Intelligent prediction method, system and medium for PM2.5 and ozone combined pollution

Abstract

The present invention discloses a PM2.5 and ozone composite pollution intelligent prediction method, system and medium, the method steps include: collecting historical air quality data, environmental meteorological data and atmospheric pollution source emission data of the area to be predicted; collecting cross-feature data formed by different environmental meteorological conditions and different associated pollution sources; forming the collected data into a training data set, and performing data matching; training a multi-task neural network model; selecting the optimal input feature parameters according to the incremental contribution of each input feature to the prediction of the trained multi-task neural network model, and constructing a final composite pollution identification model; obtaining real-time air quality data, environmental meteorological data and atmospheric pollution source emission data of the area to be predicted, and inputting them into the composite pollution identification model to predict PM2.5 and ozone concentration data. The present invention has the advantages of simple implementation method, low cost, high prediction accuracy and efficiency, strong robustness, etc.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent monitoring of air pollution, and in particular to an intelligent prediction method, system and medium for PM2.5 and ozone combined pollution. Background Art

[0002] Haze phenomena characterized by PM2.5 (high-concentration fine particulate matter with an aerodynamic diameter of less than 2.5 μm) and photochemical smog characterized by surface ozone have become the two main air pollution events in cities today. Usually, there are significant seasonal differences in the occurrence of PM2.5 pollution and ozone pollution. The former (PM2.5 pollution) usually occurs in winter, while the latter (ozone pollution) usually occurs in summer. However, affected by the emissions of various human pollution sources and meteorological conditions, the situation of air pollution has become increasingly complex. Under extremely unfavorable meteorological conditions, PM2.5 and ozone pollution often occur at the same time and the concentration is "double high", which brings great challenges to the actual air pollution control work. Therefore, if the combined pollution phenomenon of PM2.5 and ozone pollution can be accurately predicted and identified, it will be conducive to the effective and targeted implementation of air pollution prevention and control work.

[0003] In order to realize intelligent prediction and identification of air pollution, the prior art usually adopts machine learning algorithms such as deep learning, based on data pattern recognition and learning, and through training on historical data, it is possible to identify ambient air pollution and predict the concentration of air pollution sources in the future. For example, Chinese patent CN110766222A discloses a PM2.5 concentration prediction method based on particle swarm parameter optimization and random forest. This scheme is to collect meteorological data, pollution source concentration, time and related sites and other influencing factors, and realize the prediction of PM2.5 concentration based on particle swarm optimization random forest algorithm.

[0004] However, the above prediction and identification methods based on machine learning usually adopt a single-target modeling method, that is, considering the changes in a single pollution type, it can only realize the prediction and identification of a single pollution type. When PM2.5 and ozone pollution occur, the generation of PM2.5 and ozone will be affected by different meteorological conditions and emissions from different atmospheric pollution sources, that is, the occurrence of PM2.5 and ozone pollution is related to meteorological conditions, atmospheric pollution sources, etc. and will affect each other. For example, ozone generation will increase significantly under high temperature and high NOx / VOCs conditions, and humidity will play a key role in the secondary generation reaction of PM2.5. The single-target modeling method is not suitable for the detection of PM2.5 and ozone pollution. Even if the multi-target modeling method is considered, the direct application of the multi-target modeling method still cannot achieve accurate prediction of PM2.5 and ozone pollution due to the inability to fully explore the correlation and mutual influence characteristics between PM2.5 and ozone pollution and meteorological conditions, atmospheric pollution sources, etc., and is easily interfered by environmental factors. Summary of the invention

[0005] The technical problem to be solved by the present invention is: in response to the technical problems existing in the prior art, the present invention provides an intelligent prediction method, system and medium for PM2.5 and ozone combined pollution with simple implementation method, low cost, high prediction accuracy and efficiency, and strong robustness.

[0006] In order to solve the above technical problems, the technical solution proposed by the present invention is:

[0007] A method for intelligent prediction of PM2.5 and ozone combined pollution, comprising the following steps:

[0008] Collect air quality data, environmental meteorological data, and atmospheric pollution source emission data for the area to be predicted within a specified historical time period, wherein the air quality data includes PM2.5 concentration data and ozone concentration data;

[0009] Constructing cross-feature data based on the collected environmental meteorological data and atmospheric pollution source emission data, constructing meteorological-emission cross-features by combining different environmental meteorological conditions with different precursors to characterize the interaction between different environmental meteorological conditions and different precursor emissions, and constructing emission-emission cross-features by combining different precursors to characterize the composite effect of different precursor emissions, wherein the precursors are air pollution sources associated with the generation of PM2.5 and ozone;

[0010] The collected air quality data, environmental meteorological data, atmospheric pollution source emission data and the cross-feature data are formed into a training data set, and the environmental meteorological data, atmospheric pollution source emission data and the cross-feature data in the training data set are matched with the PM2.5 concentration data and the ozone concentration data in the air quality data;

[0011] Using the training data set to train the multi-task neural network model, during the training process, the environmental meteorological data, the atmospheric pollution source emission data and the cross-feature data are used as input data, and the corresponding matching PM2.5 concentration data and ozone concentration data are used as output data;

[0012] According to the incremental contribution of each input feature of the trained multi-task neural network model to the model prediction results, the input features are screened, the optimal input feature parameter combination is selected, and the final composite pollution identification model is constructed;

[0013] Real-time air quality data, environmental meteorological data and atmospheric pollution source emission data of the area to be predicted are obtained and input into the composite pollution identification model, which then outputs predicted PM2.5 concentration data and ozone concentration data.

[0014] Furthermore, the environmental meteorological data include any combination of temperature, relative humidity, atmospheric visibility, solar radiation intensity, wind speed and direction, boundary layer height, and daily precipitation; and the atmospheric pollution source emission data include any one or more of SO2, NOx, CO, NMVOC, NH3, PM2.5, PM10, BC, OC and CO2.

[0015] Furthermore, the meteorological-emission cross-signal is formed by the product combination of different environmental meteorological conditions and different precursors, and the meteorological-emission cross-signal includes any one or more of the cross-signal formed by temperature and NOx, light intensity, temperature and VOCs, light intensity, humidity and NOx, humidity and SO2, and humidity and OC. The emission-emission cross-signal includes and Cross-features, and Cross-features and and Any one or more of the cross-characteristics, and Cross-over features include , and Cross-over features include , and Crossover features include: ,in, Respectively represent NOx, 、 as well as emissions.

[0016] Further, the using the training data set to train the multi-task neural network model includes:

[0017] The training data set is input into a pre-built multi-task neural network model, which includes an input layer, a network sharing layer and a task-specific layer arranged in sequence. The air quality data, environmental meteorological data, atmospheric pollution source emission data and the cross-feature data are input into the input layer, and the common features in the input data are extracted through the shared layers composed of cascades in the network sharing layer in sequence to obtain the common feature representation. , represented by the task-specific layer according to the common features Predict PM2.5 concentration data and ozone concentration data output;

[0018] The error losses between the PM2.5 concentration data and the ozone concentration data output by the task-specific layer and the true label are calculated respectively, and the error losses are back-propagated to update the parameters of the network sharing layer and the task-specific layer;

[0019] When the error loss reaches the preset condition, the training is completed and a model with the optimal neural network model parameters is obtained.

[0020] Furthermore, the network sharing layer extracts the common features in the input data according to the following formula to obtain the common feature representation: :

[0021]

[0022]

[0023] in, Indicates the number of shared layers, is the activation function Sigmoid, and are the weight matrix and bias vector of the first layer in the network shared layer, is the input layer data, and Respectively The weight matrix and bias vector of the layer, , Respectively represent Layer, The output of the layer.

[0024] Furthermore, the task-specific layer predicts the PM2.5 concentration data and ozone concentration data output according to the following formula:

[0025]

[0026] in, Indicates output targets, i =1,2, corresponding to PM 2.5 Concentration data, ozone concentration data, Indicates the number of task-specific layers, and They are respectively i The weight matrix and bias vector of the target in the first layer, and They are respectively The output target The weight matrix and bias vector of the layer, Respectively i The goal is M layer, M- Output of layer 1.

[0027] Furthermore, the input features are screened according to the incremental contribution of each input feature of the trained multi-task neural network model to the model prediction result, and the optimal input feature parameter combination is selected, including:

[0028] Record the initial characteristic parameter combination of the composite pollution identification model obtained after training and the corresponding model error loss ;

[0029] Evaluate the incremental contribution of each input feature to the prediction result of the composite pollution identification model and rank the importance;

[0030] According to the importance ranking results of each input feature, a specified number of features with the lowest contribution are eliminated to form a feature parameter combination containing Kn features and obtain the corresponding model error loss , K is the total number of input features, n is the number of features to be removed, and then the specified number of features with the lowest contribution are removed from the remaining features to form a new feature parameter combination and obtain the corresponding model error loss until the last feature remains in the model;

[0031] The model with the lowest error loss is screened out from the models corresponding to the obtained characteristic parameter combinations, and the characteristic parameter combination of the screened model is used as the optimal input characteristic parameter combination.

[0032] Furthermore, the Shapley value is used as the incremental contribution of each input feature to the prediction result of the composite pollution identification model, and the calculation formula of the Shapley value is:

[0033]

[0034] in, It is a feature Shapley value of N, where N is the set of all input features and S is the set of features not included. A subset of features, is the size of the feature subset, represents the model output obtained by predicting the features in subset S, Indicates that the feature Model output after adding to subset S.

[0035] A computer system comprises a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to execute the computer program to perform the above method.

[0036] A computer-readable storage medium storing a computer program, wherein the computer program implements the above method when executed by a processor.

[0037] Compared with the prior art, the advantages of the present invention are as follows: the present invention collects air quality data, environmental meteorological data and atmospheric pollution source emission data, and collects cross-feature data formed by different environmental meteorological conditions and different associated pollution sources to form a training data set, trains a multi-task neural network model to build a PM2.5 and ozone composite pollution intelligent recognition model, and can fully explore the correlation and mutual influence characteristics between environmental meteorological data, atmospheric pollution source emission data and PM2.5 and ozone concentration data, and effectively enhance the model's recognition of PM 2.5 The ability to capture the complex mechanism of ozone generation, combined with the selection of the optimal input feature parameter combination based on the incremental contribution of each input feature to the model prediction results, can obtain the optimal composite pollution identification model, which enables real-time and efficient prediction of PM2.5 and ozone concentrations in the area, significantly improving PM 2.5 and ozone complex pollution prediction identification accuracy and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a schematic diagram of the implementation process of the intelligent prediction method for PM2.5 and ozone combined pollution in this embodiment.

[0039] Figure 2 Schematic diagram of the structure principle of the multi-task neural network model used in this embodiment. DETAILED DESCRIPTION

[0040] The present invention is further described below in conjunction with the accompanying drawings and specific preferred embodiments, but the protection scope of the present invention is not limited thereby.

[0041] Changes in environmental meteorological conditions are the main factor in the evolution and fluctuation of air pollution events. When air pollution occurs, the pollution sources are generally relatively stable in the short term. Therefore, the occurrence of PM2.5 and ozone compound pollution will have specific correlation characteristics with meteorological conditions and air pollution sources. For example, ozone generation is significantly related to temperature increase, especially under high temperature and high NOx / VOCs conditions, ozone generation increases significantly, and light intensity directly affects the reaction rate of VOCs and NOx, that is, light intensity directly affects the reaction rate of air pollution sources, and indirectly affects ozone generation. Humidity plays a key role in the secondary generation reaction of PM2.5, affecting the generation of inorganic components of PM2.5 on the one hand, and the generation of organic components of PM2.5 on the other hand. That is, different environmental meteorological conditions such as temperature and humidity and different atmospheric pollution sources will form different PM2.5 and ozone complex pollution. Various environmental meteorological conditions and atmospheric pollution sources will influence each other in the formation of PM2.5 and ozone. Therefore, to accurately identify PM2.5 and ozone complex pollution events, it is necessary to fully explore the correlation between PM2.5 and ozone complex pollution and environmental meteorological conditions and atmospheric pollution sources, as well as their mutual influence.

[0042] The present invention constructs an intelligent identification model for composite pollution of PM2.5 and ozone based on a multi-task neural network model, collects air quality data, environmental meteorological data and atmospheric pollution source emission data, and considers the mutual influence of quasi-environmental meteorological items and atmospheric pollution sources on the formation of PM2.5 and ozone. Based on the environmental meteorological data and the atmospheric pollution source emission data, meteorological-emission cross-features and emission-emission cross-features are constructed to characterize the interaction between different environmental meteorological conditions and different precursor emissions and between different precursor emissions. These data together constitute a training data set, and the environmental meteorological data, atmospheric pollution source emission data and cross-feature data are matched with PM2.5 concentration data and ozone concentration data respectively. Then, through model training, the correlation and mutual influence characteristics between the environmental meteorological data, atmospheric pollution source emission data and PM2.5 and ozone concentration data can be fully explored, and the model can effectively enhance the PM 2.5 The ability to capture the complex mechanism of ozone generation and select the optimal input feature parameter combination according to the incremental contribution of each input feature to the model prediction result can obtain the optimal composite pollution identification model. This model can be used to predict PM2.5 and ozone concentrations in the area in real time and efficiently, significantly improving PM 2.5 The accuracy and robustness of ozone complex pollution prediction and identification can help to timely grasp the dynamic changes of air pollution.

[0043] like Figure 1 As shown, the steps of the PM2.5 and ozone combined pollution intelligent prediction method in this embodiment include:

[0044] Step S01: Collect air quality data, environmental meteorological data and atmospheric pollution source emission data of the area to be predicted within a specified historical time period.

[0045] Air quality data mainly involves the surface PM2.5 concentration data and ozone concentration data of atmospheric complex pollution. Environmental meteorological data can specifically include temperature, relative humidity, atmospheric visibility, solar radiation intensity, wind speed and direction, boundary layer height, daily precipitation and other data, and atmospheric pollution source emission data can specifically include the emission of various types of pollution sources such as SO2, NOx, CO, NMVOC, NH3, PM2.5, PM10, BC, OC and CO2. The specific data types of environmental meteorological data and atmospheric pollution source emission data can be selected and configured according to actual needs.

[0046] As an optional implementation, the PM of the area to be predicted within a historical period (e.g., 1 year or more) can be obtained. 2.5 The daily average concentration data and the maximum eight-hour concentration data of ozone form an air quality data set, and obtain relevant meteorological data that affect the occurrence of air pollution, including temperature, relative humidity, atmospheric visibility, solar radiation intensity, wind speed and direction, boundary layer height, daily precipitation, etc., to form a meteorological data set, and at the same time obtain SO2, NOx, CO, NMVOC, NH3, PM 2.5 、PM 10 , BC, OC and CO2, etc., to form an atmospheric pollution source emission data set. In order to keep consistent with the daily air quality data in terms of time scale, except for the daily precipitation, all other meteorological data are daily average data, and the pollution source emission data are converted into daily atmospheric pollution source emissions according to the daily emission factors of atmospheric pollution sources in the target area. Further, the time information corresponding to the above data can be extracted, including Unix timestamp, Julian day, day of the week, etc.

[0047] Furthermore, it also includes normalizing the collected raw meteorological data, emission data, etc. After the raw data is standardized, the linear interpolation method can be used to process the air quality data and meteorological data containing abnormal values ​​and missing values ​​to ensure the integrity of the data set. Among them, the normalization process can use the following formula:

[0048] (1)

[0049] In the formula, is the standardized data, For Daily raw data, is the average value of the original data for all days in the historical period, is the standard deviation.

[0050] Step S02. Construct cross-feature data based on the collected ambient meteorological data and atmospheric pollution source emission data, construct meteorological-emission cross-features by combining different ambient meteorological conditions with different precursor emissions, so as to characterize the interaction between different ambient meteorological conditions and different associated pollution sources, and construct emission-emission cross-features by combining different precursors, so as to characterize the composite effect of emissions from different pollution sources, and construct cross-feature data by meteorological-emission cross-features and emission-emission cross-features. Precursors are pollution sources associated with the generation of PM2.5 and ozone, such as NOx, VOCs, SO2, and OC.

[0051] This embodiment is based on PM 2.5 The physical and chemical formation mechanism of ozone and the use of meteorological and precursor factors to construct a model that can reflect PM 2.5 The cross-characteristics of the secondary generation mechanism of ozone make the neural network model have the ability to capture different environmental meteorological conditions, atmospheric pollution sources and PM 2.5 The correlation characteristics between oxygen and ozone and their mutual influence characteristics can be enhanced, while the model’s ability to extract nonlinear relationships can be improved, thereby assisting the machine learning process and improving the model’s predictive performance.

[0052] In this embodiment, the cross-signal mainly includes two aspects: one is the interaction effect between the ambient meteorological conditions and the precursor emissions, and the other is the interaction effect between the emissions of different precursors. The use of these cross-signal can not only reflect the interaction between meteorological factors and pollution sources, but also reveal the factors that affect PM2.5. 2.5 And the emission interaction effects between precursors of ozone generation (such as NOx, VOCs, SO2, OC, etc.). Based on domain knowledge, cross-features can be constructed from relevant environmental meteorological conditions and emission sources to form new features that can reflect their interactions. Specifically, meteorological-emission cross-features can include cross-features of temperature, light intensity, etc. with NOx, VOCs, cross-features of atmospheric humidity with NOx, SO2, and OC, etc. In addition, considering the emission interaction effects between different precursors, emission-emission cross-features are constructed. For example, cross-features can be formed between NOx and VOCs, and cross-features can be formed between NH3, NOx, and SO2 to characterize the interactions between NOx and VOCs, NH3, NOx, and SO2. These interaction effects can be used to effectively improve PM 2.5 and ozone complex pollution prediction accuracy.

[0053] As an optional implementation, the following types of cross features may be used:

[0054] 1) Interaction between meteorological conditions and precursor emissions

[0055] Considering that under high temperature and strong light conditions, the photochemical reaction of NOx and VOCs is enhanced, which contributes to the formation of surface ozone. , light intensity The product of NOx and VOCs is used as the cross feature , to capture the nonlinear effect of ozone generation under high temperature and strong light conditions, namely:

[0056] (2)

[0057] (3)

[0058] Secondly, considering the influence of humidity on PM 2.5 It plays a key role in the secondary generation reaction of PM. 2.5 The formation of inorganic components, on the other hand, affects PM 2.5 The generation of organic components can be divided into humidity The product of NOx, SO2 and OC is used as the cross-signal ,Right now:

[0059] (4)

[0060] (5)

[0061] (6)

[0062] 2) Interaction effects between emissions of different precursors

[0063] Considering that the generation of ozone is nonlinearly related to NOx and VOCs, the ratio of NOx / VOCs reveals whether the ozone generation is controlled by NOx or VOCs. Based on this, the cross-characteristics of the ratio of NOx / VOCs are constructed to reflect the sensitive conditions of ozone generation, namely:

[0064] (7)

[0065] In addition, considering that ammonia as the only alkaline gas in the atmosphere will react chemically with NOx and SO2 in the atmosphere to generate secondary inorganic particles, based on this, the product characteristics of NH3, NOx and SO2 are constructed, namely:

[0066] (8)

[0067] (9)

[0068] In this embodiment, after the collected original data is normalized and dimensioned, the above method is used to construct a cross-feature data set. Of course, the cross-feature types may not be limited to the above cross-feature types.

[0069] Step S03. Form a training data set with air quality data, environmental meteorological data, atmospheric pollution source emission data and cross-feature data, and perform data matching on the environmental meteorological data, atmospheric pollution source emission data and cross-feature data in the training data set with the PM2.5 concentration data and ozone concentration data in the air quality data.

[0070] After collecting air quality data, environmental meteorological data, atmospheric pollution source emission data, and cross-feature data, the above data are combined to form a training data set. 2.5 , ozone concentration data are matched on a time series basis, for example, on a daily scale, to determine the relationship between meteorological data, atmospheric pollution source emission data, cross-feature data and PM 2.5 , and the correspondence between ozone concentration data.

[0071] Step S04. Use the training data set to train the multi-task neural network model. During the training process, environmental meteorological data, atmospheric pollution source emission data and cross-feature data are used as input data, and the corresponding matching PM2.5 concentration data and ozone concentration data are used as output data.

[0072] Use training set data to train PM 2.5 A multi-task neural network model of compound pollution with ozone is constructed to form a compound pollution identification model. During the model training process, meteorological data, emission data, time information and cross-features are used as the input data (predictor variables) of the model. 2.5 , ozone concentration data as the output data of the model (target variable), and then the matched data set is randomly divided into a training set and a test set. The training set data can be described by the following formula:

[0073] (10)

[0074] in, is the input feature vector of the training set, including all meteorological data, emission data, time information and cross-features. Features, is the output vector, including two features: PM2.5 and ozone concentration. N is the sample size of the training set.

[0075] As an optional implementation, the division of training set and test set can be completed through 5-fold cross validation, that is, all sample data are randomly divided into 5 subsets and calculated 5 times. In each calculation, 4 subsets are used to train the neural network model, and then the remaining 1 subset is used to test the atmospheric pollution characteristic results. The final result is the average of the 5 root mean square errors.

[0076] Since there is a correlation between the predictions of PM2.5 and ozone, in order to fully exploit this correlation, this embodiment sets a network sharing layer based on a multi-task neural network model to extract common features in the input data to improve the overall performance of the model. After the network sharing layer, a task-specific layer is set to extract common features from the common feature representation. The results are then further processed to generate the multi-objective task outputs of PM2.5 concentration and ozone concentration prediction.

[0077] As an optional implementation, training the multi-task neural network model using the training data set includes:

[0078] Step S401. Input the training data set into the pre-built multi-task neural network model, such as Figure 2 As shown in FIG. 1 , the multi-task neural network model includes an input layer, a network sharing layer, and a task-specific layer, which are arranged in sequence. The network sharing layer and the task-specific layer are both composed of multiple layers in cascade. The input layer inputs air quality data, environmental meteorological data, atmospheric pollution source emission data, and cross-feature data. The shared layers composed of the cascades in the network sharing layer extract the common features in the input data in sequence to obtain the common feature representation. , represented by the task-specific layer based on the common features Predict PM2.5 concentration data and ozone concentration data output;

[0079] Step S402. Calculate the error loss (e.g., root mean square error) between the PM2.5 concentration data and the ozone concentration data output by the task-specific layer and the true label, and back-propagate the error loss to update the parameters of the network sharing layer and the task-specific layer;

[0080] Step S403: When the error loss reaches a preset condition, the training is completed and a model with optimal neural network model parameters is obtained.

[0081] As an optional method, the network sharing layer extracts the common features in the input data according to the following formula to obtain the common feature representation :

[0082] (11)

[0083] (12)

[0084] in, Indicates the number of shared layers, is the activation function Sigmoid, and are the weight matrix and bias vector of the first layer in the network shared layer, is the input layer data, and Respectively The weight matrix and bias vector of the layer, , Respectively represent Layer, The output of the layer.

[0085] As an optional implementation, the task-specific layer also adopts a multi-layer cascade method. For each target task output, it can be described by the following formula:

[0086] (13)

[0087] in, Indicates output targets, i =1,2, corresponding to PM 2.5 Concentration data, ozone concentration data, Indicates the number of task-specific layers, and They are respectively i The weight matrix and bias vector of the target in the first layer, and They are respectively The output target The weight matrix and bias vector of the layer, Respectively i The goal is M layer, M- Output of layer 1.

[0088] As an optional implementation, during the training process, the task-specific layer PM 2.5 The ozone output results are compared with the true labels, and the root mean square error is calculated. Then the error loss is back-propagated using the gradient update to update the parameters of the shared layer and the task-specific layer, so as to train the neural network model to minimize the error loss and obtain the optimal neural network model parameters. Then the test sample is input into the trained model, and the PM output of the prediction model is recorded. 2.5 and ozone concentration values, and calculated PM 2.5The root mean square error between the predicted value and the true value of ozone concentration was calculated. Then, based on 5-fold cross validation, the above model training and testing were repeated 5 times. Finally, the average root mean square error of the 5 test sets was taken to measure the prediction performance of the model.

[0089] Step S05. Optimize and screen the input features according to the incremental contribution of each input feature of the multi-task neural network model trained in step S04 to the model prediction results, select the optimal input feature parameter combination, and construct the final composite pollution identification model.

[0090] Step S501: Record the initial characteristic parameter combination of the composite pollution identification model obtained after training and the corresponding model error loss , the model error loss It represents the error loss when all characteristic factors (i.e. meteorological historical data set, atmospheric pollution source emission data set, time information variables and cross-characteristics) are included in the model;

[0091] Step S502: Evaluate the incremental contribution of each input feature to the prediction result of the composite pollution identification model and rank them by importance;

[0092] Step S503: Eliminate a specified number of features with the lowest contribution according to the importance ranking results of each input feature, form a feature parameter combination containing Kn features and obtain the corresponding model error loss , K is the total number of input features, n is the number of features to be removed, and then the specified number of features with the lowest contribution are removed from the remaining features to form a new feature parameter combination and obtain the corresponding model error loss until the last feature remains in the model;

[0093] Step S504: Filter out the model with the lowest error loss from the models corresponding to the obtained characteristic parameter combinations, and use the characteristic parameter combination of the filtered model as the optimal input characteristic parameter combination to construct a final composite pollution identification model.

[0094] As an optional implementation, the Shapley value can be used as the incremental contribution of each input feature to the prediction result of the composite pollution identification model. The calculation formula of the Shapley value is:

[0095] (14)

[0096] in, It is a feature Shapley value of N, where N is the set of all input features and S is the set of features not included. A subset of features, is the size of the feature subset, represents the model output obtained by predicting the features in subset S, Indicates that the feature Model output after adding to subset S.

[0097] By using the Shapley value method to evaluate the contribution of each input feature to the model one by one and gradually eliminating unimportant features with low contribution to the model, the model structure can be simplified and the consumption of computing resources can be reduced. At the same time, the simplified model with the best prediction performance can be finally selected to predict the complex pollution characteristics and ensure the optimal prediction performance.

[0098] As an optional implementation, the initial feature parameter combination of the neural network model constructed in step S04 and the error loss of the prediction model at this time can be recorded first. (take the root mean square error loss), then calculate the Shapley value to quantitatively evaluate the impact of each feature of the input neural network model on the model, sort the features according to their importance, remove the input features with the lowest contribution to the model, and repeat step S04 to obtain the error loss when K-1 feature factors are included , and so on, until only the last characteristic factor is left in the neural network model, at which point the error loss of the model is , the neural network model with the lowest error loss is selected from the simplified model structure that eliminates the characteristic factors with low contribution to the model in turn. At this time, the neural network model is the required optimal input characteristic parameter combination, and the corresponding prediction performance of the model is also optimal.

[0099] Step S06. Obtain real-time air quality data, environmental meteorological data, and atmospheric pollution source emission data of the area to be predicted, and input them into the composite pollution identification model, which outputs predicted PM2.5 concentration data and ozone concentration data.

[0100] First, obtain real-time air quality data, environmental meteorological data, and atmospheric pollution source emission data as input data, and then input them into the constructed composite pollution identification model. The composite pollution identification model can predict PM in real time. 2.5 and ozone concentration.

[0101] Furthermore, the predicted PM 2.5 The pollution characteristics of ozone concentration can be identified and classified based on the air quality standard threshold to provide early warning of complex atmospheric pollution types and help optimize air pollution prevention and control measures. For example, PM2.5 can be classified according to the Ambient Air Quality Standard (GB 3095-2012). 2.5 Concentrations exceeding 75 μg / m 3 Or the maximum eight-hour daily ozone concentration exceeds 160 μg / m3 It is designated as a day of severe air pollution, so the air pollution characteristics can be divided into four categories: PM 2.5 Heavy pollution day (σ[PM 2.5 ]>= 75, σ[O3]<160), severe ozone pollution day (σ[PM 2.5 ]<75,σ[O3]>= 160)、PM 2.5 Combined ozone pollution day (σ[PM 2.5 ]>=75, σ[O3]>= 160) and no heavy pollution days. According to the above method, the type of atmospheric complex pollution in the area to be predicted can be accurately identified.

[0102] This embodiment also provides an intelligent prediction device for PM2.5 and ozone combined pollution, including:

[0103] A data collection module is used to collect air quality data, environmental meteorological data, and atmospheric pollution source emission data of the area to be predicted within a specified historical time period, wherein the air quality data includes PM2.5 concentration data and ozone concentration data;

[0104] A cross-feature construction module, used to construct cross-feature data based on the collected ambient meteorological data and atmospheric pollution source emission data, to construct meteorological-emission cross-features by combining different ambient meteorological conditions with different precursors to characterize the interaction between different ambient meteorological conditions and different precursor emissions, and to construct emission-emission cross-features by combining different precursors to characterize the composite effect of different precursor emissions, wherein the precursors are air pollution sources associated with the generation of PM2.5 and ozone;

[0105] A training set construction module, used to form a training data set from the air quality data, environmental meteorological data, atmospheric pollution source emission data, and the cross-feature data, and to perform data matching on the environmental meteorological data, atmospheric pollution source emission data, and the cross-feature data in the training data set with the PM2.5 concentration data and ozone concentration data in the air quality data;

[0106] A model training module, used to train a multi-task neural network model using the training data set, wherein during the training process, environmental meteorological data, atmospheric pollution source emission data and the cross-feature data are used as input data, and corresponding matching PM2.5 concentration data and ozone concentration data are used as output data;

[0107] The model building module is used to select the optimal input feature parameter combination according to the incremental contribution of each input feature of the trained multi-task neural network model to the prediction result of the trained multi-task neural network model, and construct the final composite pollution identification model;

[0108] The composite pollution prediction module obtains real-time air quality data, environmental meteorological data and atmospheric pollution source emission data of the area to be predicted, and inputs them into the composite pollution identification model, which then outputs predicted PM2.5 concentration data and ozone concentration data.

[0109] The intelligent prediction device for PM2.5 and ozone combined pollution in this embodiment corresponds one to one with the above-mentioned intelligent prediction method for PM2.5 and ozone combined pollution, and will not be described in detail here.

[0110] This embodiment further provides a computer system, including a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to execute the computer program to perform the above method.

[0111] It is understandable that the above method of this embodiment can be executed by a single device, such as a computer or server, etc., and can also be applied to a distributed scenario and completed by multiple devices in cooperation with each other. In the case of a distributed scenario, one of the multiple devices can only execute one or more steps in the above method of this embodiment, and multiple devices interact to complete the above method. The processor can be implemented in the form of a general-purpose CPU, a microprocessor, an application-specific integrated circuit, or one or more integrated circuits, etc., for executing related programs to implement the above method of this embodiment. The memory can be implemented in the form of a read-only memory ROM, a random access memory RAM, a static storage device, and a dynamic storage device. The memory can store an operating system and other applications. When the above method of this embodiment is implemented by software or firmware, the relevant program code is stored in the memory and called and executed by the processor.

[0112] This embodiment further provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the above method is implemented.

[0113] Those skilled in the art should understand that the above-mentioned embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application can take the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the process Figure 1 A process or multiple processes and / or boxes Figure 1 These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer-readable memory produce a product including an instruction device, which implements the functions specified in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide for implementing the process in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0114] The above is only a preferred embodiment of the present invention, and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment, it is not intended to limit the present invention. Therefore, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the scope of protection of the technical solution of the present invention.

Claims

1. A method for intelligent prediction of PM2.5 and ozone combined pollution, characterized in that the steps include: Collect air quality data, environmental meteorological data, and atmospheric pollution source emission data for the area to be predicted within a specified historical time period, wherein the air quality data includes PM2.5 concentration data and ozone concentration data; Constructing cross-feature data based on the collected environmental meteorological data and atmospheric pollution source emission data, wherein meteorological-emission cross-features are constructed by combining different environmental meteorological conditions with different precursors to characterize the interaction between different environmental meteorological conditions and different precursor emissions, and emission-emission cross-features are constructed by combining different precursors to characterize the composite effect of different precursor emissions, wherein the precursors are air pollution sources associated with the generation of PM2.5 and ozone; The air quality data, environmental meteorological data, atmospheric pollution source emission data, and the cross-feature data form a training data set, and perform data matching on the environmental meteorological data, atmospheric pollution source emission data, and the cross-feature data in the training data set with the PM2.5 concentration data and the ozone concentration data in the air quality data; Using the training data set to train the multi-task neural network model, during the training process, the environmental meteorological data, the atmospheric pollution source emission data and the cross-feature data are used as input data, and the corresponding matching PM2.5 concentration data and ozone concentration data are used as output data; According to the incremental contribution of each input feature of the trained multi-task neural network model to the model prediction results, the input features are screened, the optimal input feature parameter combination is selected, and the final composite pollution identification model is constructed; Real-time air quality data, environmental meteorological data and atmospheric pollution source emission data of the area to be predicted are obtained and input into the composite pollution identification model, which then outputs predicted PM2.5 concentration data and ozone concentration data.

2. The method for intelligent prediction of PM2.5 and ozone combined pollution according to claim 1, characterized in that: The environmental meteorological data include any combination of temperature, relative humidity, atmospheric visibility, solar radiation intensity, wind speed and direction, boundary layer height, and daily precipitation; the atmospheric pollution source emission data include any one or more of SO2, NOx, CO, NMVOC, NH3, PM2.5, PM10, BC, OC and CO2.

3. The method for intelligent prediction of PM2.5 and ozone combined pollution according to claim 1, characterized in that: The meteorological-emission cross-signal is formed by the product combination of different environmental meteorological conditions and different precursors. The meteorological-emission cross-signal includes any one or more of the cross-signal formed by temperature and NOx, light intensity, temperature and VOCs, light intensity, humidity and NOx, humidity and SO2, and humidity and OC. The emission-emission cross-signal includes and Cross-features, and Cross-features and and Any one or more of the cross-characteristics, and Cross-over features include , and Cross-over features include , and Cross-over features include ,in, Respectively represent NOx, 、 as well as emissions.

4. The method for intelligent prediction of PM2.5 and ozone combined pollution according to claim 1, characterized in that: The using the training data set to train the multi-task neural network model comprises: The training data set is input into a pre-built multi-task neural network model, which includes an input layer, a network sharing layer and a task-specific layer arranged in sequence. The air quality data, environmental meteorological data, atmospheric pollution source emission data and the cross-feature data are input into the input layer, and the common features in the input data are extracted through the shared layers composed of cascades in the network sharing layer in sequence to obtain the common feature representation. , represented by the task-specific layer according to the common features Predict PM2.5 concentration data and ozone concentration data output; The error losses between the PM2.5 concentration data and the ozone concentration data output by the task-specific layer and the true label are calculated respectively, and the error losses are back-propagated to update the parameters of the network sharing layer and the task-specific layer; When the error loss reaches the preset condition, the training is completed and a model with the optimal neural network model parameters is obtained.

5. The method for intelligent prediction of PM2.5 and ozone combined pollution according to claim 4 is characterized in that: The network sharing layer extracts the common features in the input data according to the following formula to obtain the common feature representation: : in, Indicates the number of shared layers, is the activation function Sigmoid, and are the weight matrix and bias vector of the first layer in the network shared layer, is the input layer data, and Respectively The weight matrix and bias vector of the layer, , Respectively represent Layer, The output of the layer.

6. The method for intelligent prediction of PM2.5 and ozone combined pollution according to claim 4, characterized in that: The task-specific layer predicts the PM2.5 concentration data and ozone concentration data output according to the following formula: in, Indicates output targets, i =1,2, corresponding to PM 2.5 Concentration data, ozone concentration data, Indicates the number of task-specific layers, and They are respectively i The weight matrix and bias vector of the target in the first layer, and They are respectively The output target The weight matrix and bias vector of the layer, Respectively i The goal is M layer, M- Output of layer 1.

7. The method for intelligent prediction of PM2.5 and ozone combined pollution according to any one of claims 1 to 6, characterized in that: The input features are screened according to the incremental contribution of each input feature of the trained multi-task neural network model to the model prediction result, and the optimal input feature parameter combination is selected, including: Record the initial characteristic parameter combination of the composite pollution identification model obtained after training and the corresponding model error loss ; Evaluate the incremental contribution of each input feature to the prediction result of the composite pollution identification model and rank the importance; According to the importance ranking results of each input feature, a specified number of features with the lowest contribution are eliminated to form a feature parameter combination containing Kn features and obtain the corresponding model error loss , K is the total number of input features, n is the number of features to be removed, and then the specified number of features with the lowest contribution are removed from the remaining features to form a new feature parameter combination and obtain the corresponding model error loss until the last feature remains in the model; The model with the lowest error loss is screened out from the models corresponding to the obtained characteristic parameter combinations, and the characteristic parameter combination of the screened model is used as the optimal input characteristic parameter combination.

8. The method for intelligent prediction of PM2.5 and ozone combined pollution according to claim 7, characterized in that: The Shapley value is used as the incremental contribution of each input feature to the prediction result of the composite pollution identification model. The calculation formula of the Shapley value is: in, It is a feature Shapley value of N, where N is the set of all input features and S is the set of features not included. A subset of features, is the size of the feature subset, represents the model output obtained by predicting the features in subset S, Indicates that the feature Model output after adding to subset S.

9. A computer system comprising a processor and a memory, wherein the memory is used to store a computer program, wherein: The processor is configured to execute the computer program to perform the method according to any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images