Identifying a suitable model for predicting hourly pollutant concentrations by using low-cost microstation data and machine learning

Accurately predicting the concentration of PM2.5 (fine particles with a diameter of 2.5 μm or less) is essential for health risk assessment and formulation of air pollution control strategies. At present, there is also a large amount of air pollution data. How to efficiently mine its hidden features to obtain the future concentration of pollutants is very important for the prevention and control of air pollution. Therefore we build a pollutant prediction model based on Lightweight Gradient Boosting Model (LightGBM) shallow machine learning and Long Short-Term Memory (LSTM) neural network. Firstly, the PM2.5 pollutant concentration data of 34 air quality stations in Beijing and the data of 18 weather stations were matched in time and space to obtain an input data set. Subsequently, the input data set was cleaned and preprocessed, and the training set was obtained by methods such as input feature extraction, input factor normalization, and data outlier processing. The hourly PM2.5 concentration value prediction was achieved in accordance with experiments conducted with the hourly PM2.5 data of Beijing from January 1, 2018 to October 1, 2020. Ultimately, the optimal hourly series prediction results were obtained after model comparisons. Through the comparison of these two models, it is found that the RMSE predicted by LSTM model for each pollutant is nearly 50% lower than that of LightGBM, and is more consistent with the fitting curve between the actual observations. The exploration of the input step size of LSTM model found that the accuracy of 3-h input data was higher than that of 12-h input data. It can be used for the management and decision-making of environmental protection departments and the formulation of preventive measures for emergency pollution incidents.