An air quality prediction method based on a hybrid deep learning model

Abstract

The application belongs to the technical field of air quality monitoring, and particularly relates to an air quality prediction method based on a hybrid deep learning model, which comprises the following steps: (1) constructing an air quality prediction data set and pre-processing air quality data; (2) decomposing the pre-processed air quality data set into a plurality of independent intrinsic mode function (IMF) sub-data sets and a residual sub-data set; (3) performing feature selection on the intrinsic mode function (IMF) sub-data sets and the residual sub-data set, selecting a sub-data set conducive to prediction and reconstructing to obtain an optimized air quality data set; (4) adopting a three-layer bidirectional long short-term memory neural network (BLSTM) and a self-attention mechanism in a series connection manner to construct a deep BALSTM model; and (5) inputting the optimized air quality data set into the constructed model for learning and prediction to obtain a final air quality prediction result.

Description

Technical Field

[0001] This invention belongs to the field of air quality monitoring technology, specifically relating to an air quality prediction method based on a hybrid deep learning model. Background Technology

[0002] With social development and improved living standards, air pollution has become a very urgent environmental problem facing the global population.

[0003] Air quality is typically described quantitatively using the Air Quality Index (AQI), which is calculated from the concentrations of PM2.5, PM10, SO2, NO2, O3, and CO. However, air quality data exhibits nonlinear and non-stationary characteristics, leading to low accuracy in traditional prediction methods. In recent years, deep learning methods have been increasingly applied to air quality prediction, demonstrating better predictive performance. However, current methods rarely handle noise in air quality data effectively and seldom consider the correlations between historical air quality data over time periods, resulting in insufficient prediction accuracy and slow convergence speed. Summary of the Invention

[0004] The purpose of this invention is to provide an air quality prediction method based on a hybrid deep learning model, so as to improve the accuracy and convergence speed of air quality prediction.

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

[0006] An air quality prediction method based on a hybrid deep learning model includes the following steps:

[0007] S1. Construct an air quality prediction dataset and preprocess the air quality data;

[0008] Collect raw air quality data for the city to be predicted, i.e., hourly air quality data, including data on six air pollutants: PM2.5, PM10, SO2, NO2, O3, and CO, and the Air Quality Index (AQI); create seven air parameter arrays consisting of the above six air pollutant data and AQI as the initial dataset for air quality prediction; perform preprocessing on the initial air quality dataset, including missing value handling, seasonality handling, and standardization.

[0009] S2. The preprocessed air quality dataset is decomposed into multiple independent intrinsic mode function (IMF) subsets and a residual subset using the adaptive noise complete set empirical mode decomposition (CEEMDAN) method, which enhances the temporal characteristics of the air quality data while eliminating noise.

[0010] S3. The improved Chaotic Binary Raven Search (CBCSA) algorithm is used to perform feature selection on the aforementioned Intrinsic Mode Function (IMF) subset and residual subset. The subset that is favorable for prediction is selected and reconstructed to obtain the optimized air quality dataset.

[0011] S4. Constructing a deep learning model, BALSTM

[0012] A deep learning BALSTM model is constructed by cascading a three-layer bidirectional long short-term memory neural network (BLSTM) with a self-attention mechanism. The model adds the Dropout method to the first two BLSTM layers to reduce the overfitting problem and selects Sigmoid as the activation function of the self-attention mechanism layer. The final result is input into the fully connected layer as the output of the final prediction result.

[0013] S5. Input the optimized air quality dataset obtained in step S3 into the deep learning BALSTM model constructed in step S4 for learning and prediction, and obtain the final air quality prediction result.

[0014] Furthermore, in step S1, the missing value handling method is as follows: for the missing air quality data, the average value of the original air quality data is used to fill in the missing data.

[0015] Seasonal processing includes defining a seasonal index, calculated using the following formula:

[0016]

[0017] Where Se represents the seasonal factor, and seA i This represents the average of the data for each season in the collected raw air quality data over the years, i = 1, 2, 3, 4, representing the four seasons of spring, summer, autumn and winter respectively, and yearA represents the average value of the collected raw air quality data.

[0018] The standardization process is as follows: First, calculate the mean (yearA) and standard deviation (σ) of all raw air quality data, and then perform Z-score standardization on them. The calculation formula is as follows:

[0019] x=(x0-yearA) / σ

[0020] Where x0 represents the air quality data before any standardization process, and x represents the air quality data after standardization process.

[0021] Furthermore, the specific process of step S2 is as follows:

[0022] S21: Set the total number of intrinsic mode functions to be decomposed, maxImf;

[0023] S22: Add adaptive noise to the seven air parameter arrays of the air quality dataset after step S1, and the resulting new dataset is denoted as X;

[0024] S23: Find all local maxima and minima in X, and fit them with cubic spline interpolation functions to form the upper and lower envelopes of X. Subtract the mean of the upper and lower envelopes from X to obtain a new dataset. If there are no negative local maxima or positive local minima in the new dataset, terminate the data processing; otherwise, continue the above process until there are no negative local maxima or positive local minima. The dataset after this decomposition is denoted as the Intrinsic Mode Function dataset IMF1.

[0025] S24: Subtract the extracted dataset IMF1 from X to obtain the remaining dataset RES1: RES1 = X - IMF1;

[0026] S25: Treat RES1 as X and repeat steps S23 to S24 until the number of decompositions reaches maxImf, thus obtaining a sequence of intrinsic mode function datasets IMF. i The dataset RES of sum items i , i = 1, 2, ..., maxImf; the last remainder dataset is the residual dataset;

[0027] S26: Combine all the above intrinsic mode function datasets (IMF) i Combined with the residual dataset, they form a generalized array, which is the air quality dataset X' decomposed using the CEEMDAN method.

[0028] Furthermore, step S3 performs feature selection on the Intrinsic Mode Function (IMF) subset and the residual subset, and the specific process for selecting the subset that is beneficial for prediction is as follows:

[0029] S31: Initialize the chaotic binary crow search algorithm;

[0030] Set the number of crows n and the maximum perception probability AP. max Minimum perceptual probability AP min Maximum flight distance fl max Minimum flight distance fl min The parameters such as the maximum number of iterations tmax are set, and the chaotic mapping value of each crow is created using the Logistic chaotic mapping function;

[0031] S32: Use each air quality data point in the decomposed air quality dataset X' output in step S2 as the initial position p of each crow. i and initial memory m i , i = 1, 2, ..., n;

[0032] S33: Update the position of each crow;

[0033] In each iteration, the perception probability and flight distance of each crow are dynamically updated using the following formula:

[0034]

[0035]

[0036] Where t is the iteration round, tmax is the maximum number of iterations, and AP t Let fl be the probability of the crow's perception in the t-th iteration. t Let be the distance the crow flies in the t-th iteration;

[0037] Then update the position of each crow using the following formula.

[0038]

[0039] Where, p i,t+1 Let p be the position of crow i in the (t+1)th iteration. i,t Let C be the position of crow i in the t-th iteration. i and C j Let m be the chaotic mapping values ​​of crow i and crow j, respectively. j,t Let be the food hiding place of crow j that is tracked in the t-th iteration, which is the memory of crow j in the t-th iteration;

[0040] S34: Convert the new position of each crow into a binary value position;

[0041] The conversion function is defined as follows:

[0042]

[0043] Where d is the average value of the air quality data after preprocessing in step S1;

[0044] The following formula converts the new position of each crow into a binary position value.

[0045]

[0046] Among them, bp i,t+1 Let be the binary value position of crow i in the (t+1)th iteration;

[0047] S35: Calculate the binary position (bp) of each crow. i,t+1 fitness value;

[0048] S36: Update the memory of each crow;

[0049] The memory of each crow is updated using the following formula:

[0050]

[0051] Where, m i,t+1 Let m be the memory value of crow i in the (t+1)th iteration. i,t Let bp be the memory value of crow i in the t-th iteration. i,t Let f(bp) be the binary position of crow i in the t-th iteration. i,t Let f(bp) be the fitness value of crow i at its binary position in the t-th iteration. i,t+1 Let be the fitness value of crow i at its binary position in the (t+1)th iteration;

[0052] S37: Repeat S33 to S36 until the maximum number of iterations is reached;

[0053] S38: The final position of each crow after the above steps is the selected air quality data. Based on the original subset, it is recombined into a generalized array to obtain the final optimized air quality dataset X used for prediction.

[0054] Furthermore, in step S35, the fitness function is set as follows:

[0055] f(p)=α*γR(p)+(1-α)*|p| / |N|

[0056] Where γR(p) represents the classification error rate of p, |p| represents the number of features selected in p, |N| represents the number of features in the entire air quality dataset, and α is a constant between 0 and 1.

[0057] The beneficial effects of this invention are:

[0058] (1) The preprocessed air quality dataset is decomposed into multiple independent intrinsic mode function (IMF) subsets and a residual subset by using the adaptive noise complete set empirical mode decomposition (CEEMDAN) method, which enhances the temporal characteristics of the air quality data while eliminating noise.

[0059] (2) The improved Chaotic Binary Raven Search (CBCSA) algorithm was used to perform feature selection on the aforementioned Intrinsic Mode Function (IMF) subset and residual subset. The subset that is favorable for prediction was selected and reconstructed, which further optimized the air quality data and accelerated the convergence speed of the prediction.

[0060] (3) A deep BLSTM model with self-attention mechanism is used to learn and predict air quality, which effectively utilizes the hidden bidirectional time series relationship in the air quality data, thereby further improving the accuracy of prediction. Attached Figure Description

[0061] Figure 1 This is a flowchart of a method according to an embodiment of the present invention.

[0062] Figure 2 This is a diagram of the deep learning BALSTM model constructed in the air quality prediction method of this invention.

[0063] Figure 3 This is a comparison chart of the predicted AQI values ​​for the next 24 hours and the actual AQI values ​​using the method of this invention. Detailed Implementation

[0064] To better understand the above-described objects, features, and advantages of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the invention; however, the invention may be practiced in other ways different from those described herein, and therefore, the invention is not limited to the specific embodiments disclosed below.

[0065] like Figure 1 As shown, the air quality prediction method based on a hybrid deep learning model in this embodiment includes the following steps:

[0066] S1. Construct an air quality prediction dataset and preprocess the air quality data;

[0067] The data collected is the raw air quality data for the city to be predicted, i.e. hourly air quality data, including data on six air pollutants: PM2.5, PM10, SO2, NO2, O3, and CO, as well as the Air Quality Index (AQI).

[0068] Seven air parameter arrays, consisting of data on the six air pollutants (PM2.5, PM10, SO2, NO2, O3, and CO) and AQI, were created as the initial dataset for air quality prediction. The initial air quality dataset was then preprocessed sequentially, including missing value processing, seasonality processing, and standardization processing.

[0069] The missing value handling method is as follows: for the missing air quality data, the average value of the original air quality data is used to fill in the missing parts;

[0070] Seasonal processing is as follows: Define a seasonal index, and calculate it using the following formula:

[0071]

[0072] Where Se represents the seasonal factor, and seA iThis represents the average of the data for each season in the collected raw air quality data over the years, where i = 1, 2, 3, 4, representing the four seasons of spring, summer, autumn, and winter respectively, and yearA represents the average value of the collected raw air quality data.

[0073] The specific processing method is as follows: first, divide the original air quality data by the corresponding seasonal factor so that it can be used for unified prediction later.

[0074] The standardization process is as follows: First, calculate the mean (yearA) and standard deviation (σ) of all raw air quality data, and then perform Z-score standardization on them, using the following formula:

[0075] x=(x0-yearA) / σ

[0076] Where x0 represents the air quality data before any standardization process, and x represents the air quality data after standardization process.

[0077] S2. The preprocessed air quality dataset is decomposed into multiple independent intrinsic mode function (IMF) subsets and a residual subset using the adaptive noise complete set empirical mode decomposition (CEEMDAN) method, which enhances the temporal characteristics of the air quality data while eliminating noise.

[0078] S21: Set the total number of intrinsic mode functions to be decomposed, maxImf;

[0079] S22: Add adaptive noise to the seven air parameter arrays of the air quality dataset after step S1, and the resulting new dataset is denoted as X;

[0080] S23: Find all local maxima and minima in X, and fit them with cubic spline interpolation functions to form the upper and lower envelopes of X. Subtract the mean of the upper and lower envelopes from X to obtain a new dataset. If there are no negative local maxima or positive local minima in the new dataset, terminate the data processing; otherwise, continue the above process until there are no negative local maxima or positive local minima. The dataset after this decomposition is denoted as the Intrinsic Mode Function dataset IMF1.

[0081] S24: Subtract the extracted dataset IMF1 from X to obtain the remaining dataset RES1: RES1 = X - IMF1;

[0082] S25: Treat RES1 as X and repeat steps S23 to S24 until the number of decompositions reaches maxImf, thus obtaining a sequence of intrinsic mode function datasets IMF. i The dataset RES of sum items i, i = 1, 2, ..., maxImf; the last remainder dataset is the residual dataset;

[0083] S26: Combine all the above intrinsic mode function datasets (IMF) i Combined with the residual dataset, they form a generalized array, which is the air quality dataset X' decomposed using the CEEMDAN method.

[0084] S3. The improved Chaotic Binary Raven Search (CBCSA) algorithm is used to perform feature selection on the aforementioned Intrinsic Mode Function (IMF) subset and residual subset, and the subset that is favorable for prediction is selected and reconstructed.

[0085] The Chaotic Binary Raven Search (CBCSA) algorithm is used to perform feature selection on the aforementioned Intrinsic Mode Function (IMF) subset and residual subset, selecting subsets that are beneficial for prediction. These subsets are then combined into a generalized array to serve as the optimized air quality dataset. The specific process is as follows:

[0086] S31: Initialize the chaotic binary crow search algorithm;

[0087] Set the number of crows n and the maximum perception probability AP. max Minimum perceptual probability AP min Maximum flight distance fl max Minimum flight distance fl min The parameters such as the maximum number of iterations tmax are set, and the chaotic mapping value of each crow is created using the Logistic chaotic mapping function;

[0088] S32: Use each air quality data point in the decomposed air quality dataset X' output in step S2 as the initial position p of each crow. i and initial memory m i , i = 1, 2, ..., n;

[0089] S33: Update the position of each crow;

[0090] In conventional chaotic binary crow search algorithms, the flight distance and perception probability of each crow are set to constants, which is detrimental to expanding the search range of crows and accelerating the iterative convergence of the algorithm. Therefore, this invention dynamically updates the perception probability and flight distance of each crow in each iteration using the following formula:

[0091]

[0092]

[0093] Where t is the iteration round, tmax is the maximum number of iterations, and AP tLet fl be the probability of the crow's perception in the t-th iteration. t Let be the distance the crow flies in the t-th iteration.

[0094] Then update the position of each crow using the following formula.

[0095]

[0096] Where, p i,t+1 Let p be the position of crow i in the (t+1)th iteration. i,t Let C be the position of crow i in the t-th iteration. i and C j Let m be the chaotic mapping values ​​of crow i and crow j, respectively. j,t Let be the food hiding place of crow j that is tracked in the t-th iteration, which is the memory of crow j in the t-th iteration.

[0097] S34: Convert the new position of each crow into a binary value position;

[0098] The conversion function is defined as follows:

[0099]

[0100] Where d is the average value of the air quality data after preprocessing in step S1.

[0101] The following formula converts the new position of each crow into a binary position value.

[0102]

[0103] Among them, bp i,t+1 Let be the binary value position of crow i in the (t+1)th iteration;

[0104] S35: Calculate the new binary position bp for each crow. i,t+1 The fitness value; the fitness function is set as follows:

[0105] f(p)=α*γR(p)+(1-α)*|p| / |N|

[0106] Where γR(p) represents the classification error rate of p, |p| represents the number of features selected in p, |N| represents the number of features in the entire air quality dataset, and α is a constant between 0 and 1. In this embodiment, α is set to 0.5.

[0107] S36: Update the memory of each crow;

[0108] Update each crow's memory using the following formula.

[0109]

[0110] Where, m i,t+1 Let m be the memory value of crow i in the (t+1)th iteration. i,t Let bp be the memory value of crow i in the t-th iteration. i,t Let f(bp) be the binary position of crow i in the t-th iteration. i,t Let f(bp) be the fitness value of crow i at its binary position in the t-th iteration. i,t+1 Let be the fitness value of crow i at its binary position in the (t+1)th iteration.

[0111] S37: Repeat S33 to S36 until the maximum number of iterations is reached.

[0112] S38: The final position of each crow after the above steps is the selected air quality data. Based on the original subset, it is recombined into a generalized array (that is, the selected air quality data that were originally in one subset are still assigned together to form a new subset, and then the new subsets are combined into a generalized array) to obtain the final optimized air quality dataset X used for prediction.

[0113] S4. Constructing a deep learning model, BALSTM

[0114] This invention constructs a deep learning model, BALSTM, by cascading a three-layer bidirectional long short-term memory neural network (BLSTM) with a self-attention mechanism. The model incorporates Dropout in the first two BLSTM layers to reduce overfitting and selects Sigmoid as the activation function for the self-attention layer. The final result is then fed into a fully connected layer as the output of the final prediction. The BALSTM model is as follows: Figure 2 As shown.

[0115] S5. Input the optimized air quality dataset X” obtained after feature selection in step S3 into the deep learning BALSTM model constructed in step S4 for learning and prediction to obtain the final air quality prediction result. After prediction, multiply the air quality prediction result by the seasonal factor Se to obtain the true air quality prediction value.

[0116] Experimental verification:

[0117] 1.1) This embodiment selects hourly air quality data in Beijing from January 1, 2020 to December 31, 2020 as experimental data, including data on six air pollutants: PM2.5, PM10, SO2, NO2, O3 and CO, and AQI.

[0118] 1.2) The relevant parameter settings of the proposed model are as follows: the maximum perception probability is set to 0.5, the minimum perception probability is 0.1, the maximum flight distance is 4, the minimum flight distance is 2, the maximum number of iterations of the improved CBCSA is 50, the chaotic mapping is the Logistic mapping function, the number of neurons in the three layers of BALSTM is set to 64, 64 and 32 respectively, the number of training iterations is 2000, the number of batches is 256, the step size is 24, the learning rate is set to decrease from 0.1, and the minimum learning rate is 0.0000001.

[0119] 1.3) The selected air quality data is preprocessed, decomposed, and optimized using the improved Chaotic Binary Raven Search (CBCSA) algorithm according to the method of this invention, and then input into the deep learning model BALSTM for learning and prediction. The comparison between the predicted AQI results for the next 24 hours and the actual AQI values ​​is shown below. Figure 3 As shown in the figure, the predicted value curve fits the true value curve very well, indicating a high prediction accuracy.

[0120] 1.4) Compare the improved CBCSA algorithm in the method of this invention with the original CBCSA. Let epoch = 2000. The running time results are shown in Table 1.

[0121] Table 1 Comparison of runtime between the improved CBCSA and the standard CBCSA

[0122]

[0123] It can be seen that the improved CBCSA calculation speed by the method of the present invention is increased by 7.82%.

[0124] The air quality prediction method of this invention eliminates adverse factors such as missing data, seasonal effects, and noise in air quality data, and effectively utilizes the hidden two-way time series relationship in air quality data, thereby enhancing the accuracy and convergence speed of air quality prediction.

Claims

1. An air quality prediction method based on a hybrid deep learning model, characterized in that, Includes the following steps: S1. Construct an air quality prediction dataset and preprocess the air quality data; collect raw air quality data for the cities to be predicted, including data on six air pollutants (PM2.5, PM10, SO2, NO2, O3, CO) and the Air Quality Index (AQI); create seven air parameter arrays consisting of the above six air pollutant data and the AQI as the initial dataset for air quality prediction; preprocess the initial air quality dataset by handling missing values, seasonality, and standardization. S2. The preprocessed air quality dataset is decomposed into multiple independent intrinsic mode function (IMF) subsets and a residual subset using the adaptive noise complete set empirical mode decomposition (CEEMDAN) method. S3. The improved Chaotic Binary Raven Search (CBCSA) algorithm is used to perform feature selection on the aforementioned Intrinsic Mode Function (IMF) subset and residual subset. The subset that is favorable for prediction is selected and reconstructed to obtain the optimized air quality dataset. The specific process is as follows: S31: Initialize the chaotic binary crow search algorithm; Set the number of crows n and the maximum perception probability. Minimum perceptual probability Maximum flight distance Minimum flight distance The maximum number of iterations tmax parameter is determined, and the chaotic mapping value for each crow is created using the Logistic chaotic mapping function; S32: Use each air quality data point in the decomposed air quality dataset X' output in step S2 as the initial position p of each crow. i and initial memory m i , i=1,2,…,n; S33: Update the position of each crow; In each iteration, the perception probability and flight distance of each crow are dynamically updated using the following formula: ， ， Where t is the iteration round, The maximum number of iterations, Let be the probability of the crow's perception in the t-th iteration. Let be the distance the crow flies in the t-th iteration; Then update the position of each crow using the following formula: ， in, Let i be the position of crow i in the (t+1)th iteration. Let C be the position of crow i in the t-th iteration. i and C j Let be the chaotic mapping values ​​of crow i and crow j, respectively. Let be the food hiding place of crow j that is tracked in the t-th iteration, which is the memory of crow j in the t-th iteration; S34: Convert the new position of each crow into a binary value position; The conversion function is defined as follows: ， Where d is the average value of the air quality data after preprocessing in step S1; The new position of each crow is converted into a binary value using the following formula: ， in, Let be the binary value position of crow i in the (t+1)th iteration; S35: Calculate the binary position of each crow fitness value; S36: Update the memory of each crow; that is, update the memory of each crow using the following formula. ， in, Let i be the memory value of crow i in the (t+1)th iteration. Let i be the memory value of crow i in the t-th iteration. Let i be the binary position of crow i in the t-th iteration. Let be the fitness value of crow i at its binary position in the t-th iteration. Let be the fitness value of crow i at its binary position in the (t+1)th iteration; S37: Repeat S33 to S36 until the maximum number of iterations is reached; S38: The final position of each crow after the above steps is the filtered air quality data. Based on the original subset, it is recombined into a generalized array to obtain the final optimized air quality dataset X'' used for prediction. S4. A deep learning BALSM model is constructed by using a three-layer bidirectional long short-term memory neural network (BLSTM) and a self-attention mechanism in series. The model adds the Dropout method to the first two BLSTM layers to reduce the overfitting problem and selects Sigmoid as the activation function of the self-attention mechanism layer. The final result is input into the fully connected layer as the output of the final prediction result. S5. Input the optimized air quality dataset obtained in step S3 into the deep learning BALSTM model constructed in step S4 for learning and prediction, and obtain the final air quality prediction result.

2. The air quality prediction method based on a hybrid deep learning model according to claim 1, characterized in that, In step S1, the missing value handling method is as follows: for the missing air quality data, the average value of the original air quality data is used to fill in the missing data. Seasonal processing includes defining a seasonal index, calculated using the following formula: ， Where Se represents the seasonal factor, and seA i This represents the average of the data for each season in the collected raw air quality data over the years, i=1,2,3,4, representing the four seasons of spring, summer, autumn and winter respectively, and yearA represents the average value of the collected raw air quality data. The standardization process is as follows: First, calculate the mean (yearA) and standard deviation (σ) of all raw air quality data, and then perform Z-score standardization on them. The calculation formula is as follows: ， Where x0 represents the air quality data before any standardization process, and x represents the air quality data after standardization process.

3. The air quality prediction method based on a hybrid deep learning model according to claim 1, characterized in that, The specific process of step S2 is as follows: S21: Set the total number of intrinsic mode functions to be decomposed, maxImf; S22: Add adaptive noise to the seven air parameter arrays of the air quality dataset after step S1, and the resulting new dataset is denoted as X; S23: Find all local maxima and minima in X, and fit them with cubic spline interpolation functions to form the upper and lower envelopes of X. Subtract the mean of the upper and lower envelopes from X to obtain a new dataset. If there are no negative local maxima or positive local minima in the new dataset, terminate the data processing; otherwise, continue the above process until there are no negative local maxima or positive local minima. The dataset after this decomposition is denoted as the Intrinsic Mode Function dataset IMF1. S24: Subtract the extracted dataset IMF1 from X to obtain the remaining dataset RES1: RES1 = X - IMF1; S25: Treat RES1 as X and repeat steps S23-S24 until the number of decompositions reaches maxImf, thus obtaining a sequence of intrinsic mode function datasets IMF. i The dataset RES of sum items i , i=1, 2, …, maxImf; the last remainder dataset is the residual dataset; S26: Combine all the above intrinsic mode function datasets (IMF) i Combined with the residual dataset, they form a generalized array, which is the air quality dataset X' decomposed using the CEEMDAN method.

Images