM2.5 concentration point prediction model, optimizing parameters according to the accuracy of multiple training, and finally obtaining the optimal parameters and the optimal model.
[0148] The parameters include the number of neurons in the input layer, the number of neurons in the hidden layer, the number of neurons in the output layer, the learning rate, the batch size and the number of iterations.
[0149] For step S300
[0150] S300, inputting the test data set into the trained PM2.5 concentration point prediction model to obtain the point prediction result of the output variable.
[0151] For step S400
[0152] S400, constructing a PM2.5 concentration prediction hybrid model based on Gaussian process regression and deep learning, inputting the point prediction result of the output variable into the trained PM2.5 concentration prediction hybrid model to obtain the point prediction result of the output variable and the corresponding probability distribution function, and the prediction interval corresponding to the point prediction result of the output variable.
[0153] The step S400 comprises the following steps:
[0154] S410, taking the input variables in the previous training set and test set as the input variables in the training set and test set, taking the point prediction result of the output variable, and constructing the output variable of the next time as the training set and test set of the second time;
[0155] S420, inputting the second training set and test set into the PM2.5 concentration prediction hybrid model to obtain the point prediction result of the output variable and the corresponding probability distribution function, and the prediction interval corresponding to the point prediction result of the output variable;
[0156] S430, determining the prediction interval of each point prediction result under the preset confidence based on the mean, standard deviation and preset confidence of the probability distribution function.
[0157] Wherein, the name explanation in the calculation of point prediction index:
[0158] (1) Root Mean Square Error (RMSE): used to calculate the square root of the square sum of the deviation between the predicted value and the observed value and the number of observations, the larger the RMSE, the larger the prediction error.
[0159] (2) Mean Absolute Percentage Error (MAPE): used to calculate the average absolute error of the predicted value and the observed value, the smaller the MAPE, the more perfect the prediction model.
[0160] (3) Correlation coefficient R: used to measure the linear correlation degree of the predicted value and the true value.
[0161] (4) Determination coefficient R 2 : used to measure the deviation degree of the predicted value and the true value, between 0-1, the closer to 1 indicates that the predicted value and the true value are more consistent.
[0162] The indicators of point prediction are calculated according to the following formula:
[0163]
[0164]
[0165]
[0166]
[0167] Where, y i is the i-th observation value, Var is the mean value of the i-th observation value, Y i is the i-th predicted value, is the variance, and N is the number of prediction samples.
[0168] Wherein, in the calculation of interval prediction:
[0169] (1) Interval coverage rate CP: used to calculate the percentage of the predicted interval covering the observed value, the closer to 1, the more observed values covered by the interval;
[0170] (2) Interval average width MAW: used to calculate the average width of the predicted interval, the smaller the MAW, the higher the reliability of the interval prediction;
[0171] (3) Comprehensive index MC of interval prediction: the index of MAW and CP, the smaller the MC value, the better the interval prediction effect.
[0172] The indicators of interval prediction are calculated according to the following formula:
[0173]
[0174]
[0175]
[0176]
[0177] wherein, is the upper limit of the prediction interval of the i-th point prediction value, is the lower limit of the prediction interval of the i-th point prediction value, and a is the reliability.
[0178] The evaluation indexes of the calculated point prediction and interval prediction are compared with those of the CNN-GPR model, the LSTM-GPR model and the GPR model, and a better model is obtained, and the Figures 5 to 8 .
[0179] Specifically, referring to Figure 9 the point prediction results in the table, in the point prediction, for the monitoring stations 1 and 2, the prediction accuracies of the above four models are good, and the GPR model performs worst, which indicates that the deep learning model performs excellently in the point prediction and obtains higher prediction accuracy. The four indexes of the CLSTM-GPR model are better than those of the other models, which indicates that the point prediction result obtained by the method of the present application has the highest accuracy.
[0180] Referring to Figure 10 the interval prediction results in the table, in the interval prediction, for the prediction of the monitoring stations 1 and 2, the comparison of the above four models presents the same trend: all have high CP, and the GPR model is slightly better; for the MWP, the prediction interval of the GPR model is the largest; but for the comprehensive index MC, the CLSTM-GPR is the smallest, which indicates that the interval prediction result obtained by the method of the present application has the best comprehensive performance.
[0181] Figure 2 is a CLSTM model structure schematic diagram. The input gate is an input layer, the forget gate is a forgetting layer, the output gate is an output layer, the C t-1 , h t-1 , x t is input, O t , C t , h t is output, Figure 2 has been updated figure; Figure 3 is an implementation schematic diagram of division and prediction of a data set. dataset1 is a data set 1, dataset2 is a data set 2, peridod is time, traindata is a training data set, and test data is a test data set; Figure 4 is a CLSTM and GPR hybrid model flowchart. test set is a test set; train set is a training set; deep learning model training is deep learning model training; GPR model training is GPR model training; interval parameters are interval parameters;Figure 5 is a point prediction result map a. Figure 6 is a point prediction result map b. Figure 7 is an interval prediction result map a. Figure 8 is an interval prediction result map b. Figure 9 is a comparison table of point prediction results of four models. Figure 10 is a comparison table of interval prediction results of four models. RMSE (Root Mean Squared Error) is the mean square error; MAE (Mean Absolute Error) is the mean absolute error; the correlation coefficient R is used to measure the strength of the correlation between two variables; the determination coefficient R 2 After linear regression of the model, the regression model coefficient fitting degree is evaluated. CP (coverage probability) is defined as the probability that the observed value falls within the prediction interval; MWP (mean width percentage) is defined as the average percentage of interval width to observation; MC is an index defined by MWP / CP. Figure 5 , 6 , 7 and 8 are CNN-GPR, LSTM is LSTM-GPR, and CLSTM is CLSTM-GPR. Figure 5 , 6 , 9 can obtain that the CLSTM-GPR model has good point prediction results for PM2.5 concentrations of the No. 1 and No. 2 monitoring stations. Mainly looking at Figure 9 point prediction indicators, Figure 5 , 6 are only auxiliary. Figure 7 , 8 , 10 can be obtained that the CLSTM-GPR model has good interval prediction results for PM2.5 concentrations of the No. 1 and No. 2 monitoring stations. Mainly looking at Figure 10 interval prediction indicators.
[0182] The convolutional neural network and the long short-term memory network in the application can be used to fully extract information of process data and improve the ability of the model to process complex nonlinear data. The Gaussian process regression is used to collect data and generate accurate interval prediction. The PM2.5 concentration prediction method based on Gaussian process regression and deep learning can realize accurate prediction of PM2.5 concentration by integrating spatial information extracted by the convolutional neural network and time characteristics extracted by the long short-term memory network, and using the Gaussian regression process model for interval prediction. The above method predicts the PM2.5 concentration at future time according to the change of daily PM2.5 concentration, and simultaneously gives point prediction results and corresponding interval prediction results, and the output results have good credibility.
[0183] In one embodiment, in step 1, the PM2.5 concentration history data is acquired, including:
[0184] A preset PM2.5 monitoring station distribution map corresponding to the target area is acquired;
[0185] The PM2.5 monitoring station distribution map is preprocessed to obtain a credible station distribution map;
[0186] Based on the credible station distribution map, a monitoring missing point is determined;
[0187] Based on the monitoring missing point and a preset regional map corresponding to the target area, a mobile monitoring route is planned;
[0188] Based on the mobile monitoring route, the mobile monitoring trolley is controlled to perform PM2.5 monitoring on the monitoring missing point to acquire first historical monitoring data;
[0189] Second historical monitoring data of PM2.5 monitoring of each credible station in the credible station distribution map is acquired;
[0190] The first historical monitoring data and the second historical monitoring data are integrated to obtain PM2.5 concentration history data.
[0191] The working principle and beneficial effects of the above technical solution are as follows:
[0192] The target area is an area in which PM2.5 concentration prediction needs to be performed, for example, Tianhe District of Guangzhou City, etc. The preset PM2.5 monitoring station distribution map corresponding to the target area is a distribution diagram of PM2.5 monitoring stations in the target area, and the PM2.5 monitoring station is a PM2.5 monitor, etc. However, the distribution of PM2.5 monitoring stations may not be uniform, and it is necessary to determine the missing point of monitoring, control the mobile monitoring car to monitor the missing point of monitoring, and the mobile monitoring car is equipped with a PM2.5 monitor, etc. The preset area map corresponding to the target area is a map marked with buildings and roads, etc. In addition, the PM2.5 monitoring station may not be accurate due to damage, old age, and lack of timely maintenance, etc., resulting in untrustworthy monitoring data, so it is necessary to screen out the trusted station. The first historical monitoring data of the trusted station and the second historical monitoring data of the mobile monitoring car are integrated to obtain the PM2.5 concentration historical data. Secondly, the trusted station and the mobile monitoring car will perform PM2.5 concentration real-time monitoring and PM2.5 concentration prediction when performing PM2.5 monitoring, and then the input variable and the output variable are obtained. The cooperation of the trusted station and the mobile monitoring car can uniformly monitor the PM2.5 concentration in the target area, and improve the comprehensiveness and accuracy of the acquisition of the PM2.5 concentration historical data. The PM2.5 concentration historical data is input into the neural network model for training until convergence to obtain an artificial intelligence model that can replace manual PM2.5 concentration prediction, and the PM2.5 concentration monitoring data in the recent time period is input into the artificial intelligence model, and then the PM2.5 concentration prediction can be performed.
[0193] In one embodiment, the PM2.5 monitoring station distribution map is preprocessed, including:
[0194] The PM2.5 monitoring station distribution map is traversed in sequence;
[0195] At each traversal, based on a preset information acquisition template, the station information of the traversed PM2.5 monitoring station is acquired;
[0196] Based on a preset first feature extraction template, the station information is feature-extracted to obtain a plurality of first features;
[0197] The first feature is matched with a second feature in a preset indication feature library, and if the matching is consistent, the preset second feature extraction template and feature requirement corresponding to the matching consistent second feature are acquired;
[0198] Based on the second feature extraction template, the station information is feature-extracted to obtain a plurality of third features;
[0199] determining whether the third feature meets the feature requirement, and if not, eliminating the PM2.5 monitoring station from the PM2.5 monitoring station distribution map;
[0200] After all the PM2.5 monitoring stations that need to be eliminated are eliminated, the PM2.5 monitoring station distribution map is taken as a trusted station distribution map.
[0201] The working principle and beneficial effects of the above technical solution are:
[0202] By comparing the preset information acquisition template, the station information that can be used to verify whether the PM2.5 monitoring station monitoring data is reliable is obtained, including device model information, use time information, and maintenance record information. By comparing the preset first feature extraction template, the first feature of the station information that can reflect which angle to verify whether the PM2.5 monitoring station monitoring data is reliable is extracted, including device model, use time, and maintenance record. The second feature in the preset indication feature library is a feature that identifies which angle to verify whether the monitoring data is reliable, for example: the second feature is a maintenance record, which is verified from the maintenance record. The first feature and the second feature are matched, if the match is correct, the verification angle is determined, and the preset second feature extraction template and feature requirement corresponding to the matched second feature are obtained, the second feature extraction template and the feature requirement are verification tools for verification from the determined verification angle, for example: the verification angle is from the maintenance record, the second feature extraction template is to extract the historical maintenance frequency and the time interval between the last maintenance time and the current time from the maintenance record information, and the feature requirement is that the historical maintenance frequency is ≥3 times / month, and the time interval between the last maintenance time and the current time is ≤12 days. Based on the verification tool, the PM2.5 monitoring station is eliminated, and after all the elimination, a trusted station distribution map is obtained. The introduction of the preset indication feature library quickly determines which angle to verify whether the PM2.5 monitoring station monitoring data is reliable, reduces the verification resources, and improves the verification efficiency. In addition, the introduction of the verification tool, i.e. the second feature extraction template and the feature requirement, quickly verifies the credibility of the PM2.5 monitoring station, and if necessary, eliminates it, thereby improving the efficiency of preprocessing.
[0203] In one embodiment, based on the trusted station distribution map, a monitoring missing point is determined, including:
[0204] In the trusted station distribution map, each trusted station is taken as the center of a circle, and a preset radius length is taken as the radius to form a circular range.
[0205] Based on a preset grid division rule, the trusted station distribution map is divided into a plurality of grid areas.
[0206] sequentially traversing each of the grid areas;
[0207] extracting, each time of traversal, a remaining area in the traversed grid area except the circular range;
[0208] performing feature extraction on the remaining area based on a preset third feature extraction template to obtain a plurality of fourth features;
[0209] aggregating the fourth features to obtain a fourth feature set;
[0210] matching the fourth feature set with a preset index feature set to obtain a matching degree;
[0211] if the matching degree is greater than or equal to a preset matching degree threshold, setting a monitoring missing point in the remaining area based on a monitoring missing point setting requirement;
[0212] taking the set monitoring missing point as a new trusted site, and taking the new trusted site as the center of a new circular range with the radius length as the radius;
[0213] continuing to traverse the grid areas;
[0214] The monitoring missing point setting requirement comprises:
[0215] a first straight line distance between the set monitoring missing point and the center of any of the circular ranges around the set monitoring missing point is greater than or equal to a preset first straight line distance threshold, a minimum straight line distance between the set monitoring missing point and any of the circular ranges around the set monitoring missing point is greater than or equal to a preset second straight line distance threshold, and a third straight line distance between any two of the set monitoring missing points is greater than or equal to a preset third straight line distance threshold.
[0216] The working principle and beneficial effects of the above technical solution are:
[0217] The preset grid division rule is to divide the trusted site distribution map into mxn (m rows and n columns) grid areas, and the row spacing and column spacing are the same. The preset radius length is, for example, 300 meters. The remaining area is an area outside the trusted site monitoring coverage. The fourth feature of the remaining area that can reflect whether a monitoring missing point needs to be set in the remaining area is extracted by comparing the preset third feature extraction template, including the total area of the remaining area and the straight line distance between the area centroid and any point on the area boundary line. The preset index feature set is composed of features that can indicate that a monitoring missing point needs to be set in the remaining area, such as the total area of the remaining area ≥ 0.3 square kilometers (reflecting that the area unable to monitor PM2.5 is larger), and the number of distances between the area centroid and any point on the area boundary line ≥ 120 meters accounts for ≥ 70% (reflecting that the area unable to monitor PM2.5 is more concentrated). The fourth feature set and the index feature set are matched to obtain a matching degree. If the matching degree is greater than or equal to the preset matching degree threshold, a monitoring missing point is set in the remaining area based on the monitoring missing point setting requirement. The set monitoring missing point is used as a new trusted site, and a new circular range is made with the new trusted site as the center and the radius length as the radius. The purpose is to avoid repeated setting when setting the monitoring missing point in the future. The monitoring missing point setting requirement is to make the set monitoring missing point form a new circular range as much as possible to minimize the overlap with the original circular range, thereby improving the monitoring efficiency and utilization. The introduction of the preset third feature extraction template and the preset index feature set quickly determines whether the monitoring missing point needs to be set in the remaining area, thereby improving the working efficiency of the system. In addition, the introduction of the monitoring missing point setting requirement improves the rationality of setting the monitoring missing point in the remaining area.
[0218] In one embodiment, based on the monitoring missing point and the preset area map corresponding to the target area, a mobile monitoring route is planned, including:
[0219] Determining a map position in the area map corresponding to the monitoring missing point;
[0220] Determining a target map position with the smallest fourth straight line distance between the map position and a preset monitoring trolley initial position;
[0221] Taking the target map position as the starting point of the route, a shortest route passing through the remaining map positions is planned and used as the mobile monitoring route.
[0222] The working principle and beneficial effects of the above technical solution are:
[0223] The preset monitoring trolley initial position is a monitoring trolley departure position, for example, a monitoring trolley warehouse, etc. The route starting point is selected with the minimum fourth straight line distance, the mobile monitoring route is planned, too much time is not spent when the monitoring trolley starts to go, and the monitoring scheduling efficiency is improved. In addition, the monitoring trolley can adaptively identify whether the monitoring missing point position is arrived, if not, the PM2.5 monitoring data of the monitoring missing point position is monitored as the PM2.5 concentration monitoring data of the nearest reachable position of the monitoring missing point position.
[0224] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application belong to the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and variations.
Claims
1. A PM2.5 concentration prediction method based on Gaussian process regression and deep learning, characterized in that, The method comprises the following steps: Step 1: obtaining PM2.5 concentration historical data; the PM2.5 concentration historical data comprises input variables which are true values of PM2.5 concentration and output variables which are predicted values of PM2.5 concentration; Step 2: dividing the PM2.5 concentration historical data into a training data set and a test data set according to a preset proportion, and preprocessing the training data set and the test data set; Step 3: constructing a PM2.5 concentration point prediction model combining a convolutional neural network and a long short-term memory network, inputting the preprocessed training data set into the PM2.5 concentration point prediction model for multiple times of training, obtaining ideal parameters, and configuring the PM2.5 concentration point prediction model based on the ideal parameters; Step 4: inputting the preprocessed test data set into the configured PM2.5 concentration point prediction model to obtain point prediction results of the output variables; Step 5: constructing a PM2.5 concentration prediction hybrid model combining a Gaussian process regression and deep learning, inputting the point prediction results into the PM2.5 concentration prediction hybrid model to obtain a probability distribution function and a prediction interval corresponding to the point prediction results; The step 5 further comprises: calculating each index of the point prediction and the interval prediction, the index of the point prediction is calculated according to the following formula: wherein y i is the ith observation, Var is the mean of the ith observation, Y i is the ith predicted value of the PM2.5 concentration prediction hybrid model, the CNN-GPR model, the LSTM-GPR model, and the GPR model, is the variance, N is the number of prediction samples, E and σ are operators, MAE is the mean absolute error, RMSE is the mean square root error, and R is the correlation coefficient. the index of the interval prediction is calculated according to the following formula: wherein U i (α) is the upper limit of the prediction interval of the i-th point prediction value, L i (α) is the lower limit of the prediction interval of the i-th point prediction value, and a is the confidence level. 2.The PM2.5 concentration prediction method based on Gaussian process regression and deep learning of claim 1, wherein, In the step 2, the preprocessing of the training data set and the test data set comprises: screening and removing abnormal values in the training data set and the test data set; performing normalization processing on the screening and removing results, and the normalization processing formula is as follows: y * = (y - min) / (max - min) wherein y is any original data value in the screening and removing results, min is the minimum original data value in the screening and removing results, max is the maximum original data value in the screening and removing results, and y* is the normalized data value corresponding to any original data value in the screening and removing results. 3.The PM2.5 concentration prediction method based on Gaussian process regression and deep learning of claim 1, wherein, In the step 3, inputting the preprocessed training data set into the PM2.5 concentration point prediction model for multiple times of training to obtain ideal parameters comprises: setting neural network parameters of the PM2.5 concentration point prediction model; inputting the training data set into the PM2.5 concentration point prediction model for multiple times of training to obtain accuracy and corresponding optimization parameters of each training; taking the optimization parameters corresponding to the maximum accuracy as the ideal parameters; wherein setting the neural network parameters of the PM2.5 concentration point prediction model comprises: setting a CNN layer initialization function of the PM2.5 concentration point prediction model as Kaiming; setting LSTM, CLSTM and GRU layer initialization functions of the PM2.5 concentration point prediction model as Orthogonal; setting an optimizer of the PM2.5 concentration point prediction model as Adam; setting a learning rate of the PM2.5 concentration point prediction model as 1e-3; setting a loss function of the PM2.5 concentration point prediction model as MSE; Set the batch size of the PM2.5 concentration point prediction model to 20; Set the dropout rate of the PM2.5 concentration point prediction model to 0.
2. 4.The PM2.5 concentration prediction method based on Gaussian process regression and deep learning of claim 1, wherein, In step 1, the PM2.5 concentration historical data is obtained, including: Obtain the preset PM2.5 monitoring station distribution map corresponding to the target area; Preprocess the PM2.5 monitoring station distribution map to obtain a trusted station distribution map; Based on the trusted station distribution map, determine the monitoring missing point position; Based on the monitoring missing point position and the preset regional map corresponding to the target area, plan a mobile monitoring route; Based on the mobile monitoring route, control the mobile monitoring car to monitor the PM2.5 of the monitoring missing point position, and obtain the first historical monitoring data; Obtain the second historical monitoring data of each trusted station in the trusted station distribution map for PM2.5 monitoring; Integrate the first historical monitoring data and the second historical monitoring data to obtain the PM2.5 concentration historical data. 5.The PM2.5 concentration prediction method based on Gaussian process regression and deep learning of claim 4, wherein, The preprocessing of the PM2.5 monitoring station distribution map includes: Iterate through each PM2.5 monitoring station in the PM2.5 monitoring station distribution map in turn; At each iteration, based on a preset information acquisition template, obtain the station information of the PM2.5 monitoring station being iterated through; Based on a preset first feature extraction template, perform feature extraction on the station information to obtain a plurality of first features; Match the first features with second features in a preset indication feature library. If the match is consistent, obtain the preset second feature extraction template and feature requirement corresponding to the matching consistent second features; Based on the second feature extraction template, perform feature extraction on the station information to obtain a plurality of third features; Determine whether the third features meet the feature requirement. If not, remove the PM2.5 monitoring station being iterated through from the PM2.5 monitoring station distribution map; After all the PM2.5 monitoring stations that need to be removed are removed, the PM2.5 monitoring station distribution map is taken as a trusted station distribution map. 6.The PM2.5 concentration prediction method based on Gaussian process regression and deep learning of claim 4, wherein, Based on the trusted station distribution map, determine the monitoring missing point position, including: In the trusted station distribution map, take each trusted station as the center, and a preset radius length as the radius to draw a circular range; Based on a preset grid division rule, divide the trusted station distribution map into a plurality of grid regions; Iterate through each grid region in turn; At each iteration, extract the remaining area in the grid region being iterated through, excluding the circular range; Based on a preset third feature extraction template, perform feature extraction on the remaining area to obtain a plurality of fourth features; Aggregate the fourth features to obtain a fourth feature set; Match the fourth feature set with a preset index feature set to obtain a matching degree; If the matching degree is greater than or equal to a preset matching degree threshold, set a monitoring missing point position in the remaining area based on monitoring missing point position setting requirements; Take the set monitoring missing point position as a new trusted station, take the new trusted station as the center, and the radius length as the radius to draw a new circular range; Continue to iterate through the grid regions; The monitoring missing point position setting requirement comprises: The first straight line distance between the set monitoring missing point position and the center of the circular range around any of the set monitoring missing point positions is greater than or equal to a preset first straight line distance threshold, the minimum straight line distance between the set monitoring missing point position and the center of the circular range around any of the set monitoring missing point positions is greater than or equal to a preset second straight line distance threshold, and the third straight line distance between any two of the set monitoring missing point positions is greater than or equal to a preset third straight line distance threshold.
7. The PM2.5 concentration prediction method based on Gaussian process regression and deep learning of claim 4, wherein, Based on the monitoring missing point position and the preset area map corresponding to the target area, a mobile monitoring route is planned, comprising: Determining a map position corresponding to the monitoring missing point position in the area map; Determining a target map position with the minimum fourth straight line distance between the map position and a preset monitoring trolley initial position; Taking the target map position as a route starting point, a shortest route passing through the remaining map positions is planned and serves as a mobile monitoring route. 8.A PM2.5 concentration prediction system based on Gaussian process regression and deep learning, characterized in that, Comprise: An acquisition module is configured to acquire PM2.5 concentration historical data, wherein the PM2.5 concentration historical data comprises input variables of PM2.5 concentration true values and output variables of PM2.5 concentration predicted values; A division module is configured to divide the PM2.5 concentration historical data into a training data set and a test data set according to a preset proportion, and to pre-process the training data set and the test data set; A first construction module is configured to construct a PM2.5 concentration point prediction model that integrates a convolutional neural network and a long short-term memory network, to input the pre-processed training data set into the PM2.5 concentration point prediction model for multiple times of training, to obtain ideal parameters, and to configure the PM2.5 concentration point prediction model based on the ideal parameters; An input module is configured to input the pre-processed test data set into the configured PM2.5 concentration point prediction model to obtain point prediction results of the output variables; A second construction module is configured to construct a PM2.5 concentration prediction hybrid model that integrates a Gaussian process regression and deep learning, to input the point prediction results into the PM2.5 concentration prediction hybrid model to obtain a probability distribution function and a prediction interval corresponding to the point prediction results; The second construction module further comprises: The indicators of the point prediction and the interval prediction calculated, The indicator of the point prediction is calculated according to the following formula: wherein y i is the ith observation, Var is the mean of the ith observation, Y i is the ith predicted value of the PM2.5 concentration prediction hybrid model, the CNN-GPR model, the LSTM-GPR model, and the GPR model, is the variance, N is the number of prediction samples, E and σ are operators, MAE is the mean absolute error, RMSE is the mean square root error, and R is the correlation coefficient. The indicator of the interval prediction is calculated according to the following formula: wherein U i (α) is the upper limit of the prediction interval of the i-th point prediction value, L i (α) is the lower limit of the prediction interval of the i-th point prediction value, and a is the reliability.
9. A computer-readable storage medium having stored thereon a computer program, characterized in that The program is executed by a processor to implement a PM2.5 concentration prediction method based on Gaussian process regression and deep learning according to any one of claims 1 to 7.