Prediction of PM_2.5 concentrations in Malaysia using machine learning techniques: a review

Category 1: deep learning

Artificial neural networks, which are algorithms inspired by the structure and function of the brain, are used in deep learning. The papers that fall within this category are listed below.

Shahriar et al.⁶ conducted a study in three Bangladeshi air pollution hotspots to evaluate the effectiveness of two hybrid models for predicting daily PM_2.5 concentrations in terms of computational efficiency and accuracy. The models presented and compared with DT and Catboost were Autoregressive Integrated Moving Average (ARIMA)-Artificial Neural Network (ANN) and ARIMA-SVM. With lower MAE and RMSE and a better R² value, the CatBoost and ARIMA-ANN models fared well.

By comparing four prediction models, Yang et al.⁷ sought to anticipate PM several days in advance in 39 Seoul stations. The gated recurrent Unit (GRU) of a convolutional neural network (CNN) and the long short-term memory (LSTM) of a CNN were the two models suggested and compared. The CNN-LSTM model provided reliable prediction by capturing hidden patterns with low RMSE and MAE values. In Ref. 8 deep neural network (DNN) based hybrid model, DNN-LSTM, was proposed. The DNN-LSTM model outperformed the multiple additive regression trees (MART) and deep feedforward neural network (DFNN) models with a highest R² and lowest values of MAE and RMSE for 48-h predictions.

For the prediction of PM_2.5, Zhang et al.⁹ suggested a DL model including an auto-encoder (AE) and a bidirectional LSTM (Bi-LSTM). The proposed method incorporates data preprocessing to improve prediction accuracy, an AE layer to extract implicit features and increase training efficiency, and a Bi-LSTM layer to predict. The results indicate that the proposed model’s prediction was better with low RMSE and higher R² values and a positive correlation does exist.

Liu et al.¹⁰ proposed a hybrid ensemble model using DBN, LSTM, and multilayer perceptron (MLP). To reduce the model's complexity, it uses complementary ensemble empirical mode decomposition (CEEMD) to extract the features from the data series. To produce the best forecast outcomes, the imperial competition algorithm (ICA) is used to alter the weights of the predictors. Two sets of hourly PM_2.5 concentrations from Shanghai are used to validate the model. According to the findings of the experiments, the proposed model performed better in terms of accuracy and resilience.

Category 2: neural network

The neural network is based on the idea of classifying input observations by linearly combining datasets. For lowering difficulties and errors, Jiang et al.¹¹ presented the group teaching optimization algorithm (GTOA): the extreme learning machine (ELM) method, which was based on a data-preprocessing strategy. Two-step decomposition is used to break down PM concentration data into high-frequency IMFs. GTOA is then used to optimise ELM. Over a 16-month period, the data covers hourly PM_2.5 levels in Beijing. The findings showed that the proposed model improved prediction accuracy while maintaining a low mean absolute error and root mean square error.

With daily records from cities in Finland and Brazil, Neto et al.¹² evaluated single and neural-based ensemble techniques. The values of MSE, MAPE, MAE, and RMSE were compared. MLP, the ensemble method, has the best overall performance. Suleiman et al.¹³ examined the performance of boosted regression trees (BRT), artificial neural networks (ANN), and support vector machines (SVM) models in forecasting roadside PM_2.5 concentrations in London at nineteen monitoring sites. For performance evaluation, RMSE and other metrics were used. In general, ANN performed better.

Ali Shah et al.¹⁴ assessed the performance of the Phase Space Reconstruction (PSR) technique, which captures multi-time scale information, using radial and linear Support Vector Regression (SVR), Feedforward Neural Network (FFNN), and Random Forest (RF) ML. RMSE and MAE were used to assess the performance of ML techniques on a dataset collected in Saudi Arabia (Masfalah) over a 21-month period. The results showed that FFNN produced a reliable prediction.

Hung et al.¹⁵ conducted research to determine how transported smoke aerosols affect air quality in New York State. ANN was used in the method, and five models with distinct sets of predictors were considered during the summer seasons of 2012–2019. When the models were evaluated using RMSE and R², the results revealed that smoke cases had higher average PM_2.5 concentrations than non-smoke cases.

Category 3: decision tree

In the decision analysis, a decision tree can be used to visually and explicitly depict decisions and decision making. Angelin Jebamalar et al.¹⁶ created a method that combined light gradient boosting and decision tree techniques. The model breaks the tree leaf by leaf using the best fit, allowing it to handle massive amounts of data with little memory and great speed. From 2017 to 2019, the PM_2.5 information for Chennai, India, was collected using IoT devices and stored in the cloud. When compared to RF, DT, and regression approaches, the suggested model outperformed with the lowest mean absolute error and root mean square error values.

Another study¹⁷ focused on PM_2.5 predictions from environmental sensor data streams using a stacked boosting ensemble (STBoost) model with z-score optimization techniques. The STBoost includes Light GBM as the meta regressor and base regressors such as XGBoost and GBM regression and RF to improve the prediction accuracy. The results indicate that the STBoost model outperformed with an accuracy score of 99.52 and an RMSE value of 0.1048.

To improve PM_2.5 perception, Luo et al.¹⁸ used an image-based technique that included CNN and GBM. Daily weather conditions, 6976 pictures, and hourly data from Shanghai were used to create the model (2016). With the proposed technique, the MAE, RMSE, and R² estimations of PM_2.5 are 3.56, 10.02, and 0.85, respectively.

Using a 1.5 years’ dataset of Malaysia,¹⁹ examined the performance of MLP and RF models in predicting PM_2.5. Confusion matrix was utilised as a performance metric. MLP was outperformed by RF overall.

Category 4: regression

The regression model was used in a number of articles to predict particulate matter. Kleine Deters et al.²⁰ present a machine learning technique for predicting PM_2.5 based on pollution and meteorological data from two Ecuadorian cities over a six-year period. Using CGM in regression analysis, it was found that PM_2.5 could be predicted more accurately during extreme weather.

Kowalski and Warchałowski²¹ provide a comparison of machine learning techniques for predicting dust-type air pollution levels. Real-time hourly data from Krakow was utilised to train and test the models using MSE and R² over the course of a year. The best prediction approach, according to the findings, was a regression model.

Gu et al.²² introduced a recurrent air quality prediction (RAQP) model, which combines a recurrent framework and SVR. The model was tested on 180 hourly records from a small Chinese town. The RAQP model was found to be more successful than state-of-the-art air quality predictors with a low RMSE value.

Aljuaid et al.²³ compared numerous forecasting approaches and strategies based on mathematics and machine learning. The one-hour and five-minute datasets were created by combining and manipulating numerous sources of Danish data. For comparison, the researchers utilised multivariate (SVR, DT, and K-nearest neighbour) and univariate (auto-regression) techniques. SVR had the lowest RMSE and MAE values for the one-hour data set, and auto-regression for the five-minutes dataset, according to the findings.

Category 5: support vector machine

For PM_2.5 ground-level forecasting in the city of Bogotá, Mogollón-Sotelo et al.²⁴ presented SVM. Statistical validation was done using RMSE, etc. The SVM model predicts with greater accuracy. An ensemble empirical mode decomposition (EEMD), least square SVM (LSSVM), and PSR were proposed in Ref. 25 as alternative method for forecasting the following day. The empirical results reveal that the EEMD-PSR-LSSVM outperformed other models in terms of MAPE and RMSE values.