Gas pollutant distribution prediction method, device and equipment and storage medium

Abstract

The invention discloses a gas pollutant distribution prediction method, device and equipment and a storage medium, a data sample enhancement effect is realized by adopting two modes of data segmentation and resampling, the effect of a model can be ensured under the condition of small sample data volume, and the problem of insufficient data acquisition quantity in the field of gas pollutants is effectively relieved. Meanwhile, the model trained through the method is also effective for data collected by discrete detectors, gas pollutant situation distribution in the whole area can be predicted under the condition of limited detector data, and the larger the detector data granularity is, the more accurate the prediction result is.

Description

Technical Field

[0001] The present invention relates to the technical field of pollutant distribution prediction, and in particular to a gas pollutant distribution prediction method, device, equipment and storage medium based on a ResNet-LSTM prediction model. Background Art

[0002] Air pollutant concentration prediction plays a key role in environmental monitoring and urban management. The technical background behind it is based on the theory of atmospheric pollution science, which collects data through environmental monitoring technology and combines data science and statistical methods to analyze the diffusion pattern of pollutants. Usually, simulation models and data models are used to help simulate the propagation and chemical reaction process of pollutants in the atmosphere. With the development of machine learning and artificial intelligence technology, feature engineering and model training have become key links to improve prediction accuracy.

[0003] Commonly used prediction methods include statistical models, simulations, and machine learning models. Statistical models are based on historical data and regression analysis. Their advantages are simple implementation and fast calculation, but they may not accurately reflect complex nonlinear relationships and are often used to predict single-point trends. Simulations, such as CFD atmospheric diffusion models, simulate physical and chemical processes based on mechanism models and are suitable for detailed research, but they have high computational costs, are sensitive to input parameters, have a high degree of specialization, and are difficult to apply in practice. They are suitable for scientific research. Machine learning technology automatically identifies data patterns through algorithms, flexibly adapts to changing conditions, and does not rely too much on expert knowledge, but may require a large amount of training data to avoid overfitting. It is currently a popular research field. Each method has its own advantages and disadvantages, and is often used in combination to improve prediction accuracy.

[0004] The technical solution of neural network in the field of air pollutant concentration prediction mainly involves using a large amount of environmental monitoring data, such as meteorological conditions and historical air quality data, improving data quality through data preprocessing, and using appropriate neural network models such as back propagation (BP) network or recurrent neural network (RNN) for modeling and training to identify the complex nonlinear relationship between pollutant concentration and environmental factors. During the training process, cross-validation and parameter tuning are used to avoid overfitting and improve the generalization ability of the model, ultimately achieving accurate short-term or long-term prediction of air quality, providing a scientific basis for environmental monitoring and management decisions.

[0005] However, the main disadvantages of neural network models in the field of air pollutant concentration prediction include reliance on large amounts of high-quality training data, which can significantly affect model performance when data is scarce or inaccurate. At the same time, the "black box" nature of neural networks makes the model lack transparency and makes it difficult to explain the specific basis of its prediction results. In addition, complex neural network models require high computing resources and may overfit, resulting in poor performance of the model on independent data sets, reducing the generalization ability and practicality of the model. This shortcoming can be solved by expanding the training data.

[0006] The technical solutions of time series models in the field of air pollutant concentration prediction usually involve the use of historical air quality data and related environmental factor data to build models. These models include autoregressive moving average (ARMA), autoregressive integrated moving average (ARIMA), seasonal autoregressive integrated moving average (SARIMA), long short-term memory (LSTM), etc., which can capture trends, periodicity and random fluctuations in time series data. Through appropriate model identification, parameter estimation, model testing and diagnosis, short-term or long-term concentration changes of air pollutants can be accurately predicted. In addition, time series models that introduce external covariates such as ARIMAX or vector autoregression (VAR) models can further improve the accuracy of predictions and provide a scientific basis for air quality management.

[0007] However, the spatial prediction performance of time series models is often poor, and it is difficult to predict the spatial distribution characteristics of air pollutants within a certain range. In general, time series models are suitable for processing trend predictions of single objects and are not often used for two-dimensional and above data. To address this shortcoming, it is necessary to enhance the spatial representation of data in the model and strengthen the spatial characteristics. Summary of the invention

[0008] In view of the above problems, the present invention provides a gas pollutant distribution prediction method, device, equipment and storage medium for overcoming the above problems or at least partially solving the above problems.

[0009] The present invention provides the following scheme:

[0010] A method for predicting gas pollutant distribution, comprising:

[0011] Use detectors to obtain gas pollutant concentration data and meteorological element data at target monitoring points;

[0012] Inputting the gas pollutant concentration data and meteorological element data into the trained target distribution prediction model to perform distribution prediction to obtain a gas pollutant distribution prediction result;

[0013] The target distribution prediction model includes a multi-layer residual network and a long short-term memory network; the multi-layer residual network is used to extract spatial information from the data, and the long short-term memory network is used to extract temporal information from the data, so that the target distribution prediction model can predict the distribution of gas pollutants based on the spatial information and the temporal information;

[0014] Among them, during the training process of the target distribution prediction model, the point diffusion and linear interpolation methods are used to expand the sample data to form rectangular grid data, and the sample enhancement processing method of data resampling and data segmentation is used to enhance the sample feature expression ability of the rectangular grid data. The structural coefficient is used as the model training indicator, the recursive feature elimination method is used for feature selection, and the autocorrelation function is used to realize the time series order selection.

[0015] Preferably: before the rectangular grid data is subjected to sample feature expression capability enhancement processing by adopting the sample enhancement processing method of data resampling and segmentation, the rectangular grid data is subjected to outlier processing and normalization processing.

[0016] Preferably: the outlier processing includes filling missing values ​​with the mean filling method, eliminating outliers in the sample data based on the 3σ principle, and calculating the value range of the data after outliers are eliminated; the normalization processing includes normalizing all maximum values ​​by 9-11% on the basis of the original maximum value of the value range.

[0017] Preferably: the data resampling includes randomly resampling the data points of each group of samples, performing multiple resampling operations on each group of samples, and randomly selecting 60% to 80% of the points for zero assignment each time; the data segmentation includes equally sized segmentation of the rectangular grid data into multiple small-sized data.

[0018] Preferably: the feature selection includes:

[0019] Determine initial model parameters;

[0020] Arrange and combine all features and enumerate all possible combinations of features;

[0021] Calculate the evaluation index of each feature combination and sort it;

[0022] Select one or more feature combinations with the highest ranking, and replace the data for re-verification;

[0023] If the feature combination is verified to be stable and good after multiple times, the iteration is terminated; otherwise, the model parameters are modified and the above steps are repeated until the final model parameters and feature combination are determined.

[0024] Preferably: the autocorrelation function includes the time series {X t}Autocorrelation coefficient of lag order kρ k Defined as:

[0025]

[0026] Where: X t is the value of the sequence at t, is the mean of the sequence, T is the total length of the sequence, k is the lag order, ρ k The value range is -1 to 1, 1 means perfect positive correlation, -1 means perfect negative correlation, and 0 means no correlation.

[0027] Preferably: the target distribution prediction model further includes a Batch Normalization layer and an activation function; the Batch Normalization layer and the activation function are used to maintain the stability of the model distribution.

[0028] A gas pollutant distribution prediction device, used to execute the above-mentioned gas pollutant distribution prediction method, the device comprising:

[0029] A data acquisition unit, used to acquire gas pollutant concentration data and meteorological element data at a target monitoring point using a detector;

[0030] A prediction result acquisition unit, used for inputting the gas pollutant concentration data and the meteorological element data into the trained target distribution prediction model to perform distribution prediction, so as to obtain the gas pollutant distribution prediction result;

[0031] The target distribution prediction model includes a multi-layer residual network and a long short-term memory network; the multi-layer residual network is used to extract spatial information from the data, and the long short-term memory network is used to extract temporal information from the data, so that the target distribution prediction model can predict the distribution of gas pollutants based on the spatial information and the temporal information;

[0032] Among them, during the training process of the target distribution prediction model, the point diffusion and linear interpolation methods are used to expand the sample data to form rectangular grid data, and the sample enhancement processing method of data resampling and data segmentation is used to enhance the sample feature expression ability of the rectangular grid data. The structural coefficient is used as the model training indicator, the recursive feature elimination method is used for feature selection, and the autocorrelation function is used to realize the time series order selection.

[0033] A gas pollutant distribution prediction device, the device comprising a processor and a memory:

[0034] The memory is used to store program code and transmit the program code to the processor;

[0035] The processor is used to execute the above-mentioned gas pollutant distribution prediction method according to the instructions in the program code.

[0036] A computer-readable storage medium is used to store program codes, and the program codes are used to execute the above-mentioned gas pollutant distribution prediction method.

[0037] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0038] A gas pollutant distribution prediction method, device, equipment and storage medium provided in the embodiments of the present application adopt data segmentation and resampling to achieve data sample enhancement effect, which can ensure the effect of the model under the condition of small sample data volume and effectively alleviate the problem of insufficient data collection in the field of gas pollutants; at the same time, the model trained by this method is also effective for data collected by discrete detectors, and can predict the distribution of gas pollutants in the entire area with limited detector data. The larger the granularity of the detector data, the more accurate the prediction result.

[0039] At the same time, the ResNet-LSTM model, which integrates the residual network and the long short-term memory network, is used for prediction. This model is optimized and innovated based on the original ResNet and LSTM network structure, so that the model has both the spatial advantages of ResNet and the temporal advantages of LSTM. The structural coefficient (SSIM) with strong spatial properties is used as the model training indicator, which greatly improves the effect of gas pollutant concentration prediction.

[0040] In addition, a complete set of process-based modeling procedures is provided. Major steps such as data sample enhancement, feature selection, time series order selection, outlier processing, and normalization processing can all be selected in a standard manner. Following this process, automated modeling capabilities with little human involvement can be achieved, which can be extended to other similar fields such as gas pollution and even water pollution.

[0041] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0043] Figure 1 is a flow chart of a method for predicting gas pollutant distribution provided by an embodiment of the present invention;

[0044] Figure 2 is a schematic diagram of data resampling provided by an embodiment of the present invention;

[0045] Figure 3 is a schematic diagram of data segmentation provided by an embodiment of the present invention;

[0046] Figure 4 is a schematic diagram of the residual network structure provided by an embodiment of the present invention;

[0047] Figure 5 Schematic diagram of the LSTM network structure provided by an embodiment of the present invention;

[0048] Figure 6 is a schematic diagram of the network structure of the target distribution prediction model provided by an embodiment of the present invention;

[0049] Figure 7 Schematic diagram of distribution of autocorrelation coefficients of various orders provided by an embodiment of the present invention;

[0050] Figure 8 This is a comparison diagram of the model effect of retaining 20% ​​of the data provided by an embodiment of the present invention;

[0051] Fig. 9 This is a comparison diagram of the model effect of retaining 40% of the data provided by an embodiment of the present invention;

[0052] Fig.10 is a schematic diagram of a gas pollutant distribution prediction device provided by an embodiment of the present invention;

[0053] Fig.11 It is a schematic diagram of a gas pollutant distribution prediction device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0054] The technical scheme in the embodiment of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiment of the present invention, not all of the embodiments. Based on the embodiment of the present invention, all other embodiments obtained by ordinary technicians in this field belong to the scope of protection of the present invention.

[0055] See also Figure 1 , is a method for predicting gas pollutant distribution provided by an embodiment of the present invention, such as Figure 1 As shown, the method may include:

[0056] S101: using the detector to obtain gas pollutant concentration data and meteorological element data at the target monitoring point;

[0057] S102: Inputting the gas pollutant concentration data and meteorological element data into the trained target distribution prediction model to perform distribution prediction to obtain a gas pollutant distribution prediction result;

[0058] The target distribution prediction model includes multiple layers of residual networks and long short-term memory networks; the multiple layers of the residual networks are used to extract spatial information from the data, and the long short-term memory networks are used to extract temporal information from the data, so that the target distribution prediction model can predict the distribution of gas pollutants based on the spatial information and the temporal information; further, an embodiment of the present application may also provide that the target distribution prediction model also includes a Batch Normalization layer and an activation function; the Batch Normalization layer and the activation function are used to maintain the stability of the model distribution.

[0059] Among them, during the training process of the target distribution prediction model, the point diffusion and linear interpolation methods are used to expand the sample data to form rectangular grid data, and the sample enhancement processing method of data resampling and data segmentation is used to enhance the sample feature expression ability of the rectangular grid data. The structural coefficient is used as the model training indicator, the recursive feature elimination method is used for feature selection, and the autocorrelation function is used to realize the time series order selection.

[0060] The gas pollutant distribution prediction method provided in the embodiment of the present application solves the problem of insufficient sample data. The source of gas pollutant data mainly comes from meteorological and environmental monitoring data, but the leakage of harmful gases is a low-probability event with extremely strong uncertainty, and it is difficult to obtain effective monitoring data. The source of meteorological data is mainly meteorological monitoring stations. There are 127 national meteorological stations in my country, distributed in various provinces, covering a wide range of areas from cities to villages, and the locations are unevenly distributed. For regional model training, the amount of data is far from enough. In addition, in addition to national meteorological stations, different organizations in different regions will also collect meteorological data in specific areas from time to time, which makes up for the problem of data quantity to a certain extent. This method uses different resampling degrees of refined grid data to simulate different data granularity environments to establish a prediction model. After the model is established, it has strong versatility and can be applied to data prediction of different granularities. The higher the data granularity, the more accurate the model prediction.

[0061] Traditional prediction methods are too complex to be applied in practice. Usually, simulation and data models are used to establish prediction models for gaseous pollutants. Simulation models require highly professional CFD simulation software, and the calculation time is long. The calculation results are less flexible and difficult to apply in practice. They are often used in scientific research. Data models often require many features such as meteorological elements to assist in modeling. In previous studies, a lot of filtering, feature screening and other work are often required to ensure the rationality of the model input features. The models obtained in this way are often more complex, and each pollutant needs to be modeled separately. Even the data given for different regions and different monitoring equipment need to be re-modeled, and the practical application ability is poor.

[0062] The method provided in the embodiment of the present application uses meteorological data as characteristic parameters to predict the distribution of gaseous pollutants. The method is based on the ResNet (residual network) and LSTM (long short-term memory network) network structures, and after combined optimization, an innovative model structure is formed. The model has both the spatial performance of ResNet and the temporal performance of LSTM, and uses the structural coefficient (SSIM) with strong spatial characteristics as an indicator, which has achieved good results in predicting the diffusion of gaseous pollutants.

[0063] The method provided in the embodiment of the present application processes the sample data through data sample enhancement, feature selection, time series order selection, etc., which can reduce the demand for data collection and enhance the feature expression ability of data in small sample data.

[0064] In a specific implementation, the embodiment of the present application may provide a sample enhancement processing method using data resampling and segmentation to perform outlier processing and normalization processing on the rectangular grid data before performing sample feature expression capability enhancement processing on the rectangular grid data.

[0065] Furthermore, the outlier processing includes filling missing values ​​with the mean filling method, eliminating outliers of sample data based on the 3σ principle, and calculating the value range of the data after outliers are eliminated; the normalization processing includes normalizing all maximum values ​​by 9 to 11% based on the original maximum value of the value range.

[0066] The data resampling includes randomly resampling the data points of each group of samples, performing multiple resampling operations on each group of samples, and randomly selecting 60% to 80% of the points for zero assignment each time; the data segmentation includes equally sized segmentation of the rectangular grid data into multiple small-sized data.

[0067] The feature selection includes:

[0068] Determine initial model parameters;

[0069] Arrange and combine all features and enumerate all possible combinations of features;

[0070] Calculate the evaluation index of each feature combination and sort it;

[0071] Select one or more feature combinations with the highest ranking, and replace the data for re-verification;

[0072] If the feature combination is verified to be stable and good after multiple times, the iteration is terminated; otherwise, the model parameters are modified and the above steps are repeated until the final model parameters and feature combination are determined.

[0073] The autocorrelation function includes the time series {X t}Autocorrelation coefficient of lag order kρ k Defined as:

[0074]

[0075] Where: X t is the value of the sequence at t, is the mean of the sequence, T is the total length of the sequence, k is the lag order, ρ k The value range is -1 to 1, 1 means perfect positive correlation, -1 means perfect negative correlation, and 0 means no correlation.

[0076] The embodiment of the present application provides a complete set of systematic modeling processes. Major steps such as data sample enhancement, feature selection, time series order selection, outlier processing, and normalization processing can all be selected in a standard manner. According to this system, automated modeling capabilities with little human participation can be achieved, which can be extended to other similar fields such as gas pollution and even water pollution.

[0077] The method provided in the embodiments of the present application is described in detail below.

[0078] The main model building and training process in the method provided in the embodiment of the present application is as follows:

[0079] Step 1: Data acquisition and generation.

[0080] Various detectors (sensors) are used to obtain gas pollutant concentration data and meteorological element data (wind speed, wind direction, temperature, humidity, air pressure, etc.) at N monitoring points, and the sample data are expanded into rectangular grid data using methods such as point diffusion and linear interpolation.

[0081] The purpose of this step is to generate model input data. Therefore, the more acquisition points there are, the more accurate the generated mesh data will be, and the more accurate the model input will be. The fewer acquisition points there are, the sparser the generated mesh data will be, and the less model input information will be.

[0082] The following is an example of the China High-Resolution Air Quality Reanalysis Dataset (CAQRA) (2013-2018), which provides surface grid fields of six conventional air pollutants (i.e. PM2.5, PM10, SO2, NO2, CO and O3), as well as surface wind speed (u, v), air pressure (psfc), relative humidity (RH) and temperature (temp) fields simulated by the WRF model. The spatial and temporal resolutions are 15 km and 1 hour, respectively. The time period of the dataset is from 2013 to 2019. Modeling is carried out using the above data, and the diffusion trends of the six air pollutants are predicted using wind speed (u, v), air pressure (psfc), relative humidity (RH) and temperature (temp) fields.

[0083] Step 2: data preprocessing.

[0084] It mainly includes outlier processing, normalization processing and data sample enhancement.

[0085] Outlier processing: The data uses the mean filling method to fill missing values, and the outliers of the sample data are eliminated based on the 3σ principle, and the value range of the data after outliers are eliminated is calculated;

[0086] Normalization: To ensure the stability of the model, the data is normalized. Based on redundancy considerations, it is best to normalize all maximum values ​​by about 10% based on the original maximum value of the range. Therefore, the training data used is the normalized data with a range of [0,1). Taking the CAQRA dataset as an example, the normalization thresholds are shown in Table 1 and Table 2:

[0087] Table 1 Normalized thresholds of various meteorological elements

[0088] Type of pollutant u v PSFC RH Temp 3σ processing range [-15.72,17.12] [-14.98,14.66] [55543.83,106517.43] [0,100.00] [239.82,320.22] Normalized range [-18,18] [-15,15] [50000,120000] [0,100] [220,350]

[0089] Table 2 Normalized thresholds for each pollutant

[0090] Type of pollutant PM2.5 PM10 <![CDATA[SO2]]> <![CDATA[NO2]]> CO <![CDATA[O3]]> 3σ processing range [0,99.41] [0,121.77] [0,42.41] [0,32.16] [0,1.17] [0,149.05] Normalized range [0,110] [0,130] [0,45] [0,35] [0,1.25] [0,160]

[0091] (3) Data sample enhancement: In order to reduce the demand for data collection, the feature expression ability of data in small sample data is enhanced, and sample enhancement processing such as data resampling and segmentation is performed on the data.

[0092] See also Figure 2Data resampling is to randomly resample the data points of each group of samples. Each group of samples is resampled 10 times, and 60% to 80% of the points are randomly selected each time for assignment to 0. 0 means that there is no data at this point. The purpose of expanding the sample size is achieved by significantly and randomly reducing the amount of information in each group of samples. At the same time, the model trained on the data with reduced information has stronger generalization prediction ability. At the same time, the non-grid data collected by the detector is simulated. During the simulation, the scope of the data assignment to 0 can be expanded to increase the sample size of different data granularities.

[0093] See also Figure 3 Data segmentation is to segment the original grid data into multiple small-sized data of equal size to expand the sample size of the data set. Taking the CAQRA data set as an example, the original 339×432 grid data is segmented into 12 113×108 grid data to enhance the diversity of the data and improve the feature expression ability of the model.

[0094] Step 3: Model building.

[0095] The embodiment of the present application provides a combined model of a residual network (ResNet) and a long short-term memory network (LSTM) as a target distribution prediction model, which improves the accuracy of short-term distribution prediction of air pollutants by combining the powerful ability of ResNet in extracting spatial features and the advantages of LSTM in processing time series data. The model reconstructs data with multiple feature parameters and inputs them into a multi-layer ResNet network for feature extraction, and processes the time relationship through the LSTM network to predict the short-term distribution of air pollutants.

[0096] ResNet is a deep convolutional neural network architecture that effectively solves the degradation problem in deep network training by introducing "residual learning". Thanks to its residual block design and deep network structure, it can build and train networks with hundreds or even thousands of layers, and has significant advantages in extracting spatial features. Figure 4 .

[0097] LSTM is a special type of recurrent neural network (RNN) designed to solve the long-term dependency problem in sequence data. It controls the entry, memory, forgetting, and outflow of information through gating mechanisms and memory units, captures the long-term dependency in the sequence, and has excellent performance in extracting time features. The overall structure of LSTM is as follows Figure 5 As shown, the model is a classic structure and will not be described in detail here.

[0098] The model provided in the embodiment of the present application introduces a residual block in the multi-layer network structure to fully extract the spatial information in the data, then extracts the temporal information through the LSTM network, and finally maintains the stability of the model distribution through the BatchNormalization layer and the activation function. The overall model structure is as follows Figure 6 shown.

[0099] Step 4: Indicator selection. Conventional neural network models generally use indicators such as MSE and RMSE to train models. These indicators are suitable for optimizing pixel-level accuracy, but they cannot reflect the structural nature of the overall spatial distribution. Since the concentration distribution of gas pollutants has strong spatial properties, the structural similarity index (SSIM) is used as a model training indicator. SSIM is an indicator that measures the visual similarity between two images. It is based on the human visual system's perception of image structural information and calculates similarity by comparing the brightness, contrast, and structure of the images. The value range of SSIM is between 0 and 1, where 0 means that the two images are completely dissimilar and 1 means that the two images are exactly the same. The calculation process is a conventional method, which is not the focus of this application and will not be repeated.

[0100] Step 5: Feature selection.

[0101] The diffusion of gaseous pollutants is often affected by meteorological factors. Taking PM2.5 as an example, its diffusion may be related to factors such as wind speed (u, v), air pressure (psfc), relative humidity (RH) and temperature (temp), but the specific impact is unknown. According to the prediction requirements, the pollutant concentration at the previous moment, wind speed (u, v)), air pressure (psfc), relative humidity (RH) and temperature (temp) data are needed to predict the pollutant concentration at the subsequent moment. Therefore, the features need to be selected from a total of 6 items: pollutant concentration, wind speed (u, v)), air pressure (psfc), relative humidity (RH) and temperature (temp). Since the total number of features is not large, the recursive feature elimination method is used for feature selection. The specific steps are as follows:

[0102] (1) Select a model: The model provided in the embodiment of the present application is the above-mentioned ResNet-LSTM model, and the initial model parameters are set.

[0103] (2) Initialize the feature set: Arrange and combine all features and enumerate all possible combinations of features.

[0104] (3) Feature importance evaluation and ranking: Calculate the evaluation index of each feature combination and rank them.

[0105] (4) Feature selection: Select the highest-ranked feature or a combination of multiple features, then replace the data and re-verify.

[0106] (5) Check the stopping condition: If the feature combination is verified to be stable and good after multiple verifications, terminate the iteration; otherwise, modify the model parameters and repeat the above steps until the final model parameters and feature combination are determined.

[0107] Taking the CAQRA dataset as an example, four features, including pollutant concentration, wind speed (u, v), and temperature (temp), were finally selected as model training features.

[0108] Step 6: Time series order selection. This step mainly determines the previous time order of the prediction model to extract the optimal time series features, which is mainly achieved by using the autocorrelation function (ACF). t}, the autocorrelation coefficient of its lag order k is ρ k Defined as:

[0109]

[0110] Among them, X t is the value of the sequence at t, is the mean of the sequence, T is the total duration of the sequence, k is the lag order, ρ k The value range is -1 to 1, 1 means perfect positive correlation, -1 means perfect negative correlation, and 0 means no correlation.

[0111] Using the above formula, the present invention calculates the autocorrelation coefficient of the two-dimensional time series matrix of pollutant concentration to generate a time series autocorrelation coefficient matrix ρ kij , and then determine the time correlation of the sample data by analyzing the spatial distribution characteristics of the time series autocorrelation coefficient matrix.

[0112] Taking the CAQRA dataset as an example, the relationship between the autocorrelation coefficient matrix of pollutant concentration and the time series order is as follows: Figure 7 As shown in the figure, in general, the larger the lag order, the lower the time series autocorrelation and the more difficult the model training is. The appropriate lag order can be selected according to the actual prediction time requirements.

[0113] Step 7: Model training.

[0114] After the above steps, the data and model preparation work has been completed, and the model training is directly carried out. The model training parameters are determined according to the target accuracy requirements and the environmental performance. Taking the environment 3050Ti graphics card and 16G memory provided in the embodiment of this application as an example, the model parameters with a deviation rate mean less than 0.1 and an SSIM value of about 0.6 are as shown in the following table, where the batch size can be determined according to the performance of the training environment. Within a certain range, the larger the batch, the greater the performance consumption, but the better the model effect. The model parameters are shown in Table 3.

[0115] Table 3 Model parameters

[0116] Serial number Parameter name Parameter settings 1 Optimizing functions Adam algorithm, initial learning rate 0.0001 2 Loss Function Structural Similarity Index (SSIM) 3 Activation Function Leaky ReLU 4 Number of training sessions 1000 5 Batch size 1024 6 Time step 1 7 Discard rate 0.001～0.1 8 Number of LSTM layers 1

[0117] After training, the target distribution prediction model can be used to predict the distribution of gas pollutants at the target detection point. During the prediction, various detectors (sensors) can also be used to obtain the gas pollutant concentration data and meteorological element data of the target detection point, and then the gas pollutant concentration data and meteorological element data can be input into the target distribution prediction model to obtain the gas distribution prediction result.

[0118] Step 8: Model verification.

[0119] SSIM and mean deviation rate are used to verify the model effect. To ensure the model verification effect, the longer the interval between the verification data and the training data, the better, and the correlation between the two should be avoided as much as possible. Taking the CAQRA dataset as an example, the training data is the 2019 data and the test data is the 2013 data. The mean deviation rate calculation formula is as follows:

[0120] Deviation rate = ((actual value - target value) / target value) * 100%

[0121] Deviation rate mean = Mean(deviation rate)

[0122] Taking the CAQRA dataset with a lag order of 4 and PM2.5 as an example, the comparison of model prediction results is shown below. The measured data are highly resampled, and the two sets of data retain 20% and 40% of the data respectively, which highly simulates the monitoring granularity data of different detectors. However, the model can still predict the distribution of pollutants in the entire area, and the larger the input data granularity, the more detailed the prediction. Figure 8 , Fig. 9 shown.

[0123] In summary, the gas pollutant distribution prediction method provided in this application adopts data segmentation and resampling to achieve data sample enhancement effect, which can ensure the effect of the model under the condition of small sample data volume and effectively alleviate the problem of insufficient data collection in the field of gas pollutants; at the same time, the model trained by this method is also effective for data collected by discrete detectors, and can predict the distribution of gas pollutants in the entire area with limited detector data. The larger the granularity of the detector data, the more accurate the prediction result.

[0124] At the same time, the ResNet-LSTM model, which integrates the residual network and the long short-term memory network, is used for prediction. This model is optimized and innovated based on the original ResNet and LSTM network structure, so that the model has both the spatial advantages of ResNet and the temporal advantages of LSTM. The structural coefficient (SSIM) with strong spatial properties is used as the model training indicator, which greatly improves the effect of gas pollutant concentration prediction.

[0125] In addition, a complete set of process-based modeling procedures is provided. Major steps such as data sample enhancement, feature selection, time series order selection, outlier processing, and normalization processing can all be selected in a standard manner. Following this process, automated modeling capabilities with little human involvement can be achieved, which can be extended to other similar fields such as gas pollution and even water pollution.

[0126] See also Fig.10 , the embodiment of the present application can also provide a gas pollutant distribution prediction device, such as Fig.10 As shown, the device for executing the above-mentioned gas pollutant distribution prediction method may include:

[0127] The data acquisition unit 1001 is used to acquire the gas pollutant concentration data and meteorological element data of the target monitoring point by using the detector;

[0128] The prediction result acquisition unit 1002 is used to input the gas pollutant concentration data and the meteorological element data into the trained target distribution prediction model to perform distribution prediction to obtain the gas pollutant distribution prediction result;

[0129] The target distribution prediction model includes a multi-layer residual network and a long short-term memory network; the multi-layer residual network is used to extract spatial information from the data, and the long short-term memory network is used to extract temporal information from the data, so that the target distribution prediction model can predict the distribution of gas pollutants based on the spatial information and the temporal information;

[0130] Among them, during the training process of the target distribution prediction model, the point diffusion and linear interpolation methods are used to expand the sample data to form rectangular grid data, and the sample enhancement processing method of data resampling and data segmentation is used to enhance the sample feature expression ability of the rectangular grid data. The structural coefficient is used as the model training indicator, the recursive feature elimination method is used for feature selection, and the autocorrelation function is used to realize the time series order selection.

[0131] The embodiment of the present application may also provide a gas pollutant distribution prediction device, the device comprising a processor and a memory:

[0132] The memory is used to store program code and transmit the program code to the processor;

[0133] The processor is used to execute the steps of the above-mentioned gas pollutant distribution prediction method according to the instructions in the program code.

[0134] like Fig.11 As shown, a gas pollutant distribution prediction device provided by an embodiment of the present application may include: a processor 10, a memory 11, a communication interface 12 and a communication bus 13. The processor 10, the memory 11 and the communication interface 12 communicate with each other through the communication bus 13.

[0135] In the embodiment of the present application, the processor 10 may be a central processing unit (CPU), an application specific integrated circuit, a digital signal processor, a field programmable gate array or other programmable logic devices, etc.

[0136] The processor 10 may call a program stored in the memory 11. Specifically, the processor 10 may execute operations in an embodiment of the gas pollutant distribution prediction method.

[0137] The memory 11 is used to store one or more programs, which may include program codes, and the program codes include computer operation instructions. In the embodiment of the present application, the memory 11 at least stores programs for implementing the following functions:

[0138] Use detectors to obtain gas pollutant concentration data and meteorological element data at target monitoring points;

[0139] Inputting the gas pollutant concentration data and meteorological element data into the trained target distribution prediction model to perform distribution prediction to obtain a gas pollutant distribution prediction result;

[0140] The target distribution prediction model includes a multi-layer residual network and a long short-term memory network; the multi-layer residual network is used to extract spatial information from the data, and the long short-term memory network is used to extract temporal information from the data, so that the target distribution prediction model can predict the distribution of gas pollutants based on the spatial information and the temporal information;

[0141] Among them, during the training process of the target distribution prediction model, the point diffusion and linear interpolation methods are used to expand the sample data to form rectangular grid data, and the sample enhancement processing method of data resampling and data segmentation is used to enhance the sample feature expression ability of the rectangular grid data. The structural coefficient is used as the model training indicator, the recursive feature elimination method is used for feature selection, and the autocorrelation function is used to realize the time series order selection.

[0142] In one possible implementation, the memory 11 may include a program storage area and a data storage area, wherein the program storage area can store an operating system and application programs required for at least one function (such as a file creation function, a data reading and writing function), etc.; the data storage area can store data created during use, such as initialization data, etc.

[0143] In addition, the memory 11 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device or other volatile solid-state storage device.

[0144] The communication interface 12 may be an interface of a communication module, and is used to connect to other devices or systems.

[0145] Of course, it should be noted that Fig.11 The structure shown does not constitute a limitation on the gas pollutant distribution prediction device in the embodiment of the present application. In actual applications, the gas pollutant distribution prediction device may include Fig.11 More or fewer components than shown, or combinations of certain components.

[0146] The embodiment of the present application may also provide a computer-readable storage medium, wherein the computer-readable storage medium is used to store program code, and the program code is used to execute the steps of the above-mentioned gas pollutant distribution prediction method.

[0147] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0148] It can be known from the description of the above implementation methods that those skilled in the art can clearly understand that the present application can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product can be stored in a storage medium such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments of the present application or certain parts of the embodiments.

[0149] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can refer to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system or system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can refer to the partial description of the method embodiment. The system and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without creative work.

[0150] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A method for predicting the distribution of gaseous pollutants, characterized in that: include: Use detectors to obtain gas pollutant concentration data and meteorological element data at target monitoring points; Inputting the gas pollutant concentration data and meteorological element data into the trained target distribution prediction model to perform distribution prediction to obtain a gas pollutant distribution prediction result; The target distribution prediction model includes a multi-layer residual network and a long short-term memory network; the multi-layer residual network is used to extract spatial information from the data, and the long short-term memory network is used to extract temporal information from the data, so that the target distribution prediction model can predict the distribution of gas pollutants based on the spatial information and the temporal information; Among them, during the training process of the target distribution prediction model, the point diffusion and linear interpolation methods are used to expand the sample data to form rectangular grid data, and the sample enhancement processing method of data resampling and data segmentation is used to enhance the sample feature expression ability of the rectangular grid data. The structural coefficient is used as the model training indicator, the recursive feature elimination method is used for feature selection, and the autocorrelation function is used to realize the time series order selection.

2. The method for predicting gas pollutant distribution according to claim 1, characterized in that: Before the rectangular grid data is subjected to sample feature expression capability enhancement processing by adopting the sample enhancement processing method of data resampling and segmentation, the rectangular grid data is subjected to outlier processing and normalization processing.

3. The method for predicting gas pollutant distribution according to claim 2, characterized in that: The outlier processing includes filling missing values ​​with the mean filling method, eliminating outliers of sample data based on the 3σ principle, and calculating the value range of the data after outliers are eliminated; the normalization processing includes normalizing all maximum values ​​by increasing the original value range maximum by 9 to 11%.

4. The method for predicting gas pollutant distribution according to claim 1, characterized in that: The data resampling includes randomly resampling the data points of each group of samples, performing multiple resampling operations on each group of samples, and randomly selecting 60% to 80% of the points for zero assignment each time; the data segmentation includes equally sized segmentation of the rectangular grid data into multiple small-sized data.

5. The method for predicting gas pollutant distribution according to claim 1, characterized in that: The feature selection includes: Determine initial model parameters; Arrange and combine all features and enumerate all possible combinations of features; Calculate the evaluation index of each feature combination and sort it; Select one or more feature combinations with the highest ranking, and replace the data for re-verification; If the feature combination is verified to be stable and good after multiple times, the iteration is terminated; otherwise, the model parameters are modified and the above steps are repeated until the final model parameters and feature combination are determined.

6. The method for predicting gas pollutant distribution according to claim 1, characterized in that: The autocorrelation function includes the time series {X t }Autocorrelation coefficient of lag order kρ k Defined as: Where: X t is the value of the sequence at t, is the mean of the sequence, T is the total length of the sequence, k is the lag order, ρ k The value range is -1 to 1, 1 means perfect positive correlation, -1 means perfect negative correlation, and 0 means no correlation.

7. The method for predicting gas pollutant distribution according to claim 1, characterized in that: The target distribution prediction model also includes a Batch Normalization layer and an activation function; the Batch Normalization layer and the activation function are used to maintain the stability of the model distribution.

8. A gas pollutant distribution prediction device, characterized in that: Used to execute the gas pollutant distribution prediction method according to any one of claims 1 to 7, the device comprises: A data acquisition unit, used to acquire gas pollutant concentration data and meteorological element data at a target monitoring point using a detector; A prediction result acquisition unit, used for inputting the gas pollutant concentration data and the meteorological element data into the trained target distribution prediction model to perform distribution prediction, so as to obtain the gas pollutant distribution prediction result; The target distribution prediction model includes a multi-layer residual network and a long short-term memory network; the multi-layer residual network is used to extract spatial information from the data, and the long short-term memory network is used to extract temporal information from the data, so that the target distribution prediction model can predict the distribution of gas pollutants based on the spatial information and the temporal information; Among them, during the training process of the target distribution prediction model, the point diffusion and linear interpolation methods are used to expand the sample data to form rectangular grid data, and the sample enhancement processing method of data resampling and data segmentation is used to enhance the sample feature expression ability of the rectangular grid data. The structural coefficient is used as the model training indicator, the recursive feature elimination method is used for feature selection, and the autocorrelation function is used to realize the time series order selection.

9. A gas pollutant distribution prediction device, characterized in that: The device comprises a processor and a memory: The memory is used to store program code and transmit the program code to the processor; The processor is used to execute the gas pollutant distribution prediction method according to any one of claims 1-7 according to the instructions in the program code.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium is used to store program codes, and the program codes are used to execute the gas pollutant distribution prediction method according to any one of claims 1-7.

Images