Big Data primarily uses the Machine Learning method and Natural Language Processing to process and mine data content from the Internet. A large amount of real-time, multi-source data is conducive to depicting reality from different perspectives to obtain the most realistic description, laying a data source foundation for the application of AI. With sufficient data sources, AI can achieve continuous learning, optimization, and practical applications. As one of the most important branches of environmental monitoring data, air quality is closely related to people's work, life, and physical health. The application of Big Data to monitor air quality in the living environment can comprehensively analyze meteorological data and combine the relationship between environmental protection and ecological civilization construction. The Big Data technology can conduct an in-depth analysis of the root causes of environmental problems and integrate environmental indicators and emission information of environmental pollution sources. After scientific analysis, the emission intensity, pollution source distribution, and impact on the surrounding environment of each enterprise can be analyzed to formulate a scientific and environmentally friendly governance plan.

In addition, applying Big Data to air quality monitoring can effectively improve the ability of early ecological warning. Environmental monitoring and governance refer to the use of professional equipment to detect the content and emissions of different harmful substances in the environment to track the changing trend in air quality. Environmental monitoring platforms and algorithms can be used to comprehensively collect, quickly process, and analyze environmental data to improve the efficiency of environmental governance. Fig. 1 reveals the process of using Big Data to process environmental information.

Fig. 1.

Process procedures of the environmental monitoring information via Big Data.

(0.3MB).

According to Fig. 1, obtaining high-accuracy environmental data is the premise for practical innovation decision-making analysis. In practice, Big Data should collect, process, and analyze the environmental monitoring data. Finally, countermeasures, suggestions, and predictions are presented according to the analysis results. The Big Data environment analysis system's critical step is collecting environmental information. For example, it is essential to distinguish the monitoring range of video monitors and sensors installed at specific deployment points and determine whether natural conditions rather than manufactured issues cause environmental information.

Big Data effectively analyzes massive environmental information, mines valuable environmental information, and finally displays it dynamically through tools to achieve environmental protection and governance in line with the actual situation. Moreover, Big Data has a fast data processing speed, improving the timeliness of various early warnings and effectively predicting various pollution events to take specific preventive measures in advance. Big Data technology for environmental monitoring can continuously obtain all-around environmental monitoring-related data, enhancing the data's scalability and increasing the data's value for a perfect presentation of the analysis results. At the same time, the real-time results can provide strong data support for decision-making in environmental governance, monitoring, and other operations. In addition, Big Data technology can build a real-time environmental monitoring model to accurately monitor environmental data to implement timely and effective prevention and management. These measures play a critical part in saving management costs and improving the decision-making power of environmental management.

Big Data is a problem-oriented approach to analyzing the correlation between things. It uses relevant data analysis to comprehensively describe things with data association and data trends. The error and ambiguity of environmental information will not contract the analysis accuracy of Big Data. Another function of data is to predict trends. Qualitative research is common in the existing management concepts. After the emergence of data analysis, the judgment and expectation of decision-making can be realized through quantitative analysis methods because Big Data provides various data processing tools and algorithms. Compared with traditional environmental protection methods, combined with Big Data technology, environmental governance can comprehensively collect data and improve data transparency. Besides, various environmental monitoring indicators can be disclosed to the public through proper channels after being analyzed by Big Data, improving the authenticity and reliability of the data. AI algorithms can establish air quality monitoring models. Combining the two can realize the early warning of air quality. In response to the problems and deficiencies in the current air quality forecast and atmospheric environmental impact assessment, this work introduces various advanced technologies of AI, improves existing methods, and proposes new algorithms to support system modeling for air quality innovative decision-making methods.

As the core of AI, Machine Learning can simulate human behavior through a computer and improve the model's performance. Machine Learning performs well in regression tasks and is regarded as a favorable tool for pollutant concentration prediction due to its excellent accuracy and fault tolerance. The widely recognized statistical prediction belongs to the category of Machine Learning (Chui et al., 2021). However, data missing is a common problem in air pollutant concentration monitoring, which will destroy the integrity of data and affect the effect of data mining. Therefore, data pre-processing is critical for air quality modeling. This work constructed an innovative combined model based on Machine Learning for air quality forecast, named Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise-Whale Optimization Algorithm-Extreme Learning Machine (ICEEMDAN-WOA-ELM). This model consists of three modules: pre-processing, forecasting, and assessment. Fig. 2 displays the model structure.

Fig. 2.

Air quality forecast model based on Machine Learning.

(0.45MB).

It can be seen from Fig. 2 that the model uses the autoregressive method to predict pollutant concentration. Its calculation is simple. The signal decomposition and swarm intelligence algorithm are combined to improve the prediction accuracy of the model. The pre-processing module uses a cubic spline to interpolate global segments. The signal decomposition part of the forecast module adopts a fully adaptive method, Ensemble Empirical Mode Decomposition, to optimize the residual noise and pseudo-modal problems. The decomposition results have a little noise and abundant physical meaning. The network optimization part adopts the Whale Optimization Algorithm (WOA). It simulates the hunting behavior of humpback whales and performs well in exploring, utilizing, and avoiding the global optimum. Humpback whales can identify and surround prey. Because the position of the optimal solution in the search space is not a priori, WOA assumes that the current optimal candidate solution is the target prey. After the optimal search agent is determined, other search agents update their positions according to the optimal search agent. WOA simulates the feeding strategy of the spiral bubble network for performance optimization. Tests on mathematical optimization and structural engineering problems show that WOA has an excellent performance in exploring, exploiting, avoiding local optima, and converging.

Feedforward Neural Networks typically have low learning rates because they primarily use a slow gradient-based algorithm for training through which all parameters need to be adjusted iteratively. Therefore, it cannot meet the practical needs, limiting its practical application. The Extreme Learning Machine (ELM) with a single hidden layer can randomly select the hidden layer nodes and the output layer weights and has a good generalization ability, which has received extensive attention. N independent samples (xi,ti) can be expressed as:

Then, the network can be expressed as Eq. (7).

In Eq. (7), wi represents the weight vector between the input layer neurons and the ith hidden layer neurons; bi denotes the threshold of the ith hidden layer neuron; bi signifies the activation function;βi indicates the weight vector between the output layer neurons and the ith hidden layer neurons. Besides, Eq. (8) is workable.

In Eq. (8), H stands for the hidden layer's output matrix; β represents the weight vector between the neurons in the output layer and in the hidden layer; T refers to the desired network output. The evaluation module adopts five general indicators to evaluate the performance of the model: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), Theil Inequality Coefficient (TIC), and Correlation Coefficient (r), which are calculated according to:

Therefore, in view of the problems of insufficient accuracy and high computational cost of the current air quality forecasting methods, the Machine Learning method is introduced to establish a combined model of air quality forecasting Machine Learning, ICEEMDAN-WOA-ELM. It performs forecasting through the autoregressive method of pollutant concentration. This model improves prediction accuracy by combining Signal decomposition and swarm intelligence algorithms while maintaining simple calculation.

In recent years, air pollution has been severe in all provinces, especially during the heating season in Xi'an and its surrounding areas. Therefore, the daily average concentration data of six conventional air pollutants, PM2.5, PM10, NO2, SO2, CO, and O3 in Xi'an from September 2019 to September 2021, were selected as the experimental dataset. The data from September 2019 to July 2021 was used as the training set, and the data from August and September 2021 was taken as the test set. The experimental environment selected the Intel Core i7 processor, 8GB memory, and the language chose Python 3.5 to conduct the simulation experiment of the ICEEMDAN-WOA-ELM model. The experimental parameters were set as follows: the maximum number of iterations of the ICEEMDAN model is 1000, the maximum number of iterations of WOA is 200, and the number of search agents is 10.

While Machine Learning has achieved some results in air quality decision analysis and prediction, the advent of Deep Learning has brought Machine Learning closer to AI. Deep Learning can train Deep Neural Networks, extract features, and transform, abstract, and process information. Thus, it has advantages in solving practical problems. Compared with traditional Machine Learning, Deep Learning strengthens the extraction of features and transforms the features of the original space into a new feature space through layer-by-layer transformation, making regression, classification, etc., easier to achieve. The application of Big Data can mine effective information from samples. It also provides the basis for the study of time series forecasting. In this paper, a Deep Learning-based air quality prediction model is innovatively established by combining atmospheric theory and Deep Learning methods, namely Time-Space-Type-Meteorology (TSTM). It consists of three modules: feature engineering, forecasting, and performance evaluation. Fig. 3 reveals this innovative air quality decision method.

Fig. 3.

Structure of the combined air quality forecast model based on Deep Learning.

(0.54MB).

It can be found from Fig. 3 that the feature engineering module of the combined TSTM model includes data pre-processing, feature selection, and feature building modules. The Expectation Maximization (EM) algorithm is adopted for data pre-processing to avoid missing data. The input data needs to be normalized to eliminate the magnitude difference of different features and improve the accuracy and speed of the model. Here, the Min-Max Normalization algorithm is used. The output data is de-normalized for model evaluation. Six representative air pollutant concentrations were selected as the features based on atmospheric knowledge: O3, CO, SO2, NO2, PM10, and PM2.5, as well as the factors affecting meteorology: wind speed, temperature, humidity, and precipitation. Besides, the forecast time lag was set to 24h because the hourly concentration of air pollutants presents a significant diurnal variation (24h).

The forecast module adopts the ConvLSTM composed of the Long Short-Term Memory (LSTM) network and the CNN. It has the advantages of LSTM in time series processing and retains the feature extraction ability of CNNs. CNN usually consists of five parts: the output layer, fully connected layer, pooling layer, convolution layer, and input layer. Fig. 4 displays the structure.

Fig. 4.

Model architecture of CNN.

(0.14MB).

The calculation methods of convolution and pooling are as follows:

Fig. 5 reveals the structure of the LSTM network.

Fig. 5.

Model architecture of the LSTM network.

(0.22MB).

The performance evaluation module adopts the concentration evaluation indicators of three air pollutants: Normalized Mean Bias (NMB), RMSE, and Correlation Coefficient r. NMB is calculated according to Eq. (22).

In Eq. (22), N stands for the number of samples; Fiand Oi represent the actual value and the predicted value of the ith sample, respectively; F¯ and O¯ signify the average value of the predicted value and the actual value, respectively.

Eq. (23) indicates the calculation method of the range forecast accuracy of the Air Quality Index (AQI).

In Eq. (23), nAQI refers to the number of samples for which the range forecast of the AQI is accurate, and N is the total number of samples. The forecast accuracy AAQI_level of the AQL is calculated via Eq. (24).

In Eq. (24), nAQI_level represents the number of samples with an accurate forecast of the AQL, and N stands for the total number of samples.

The accuracy rate Acp of primary pollutant forecast is calculated according to Eq. (25).

In Eq. (25), ncp represents the number of samples with accurate forecasts in the evaluation time period, and N refers to the number of samples of the AQL ≥2 at the time.

For heavy pollution, when the AQI value is above 200, the forecast accuracy HAAQI_level of the AQL is calculated according to Eq. (26).

In Eq. (26), nAQI_level represents the number of samples with accurate forecasts in the evaluation time period, and NOH represents the actual number of weather samples with moderate and severe pollution.

Experiments were carried out on the above model, and 15 cities involved in Shaanxi Province were selected as the research objects. The data set collected data on six air pollutants and four meteorological elements from December 2019 to February 2021. The data from January to January 2020 was used as the training set, the data from February 2021 was used as test set 1, and the data from June 2021 was added as test set 2 to test the generalization ability of the model. The Big Data method is used to model the regional multi-step forecast of the hourly concentration of conventional air pollutants. Besides, the performance of the Deep Learning model reported here is compared with other benchmark models. During the experiment, the prediction lag is 24, the training Epoch is 100, and the training Batch Size is set to 24.

Fig. 6 presents the evaluation results of the ICEEMDAN-WOA-ELM model and the other six benchmark models on the data set collected here.

Fig. 6.

Comparison of forecast performance of daily average concentration of conventional air pollutants in Xi'an (a. PM2.5; b. PM10; c. NO2; d. SO2; e. CO; f. O3).

(1.48MB).

As can be seen from Fig. 6, the classical time-series Autoregressive Integrated Moving Average (ARIMA) model (Piccolo, 1990) is superior in performance, even better than the single AI model. The comparison of ELM (Huang et al., 2006), Genetic Algorithm-Extreme Learning Machine (GA-ELM) (Krishnan & Kamath, 2019), and WOA-ELM (Li et al., 2019) suggests that the prediction indicators of air pollutants by the network have improved after the swarm intelligence algorithm optimizes the network. In addition, the optimization effect of WOA is generally better than the Genetic Algorithm (GA) (Whitley, 1994). This is because WOA can avoid falling into local optimum, which is more flexible and efficient, and the convergence speed is faster. The overall performance of Ensemble Empirical Mode Decomposition- Whale Optimization Algorithm-Extreme Learning Machine (EEMD-WOA-ELM) and ICEEMDAN-WOA-ELM are better than other models.

EEMD is a noise-assisted data analysis method. The added noise cannot be completely neutralized, and there is residual noise. The ICEEMDAN model further handles the problem of residual noise, so it performs best in prediction accuracy. For example, when predicting O3, under the same prediction method, the values of the five evaluation indicators of ICEEMDAN-WOA-ELM are 14.187, 17.235, 0.140, 0.067, and 0.946, which are higher than those of EEMD-WOA-ELM of 21.157, 24.596, 0.219, 0.094, and 0.885. The experimental results indicates the excellent robustness of the ICEEMDAN-WOA-ELM model. In other words, it maintains high accuracy in the face of complex environments and is competent for forecasting different environments and pollutants. The results are better than other benchmark models, and the signal decomposition algorithm and optimization algorithm can significantly improve the prediction performance of the neural network. It can be seen that a single AI model cannot meet the requirements of air quality forecasting. Meanwhile, the combined model needs to be used to exert their respective advantages to improve the overall forecasting performance.

Fig. 7 provides the multi-step prediction results of the above Deep Learning air quality forecast model, TSTM, on different pollutants, steps, and test sets in 15 cities in the study area.

Fig. 7.

The standard deviation of TSTM's prediction effect of pollutant concentration in 15 cities around Xi'an (a. CO; b. SO2; c. NO2; d. O3; e. PM2.5; f. PM10).

(1.41MB).

Fig. 7 indicates a positive correlation between the predicted and actual values of the TSTM model. Moreover, the performance results of TSTM are very close to the forecast effect of different cities, and the performance is relatively stable. Unlike single-step forecasting, multi-step forecasting uses the same model and input to perform multiple outputs without needing to build additional models or wait for the previous forecast results. However, it is more complicated than single-step forecasting, and the error is usually larger. The experimental results suggest that TSTM has good robustness and generalization ability. Because the forecasting effects of different cities are similar, Xi'an is selected as the representative city for research. Xi'an's daily air pollution level is also more severe than surrounding cities, especially in winter. Fig. 8 presents the comparative analysis of air quality forecasts in Xi'an under the same conditions as the traditional Radial Basis Function (RBF) model (Musavi et al., 1992), the Deep Learning model Deep Belief Network (DBN) (Chen et al., 2015), Elman, and the TSTM model reported here.

Fig. 8.

Forecast performance evaluation of four models under heavy pollution weather in Xi'an (a. TSTM; b. RBF; c. DBN; d. Elman).

(1.31MB).

According to Fig. 8, the prediction accuracy of RBF and DBN for AQI range and air quality level is similar, but the prediction effect of DBN for primary pollutants is poor. The single-step prediction accuracy of the three benchmark models is lower than 0.6, and the performance of TSTM is the best at 0.88. The multi-step forecast of TSTM adopts a multi-output strategy. In other words, TSTM obtains multiple outputs simultaneously based on the same model and input. Besides, it does not need to build an additional forecast model or wait for the forecast of the previous step as the input of the next step, so the performance is good. In addition, the four models show considerable differences in performance in extreme weather. Deep Learning has a higher forecast upper limit than traditional Machine Learning. The forecast accuracy of RBF for heavily polluted weather is lower than the other three Deep Learning models. Elman remains in second place, but the performance degrades significantly as the prediction step size increases, with a prediction accuracy of <0.6 at a step size of 3. The single-step accuracy and average of the TSTM model proposed here almost reach full marks in the weather forecast results of heavy pollution, with a maximum of 1.00. The performance also decreases as the step size increases but remains above 0.86. It can be seen that the heavily polluted weather contains more extreme values of gas content concentration, which is also the key to testing the model's performance.

Fig. 9 illustrates the Accuracy, Precision, Recall, and F1 value predicted by the four models under heavy pollution weather in Xi'an.

Fig. 9.

Evaluation of forecast Accuracy of four models under heavy pollution weather in Xi'an.

(0.52MB).

According to Fig. 9, TSTM has the best performance compared with the other three models. Among them, the prediction Accuracy of the TSTM model is 95.2%, and the Precision is 89.5%, which is at least 5.1% better than other models. It also outperforms the other three models in terms of Recall and F1 value. The results are consistent with the above research results, proving that the Deep Learning air quality prediction model performs better in predicting different pollutants in 15 cities in the study area. TSTM ranks first in various evaluation indicators for different pollutants, maintaining a high forecast accuracy.