Deep Learning-Based Approach for Air Quality Forecasting by Using Recurrent Neural Network with Gaussian Process in Taiwan

Abstract

Time series prediction (forecasting) has become the essential issue in many fields, such as stock market, supply chain management, speech recognition, traffic problem and etc. Forecasting the dynamics of sequential events can be applied by using different methods depending on how much detail of the issues have on the probability distribution of the data that is going to be forecasted. Deep learning as one of the approaches of machine learning offers a lot of promise for time series prediction. One of the most popular deep learning methods is the recurrent neural network (RNN). RNNs deal with sequence issues because this algorithm is good at extracting patterns in input feature space, where the input data for this algorithm spans over long sequences. Therefore, this study applies the RNN to the air quality prediction in Taipei, Taiwan. The goal of the model is to minimize mean square error (MSE), which consists of the difference of prediction value and correct value. The data is collected from Taiwanese government. The model consisting of 9.216 time windows (1 year) is implemented using RNN. A Gaussian process algorithm is employed to choose the best parameters in the model. Computational results indicate that RNN with Gaussian process outperforms back-propagation neural network and basic RNN.