Research on the usability of different machine learning methods in visibility forecasting

Abstract

Haze pollution, mainly characterized by low visibility, is one of the main environmental problems currently faced by China. Accurate haze forecasts facilitate the implementation of preventive measures to control the emission of air pollutants and thereby mitigate haze pollution. However, it is not easy to accurately predict low visibility events induced by haze, which requires not only accurate prediction for weather elements, but also refined and real-time updated source emission inventory. In order to obtain reliable forecasting tools, this paper studies the usability of several popular machine learning methods, such as support vector machine (SVM), k-nearest neighbor, and random forest, as well as several deep learning methods, on visibility forecasting. Starting from the main factors related to visibility, the relationships between wind speed, wind direction, temperature, humidity, and visibility are discussed. Training and forecasting were performed using the machine learning methods. The accuracy of these methods in visibility forecasting was confirmed through several parameters (i.e., root-mean-square error, mean absolute error, and mean absolute percentage error). The results show that: (1) among all meteorological parameters, wind speed was the best at reflecting the visibility change patterns; (2) long short-term memory recurrent neural networks (LSTM RNN), and gated recurrent unit (GRU) methods perform almost equally well on short-term visibility forecasts (i.e., 1, 3, and 6 h); (3) a classical machine learning method (i.e., the SVM) performs well in mid- and long-term visibility forecasts; (4) machine learning methods also have a certain degree of forecast accuracy even for long time periods (e.g., 7 2h).