Complex environment ozone concentration prediction method fusing selective state space and multi-scale frequency domain features

Abstract

This invention relates to a method for predicting ozone concentration in complex environments by integrating selective state-space and multi-scale frequency domain features, belonging to the field of air pollution prediction and intelligent environmental monitoring technology. Addressing the problems of high computational cost of traditional chemical transport models, insufficient capture of long-sequence dependencies by machine learning models, and inadequate multi-scale feature extraction, this invention constructs a multi-branch heterogeneous architecture, integrates a selective state-space model to achieve linear complexity long-time series modeling, combines fast Fourier transform frequency domain analysis to capture periodic patterns, employs a wavelet Kolmogorov-Arnold network to characterize nonlinear threshold effects, and utilizes dynamically fused gated adaptive weighted multi-branch outputs. In a field study in a basin, this method reduced the root mean square error of ozone concentration prediction to 11.1 μg·m³. ‑3 The coefficient of determination was improved to 0.912, significantly optimizing prediction accuracy, computational efficiency and model interpretability.

Description

Technical Field

[0001] This invention belongs to the field of atmospheric pollution prediction and intelligent environmental monitoring technology, and relates to a method for predicting complex environmental ozone concentration by integrating selective state space and multi-scale frequency domain characteristics. Background Technology

[0002] In recent years, ozone pollution has become increasingly prominent, becoming a key factor restricting the continuous improvement of air quality. As a typical secondary atmospheric pollutant, ozone's formation mechanism is extremely complex, influenced not only by the emissions of precursors such as volatile organic compounds (VOCs) and nitrogen oxides (NOx), but also by photochemical reaction rates, meteorological conditions, and complex inter-regional transport processes. In a certain basin with enclosed topography, high humidity, and variable meteorological conditions, ozone pollution exhibits a significant characteristic of both local accumulation and regional transport. Previous studies have shown that the diurnal ozone concentration in this basin is mainly driven by the synergistic effects of multiple levels of temperature, relative humidity, and geopotential height. Particularly on days with ozone exceedances, the descending of warm air caused by upper-level warming and the convergence and uplift of gases in the middle and lower atmosphere provide a favorable physical environment for ozone formation and accumulation.

[0003] Currently, three main methods are used domestically and internationally in the field of air pollutant prediction: Chemical Transport Models (CTM), statistical models, and machine learning models. Chemical transport models, such as the Community Multiscale Air Quality Model (CMAQ), predict air quality by numerically simulating atmospheric physical and chemical processes. However, they heavily rely on the accuracy of emission inventories and incur enormous computational costs, making them unsuitable for real-time, high-resolution prediction. Statistical models, such as Multiple Linear Regression (MLR), perform poorly when dealing with highly nonlinear chemical reaction mechanisms. While machine learning models based on gradient boosting decision trees, such as CatBoost (Categorical Boosting) and XGBoost (Extreme Gradient Boosting), demonstrate robust performance in handling nonlinear relationships and high-dimensional data, they still face technical bottlenecks when dealing with long-series atmospheric monitoring data exhibiting significant periodicity, trends, and sudden fluctuations. These bottlenecks include insufficient long-term capture capabilities, inadequate multi-scale feature mining, rigid nonlinear mapping mechanisms, and a lack of dynamism in feature interaction and weight allocation. Traditional recurrent neural networks (RNNs) face the vanishing gradient problem, while the self-attention mechanism of the Transformer architecture has a computational complexity that increases quadratically with the sequence length, resulting in low efficiency when processing extremely long historical sequences.

[0004] Therefore, in response to the above problems, there is an urgent need for a novel prediction method that can achieve a breakthrough in both accuracy and efficiency through physical heuristic structural design, so as to better address the challenges of ozone concentration prediction in complex environments. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a method for predicting ozone concentration in complex environments that integrates selective state space and multi-scale frequency domain features, aiming to achieve a breakthrough in both accuracy and efficiency through a physics-inspired structural design.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for predicting complex environmental ozone concentrations by integrating selective state space and multi-scale frequency domain features includes the following steps: S1, data preprocessing step, uses a learnable feature weight layer to label the feature importance of the input historical meteorological and pollutant sequences, and then uses RMS normalization (Root Mean Square Normalization, RMSNorm) for normalization. S2, the multi-branch feature extraction step, inputs the preprocessed data in parallel to the Mamba time-domain sequence enhancement branch, the Fast Fourier Transform (FFT) frequency-domain feature extraction branch, the Wavelet Kolmogorov-Arnold Network (WavKAN) nonlinear fitting branch, the cross-feature interaction branch based on the Gated Linear Unit (GLU) and the original linear feature branch to extract multiple feature vectors; S3, Feature aggregation step, weighted concatenation of the multiple feature vectors through the ReZero mechanism, and recalibrated through the channel attention mechanism to generate an enhanced feature tensor; S4, Prediction Generation Step: The enhanced feature tensor is input into the depth probe and processed by linear transformation and SiLU activation function (Sigmoid Linear Unit, SiLU) to obtain the preliminary prediction result DeepPred; S5, Dynamic fusion step: The preliminary prediction result DeepPred is weighted and summed with the linear residual correction term LinearRes and the nonlinear residual correction term KANRes based on WavKAN through a dynamic fusion gating mechanism, and the final ozone concentration prediction value FinalPrediction is output.

[0007] Furthermore, in step S1, the learnable feature weight layer is determined by the formula... Feature importance is determined, among which X It is the input sequence. It is a learnable parameter vector. It is the Sigmoid activation function, and ⊙ represents element-wise multiplication.

[0008] Furthermore, in step S2, the Mamba time-domain sequence enhancement branch adopts a Selective State Space Model (SSM), and its discretization calculation process includes:

[0009]

[0010]

[0011]

[0012] in It is the step size parameter. A It is a state matrix. B It is the input matrix. C It is the output matrix. It is a moment t state, It is a moment t Input, It is a moment t The output.

[0013] Furthermore, in step S2, the FFT frequency domain feature extraction branch converts the signal to the frequency domain through fast Fourier transform, modulates the spectrum using complex weights, and retains the amplitude features by taking the absolute value.

[0014] Furthermore, in step S2, the WavKAN nonlinear fitting branch uses the Mexican hat wavelet as the basis function, and its activation function is... ,in x It is an input variable.

[0015] Furthermore, in step S2, the cross-feature interaction branch based on GLU includes two projection matrices, which are multiplied element-wise after nonlinear transformation of the input data and normalized by RMSNorm.

[0016] Furthermore, in step S3, the ReZero mechanism utilizes a learnable scalar. 1, 2, 3, 4, 5. Perform weighted concatenation of the feature vectors of each branch.

[0017] Furthermore, in step S4, the depth probe includes a linear layer, an RMSNorm layer, a SiLU activation function layer, and a Dropout layer.

[0018] Furthermore, in step S5, the dynamic fusion gating mechanism generates weights through a gating network. w 1, w 2, w 3, and through the formula

[0019] Perform a weighted summation, where w1, w2, and w3 are the normalized weights.

[0020] Furthermore, the input historical meteorological and pollutant sequences include temperature, humidity, pressure, geopotential height, wind field, and precursor concentration data.

[0021] The beneficial effects of this invention are as follows: (1) The traditional Transformer architecture suffers from computational complexity that increases quadratically with sequence length due to its self-attention mechanism, making it difficult to handle extremely long historical meteorological sequences. This invention introduces the selective state-space model Mamba, which, through its linear computational complexity characteristics, effectively captures the lag effects spanning several days in atmospheric chemical evolution while maintaining efficient memory usage. This design is particularly suitable for the accumulation process of ozone precursors and the inertial analysis of cross-regional transport, enabling the model to achieve accurate predictions over long windows even at environmental monitoring stations with limited computing resources.

[0022] (2) In view of the characteristics of atmospheric data, which include diurnal periodicity, seasonal trends, and instantaneous abrupt changes, this invention innovatively designs a fast Fourier transform frequency domain branch and a time domain processing branch in parallel. The frequency domain branch extracts the periodicity of meteorological factors through spectral modulation, effectively filtering out high-frequency sensor noise; the time domain branch focuses on capturing sudden pollution events such as sharp drops in air pressure. This time-frequency domain complementary mechanism significantly enhances the model's adaptability under non-stationary meteorological conditions such as a basin.

[0023] (3) By integrating the wavelet Kolmogorov-Arnold network WavKAN module, the Mexican hat wavelet basis function is used to replace the traditional fixed activation function. This design can accurately describe the threshold effect and discontinuous reaction process in ozone generation, such as the abrupt characteristics of light intensity-triggered reactions. At the same time, based on the mathematical foundation of the Kolmogorov-Arnold representation theorem, an analytical mapping relationship is formed between the model weights and physical variables, effectively breaking through the black box limitation of deep learning models.

[0024] (4) The designed dynamic fusion gating system can automatically adjust the branch weights according to the different pollution characteristics of Chengdu and Chongqing. In Chongqing's summer pollution days, which are dominated by temperature, the model increases the weight of the WavKAN branch to capture the thermodynamic nonlinear effects; in Chengdu, which is significantly affected by local circulation, the dynamic capture capability of the Mamba branch is enhanced. This adaptive characteristic ensures the universality of the single model under different geographical and climatic conditions.

[0025] (5) The structural design using channel attention mechanism and dual Shortcut connections effectively solves the common gradient vanishing problem in multi-branch neural networks. Combining residual connections and RMS normalization technology ensures stable gradient flow in deep networks, which not only accelerates model convergence speed but also significantly reduces the risk of overfitting. Experimental verification shows that this design improves training efficiency by about 30% while maintaining model accuracy.

[0026] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0027] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 The overall architecture of the Spectral-Mamba-KAN model; Figure 2 The distribution of stations in Sichuan and Chongqing. Detailed Implementation

[0028] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0029] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0030] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0031] This invention proposes an ozone concentration prediction model that integrates selective state space and multi-scale frequency domain features. Its overall architecture is as follows: Figure 1 As shown, this model employs a multi-branch heterogeneous feature extraction design, aiming to comprehensively learn the representation of the input historical meteorological and pollutant sequences.

[0032] I. Data Preprocessing and Feature Importance Labeling The model receives a three-dimensional tensor as input, whose dimensions are represented as (B, L, Din), where B is the batch size, L is the lookback window length, and Din is the input feature dimension, which includes multiple layers of temperature, humidity, pressure, potential height, wind field, and precursor concentration.

[0033] The input data first passes through a learnable feature weights layer:

[0034] in It is a learnable parameter vector. The Sigmoid activation function is used. This mechanism allows the model to automatically learn and strengthen features that contribute more to ozone concentration prediction (such as high temperature and low humidity) during training, while suppressing interference from noise dimensions. Subsequently, the data is normalized using RMSnorm to enhance numerical stability and then split into X segments for sequence stream processing. seq and X used for final slice processing last .

[0035] II. Multi-branch heterogeneous feature extraction architecture The core of this invention lies in five parallel feature extraction branches, which model the data from five dimensions: time-domain dependence, frequency-domain regularity, higher-order nonlinear space, cross-feature interaction, and original linear representation. 1. Mamba Temporal Sequence Enhancement Branch This branch focuses on capturing long-range dependencies in atmospheric chemical evolution. Data is fed into cascaded two-layer Mamba blocks, the core of which is a selective scan mechanism. Mamba uses parameters from a state-space model (SSM) to... B and C are set as input functions, enabling the model to selectively filter information. Its discretization calculation process follows:

[0036]

[0037]

[0038]

[0039] The branch parameter is set to d. model=64, d state =16, d conv =4. The processed features are then processed again through RMSnorm and weighted according to the temporal dimension using the temporal attention mechanism. The most predictive temporal representation attn_vec is extracted by using the last time step of the sequence as the query vector.

[0040] 2. FFT Frequency Domain Feature Extraction Branch To address the periodicity of atmospheric factors (such as diurnal variations and seasonal trends), this branch transforms the signal to the frequency domain using FFT. Within the frequency domain, the spectrum is modulated using complex weights, and then the absolute values ​​(Abs) are taken to preserve amplitude characteristics. A frequency domain feature vector, freq_vec, is generated through Linear / SiLU layer mapping. This design effectively filters out high-frequency sensor noise and captures a stable climate baseline.

[0041] 3. WavKAN nonlinear fitting branch This scheme integrates a lightweight wavelet Kolmogorov-Arnold network (WavKAN). Unlike traditional MLPs, WavKAN deploys learnable activation functions on the edges and uses Mexican Hat wavelets as basis functions. Its activation function formula is:

[0042] By adjusting the translation and scaling parameters, the model can accurately characterize the complex nonlinear threshold effects in ozone generation (such as the discontinuity of light intensity-triggered reactions) and generate spatial feature vectors feat_vec.

[0043] 4. InteractionBlock Cross-feature Interaction Branch To model the coupling relationships between different variables (such as the synergistic effect of temperature and humidity), an interaction block based on gated linear units (GLUs) was designed. This module contains two projection matrices, Projection A and Projection B, which are used to model the coupling relationships between different variables (such as the synergistic effect of temperature and humidity). last After performing a nonlinear transformation, element-wise multiplication is performed, and finally, the interaction feature vector interact_vec is generated by normalization through RMSNorm.

[0044] 5. Original linear feature branches By preserving the original linear mapping of the input, raw_vec is generated as a robust basis for the prediction results, ensuring that the model does not deviate from the basic prediction direction in extreme cases.

[0045] III. Feature Aggregation and Depth Probe The above five feature vectors are processed by the ReZero mechanism through a learnable scalar. 1, 2, 3, 4, 5. Weighted concatenation is performed to form an enhanced feature tensor with dimensions 5 × 64 = 320. This tensor then enters the channel attention mechanism, which includes: 1. AvgPool + MaxPool: Global average pooling and max pooling are used in parallel to capture channel statistics.

[0046] 2. MLP: Learns the dependencies between channels through a multilayer perceptron.

[0047] 3. Sigmoid: Generates channel weight coefficients.

[0048] 4. Element-wise Multiply: Recalibrate the channel dimension of each branch feature.

[0049] The recalibrated feature input Output Head uses a cascaded structure of Linear→RMSNorm→SiLU→Dropout→Linear to generate the preliminary prediction result DeepPred.

[0050] IV. Dynamic Fusion and Residual Correction The model incorporates a residual connection mechanism to enhance training stability. last Two correction terms, LinearRes and KANRes, are generated respectively through Linear Shortcut (linear basis) and KAN Shortcut (nonlinear basis). The final prediction result is determined by a Dynamic Fusion Gate. This gated network receives input features and generates three normalized weights through Linear→SiLU→Dropout→Linear→Softmax. w 1, w 2, w 3) The final predicted value is obtained by weighted summation:

[0051] DeepPred is the master prediction after multi-branch fusion, LinearRes is the linear shortcut path output, and KANRes is the nonlinear shortcut output based on WavKAN. This design allows the model to adaptively adjust its decision logic based on whether it is in a pollution outbreak phase (emphasizing Deep and KAN paths) or a stable phase (emphasizing Linear paths).

[0052] V. Preparation of Experimental Data The experiment utilized the NCEP GDAS / FNL 0.25° Global Tropospheric Analysis and Prediction Grid Dataset (UTC+00) from the National Center for Environmental Prediction (NCEP) and the European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5 post-processed daily statistical data on pressure levels from 1940 to the present (UTC+08). Various near-surface and upper-air meteorological data were selected as input variables (as shown in Table 1). The study focused on atmospheric dynamics analysis at three isobaric surfaces: 925 hPa, 850 hPa, and 500 hPa. The 500 hPa level significantly characterized the surface temperature field, precipitation systems, pressure distribution, and extreme weather events. Notably, the trough-ridge system formed in this atmospheric layer is closely related to large-scale circulation evolution: when a low-pressure trough forms in this layer, it easily triggers cooling, precipitation, and severe convective weather; conversely, the establishment of a high-pressure ridge corresponds to the control of a warm air mass, often leading to descending air and sunny, hot, and dry weather conditions. The key stratifications at the 925 hPa and 850 hPa levels, covering the area from the near-surface to the top of the boundary layer, can reveal relevant insights into boundary layer height meteorology.

[0053] Table 1. List of multidimensional meteorological parameters, emissions, and auxiliary factors for long-term series.

[0054] We calculated the daily average and daily maximum values ​​of several meteorological characteristics, where the daily maximum value is the maximum value of the three-hour average for the day. To characterize the ozone response to emissions, we used monitoring stations in the Sichuan-Chongqing region provided by the National Urban Air Quality Real-Time Release Platform, such as... Figure 2 As shown, red circles represent stations within Chongqing and Chengdu, while blue circles represent stations in other areas; long-term observation data (UTC+08) includes O3 and PM2.5. 2.5Long-term trend data for pollutants such as SO2, CO, and NO2 were collected, with O3 represented by the maximum 8-hour moving average and the remaining pollutants by daily averages. Considering the spatiotemporal variability of ozone, spatiotemporally related auxiliary variables were also introduced. The overall data study spanned from July 8, 2015 to December 31, 2020. During this period, some stations had a few missing values ​​within the entire period, and observations at a few stations did not cover the entire period. Therefore, KNN was used to ensure data continuity across the observation periods of each station. The data were standardized, with a lookback window of L=5, and the prediction target was the future daily ozone concentration at each station.

[0055] 5.1 Model Parameter Setting The model configuration strictly follows the actual settings in the code to ensure performance reproducibility: Table 2 Spectral-Mamba-KAN model parameter settings

[0056] 5.2 Performance Evaluation Indicators The coefficient of determination (R²), mean absolute error (MAE), and root mean square error (RMSE) were used as evaluation indicators.

[0057] 5.3 Comparison of Predictive Performance Results To verify the superiority of the present invention, it was compared and evaluated with the strongest benchmark model in the field of time series forecasting between 2022 and 2026 on a basin test set.

[0058] Table 3. Model Performance Comparison Results

[0059] Experimental results show that our model achieves leading performance across multiple metrics. Compared to the popular iTransformer model from 2024, our model reduces MAE by approximately 8.5% and RMSE by approximately 11.2%, while maintaining linear computational complexity and improving the capture rate of high-concentration ozone events by over 12%. This is mainly attributed to WavKAN's sensitive capture ability of local nonlinear abrupt changes and Mamba's efficient memory.

[0060] 5.4 Ablation Experiment To verify the contribution of each module, ablation experiments were conducted:

[0061] Ablation experiments further confirm that this invention is not a simple stacking of modules, but rather achieves a scientific balance in model depth and breadth through a physically inspired structural design (frequency domain stable substrate + KAN enhances local nonlinearity + Mamba captures time dependence), which can significantly improve the accuracy of atmospheric ozone concentration prediction in complex environments.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting complex environmental ozone concentration by integrating selective state space and multi-scale frequency domain features, characterized in that: Includes the following steps: S1, data preprocessing step, uses a learnable feature weight layer to label the feature importance of the input historical meteorological and pollutant sequences, and uses RMS normalization for normalization. S2, the multi-branch feature extraction step, inputs the preprocessed data in parallel to the Mamba time-domain sequence enhancement branch, the Fast Fourier Transform (FFT) frequency-domain feature extraction branch, the wavelet Kolmogorov-Arnold network (WavKAN) nonlinear fitting branch, the cross-feature interaction branch based on gated linear unit (GLU), and the original linear feature branch to extract multiple feature vectors. S3, Feature aggregation step, weighted concatenation of the multiple feature vectors through the ReZero mechanism, and recalibrated through the channel attention mechanism to generate an enhanced feature tensor; S4, Prediction Generation Step: Input the enhanced feature tensor into the depth probe and perform linear transformation and SiLU activation function processing to obtain the preliminary prediction result DeepPred; S5, Dynamic fusion step: The preliminary prediction result DeepPred is weighted and summed with the linear residual correction term LinearRes and the nonlinear residual correction term KANRes based on WavKAN through a dynamic fusion gating mechanism, and the final ozone concentration prediction value FinalPrediction is output.

2. The method for predicting complex environmental ozone concentration by fusing selective state space and multi-scale frequency domain features according to claim 1, characterized in that: In step S1, the learnable feature weight layer is obtained through the formula... Feature importance is determined, among which X It is the input sequence. It is a learnable parameter vector. It is the Sigmoid activation function, and ⊙ represents element-wise multiplication.

3. The method for predicting complex environmental ozone concentration by fusing selective state space and multi-scale frequency domain features according to claim 1, characterized in that: In step S2, the Mamba time-domain sequence enhancement branch adopts the Selective State-Space Model (SSM), and its discretization calculation process includes: in It is the step size parameter. A It is a state matrix. B It is the input matrix. C It is the output matrix. It is a moment t state, It is a moment t Input, It is a moment t The output.

4. The method for predicting complex environmental ozone concentration by fusing selective state space and multi-scale frequency domain features according to claim 1, characterized in that: In step S2, the FFT frequency domain feature extraction branch converts the signal to the frequency domain through Fast Fourier Transform, modulates the spectrum using complex weights, and retains the amplitude features by taking the absolute value.

5. The method for predicting complex environmental ozone concentration by fusing selective state space and multi-scale frequency domain features according to claim 1, characterized in that: In step S2, the WavKAN nonlinear fitting branch uses the Mexican hat wavelet as the basis function, and its activation function is: ,in x It is an input variable.

6. The method for predicting complex environmental ozone concentration by fusing selective state space and multi-scale frequency domain features according to claim 1, characterized in that: In step S2, the cross-feature interaction branch based on GLU includes two projection matrices, which are multiplied element-wise after nonlinear transformation of the input data and normalized by RMSNorm.

7. The method for predicting complex environmental ozone concentration by fusing selective state space and multi-scale frequency domain features according to claim 1, characterized in that: In step S3, the ReZero mechanism uses a learnable scalar. 1, 2, 3, 4, 5. Perform weighted concatenation of the feature vectors of each branch.

8. The method for predicting complex environmental ozone concentration by fusing selective state space and multi-scale frequency domain features according to claim 1, characterized in that: In step S4, the depth probe includes a linear layer, an RMSNorm layer, a SiLU activation function layer, and a Dropout layer.

9. The method for predicting complex environmental ozone concentration by fusing selective state space and multi-scale frequency domain features according to claim 1, characterized in that: In step S5, the dynamic fusion gating mechanism generates weights through a gating network. w 1, w 2, w 3, and through the formula Perform a weighted summation, where w1, w2, and w3 are the normalized weights.

10. The method for predicting complex environmental ozone concentration by fusing selective state space and multi-scale frequency domain features according to claim 1, characterized in that: The input historical meteorological and pollutant sequences include temperature, humidity, pressure, geopotential height, wind field, and precursor concentration data.

Images