Suitable Recurrent Neural Network for Air Quality Prediction With Backpropagation Through Time

Air pollution currently occurs in developed and developing countries and can disrupt environmental conditions and public health. Determining the level of air pollution (air pollutants) or air quality can be seen from a group of sensitive parameters such as NO2, O3, PM10, PM2.5, and SO2. This study predicts data on air pollutant concentrations over time (time series data) to determine future air quality conditions that are good or bad for health and the environment. Data predictions can use algorithms from artificial neural networks, one of which is the Backpropagation Through Time (BPTT) algorithm. BPTT is a learning algorithm developed from the backpropagation algorithm that is applied to the Recurrent Neural Network (RNN) network architecture. BPTT algorithm and RNN architecture have the advantage of predicting time series data because they not only consider the latest inputs, but also all previous inputs in the network. This study proposes to apply the BPTT algorithm by comparing Elman RNN, Jordan RNN, and hybrid network architecture to predict the time series data of air pollutant concentration in determining air quality. The architecture that is suitable for predicting air pollutant concentrations in determining air quality is Jordan RNN which is based on MAPE testing of 6.481% to 7.177% for each data, and the average MAPE prediction for new input data is 5.9024%. Based on the air quality category, the prediction category of the three architectures produces the same category between prediction categories with air quality categories from real data or in other words, the three architectures are suitable for predicting air quality.