Improved ozone pollution prediction using extreme learning machine with tribas regularization activation function

Abstract

Nowadays, increasing ozone (O3) pollution is becoming a global issue. The increasing of these pollutions has a huge negative impact on human health and also to the ecosystem. In order to reduce the risk of high O3 pollution, an accurate O3 forecasting model should be developed, so that a preventive measure can be taken earlier. Therefore, this study proposed an accurate O3 prediction model using improvement Extreme Learning Machine algorithm based on Regularization Activation Function (RAF-ELM). The experiment conducted by investigating RAF-ELM performance use four types of activation function, i.e., sigmoid, sine, tribas, and hardlim. In this study, RAF-ELM uses single hidden layer feedforward neural networks to predict the air quality index for O3 pollutant based on meteorological variables (Temperature and Wind Speed) and other pollutants (NOx, NO, NO2, CO, PM10, SO2, CH4, NMHC, and THC) in Malaysia using O3 hourly time series data collected at Shah Alam station. It has 107,329 instances recorded from the year 1998 to 2010. The input weight and bias for hidden layers are randomly selected, whereas the best neurons’ number of hidden layer is determined from 5 to 20. The number of neurons (11) with regularization (0.8) using tribas activation function showed the best model. The proposed model has obtained better accuracy performance (0.007999 MSE) and better processing time (2.699 s) compared with conventional MPE. It can be concluded that the proposed algorithm can be used as a good prediction technique for time series data.