Time-series forecasting has been a challenging task in real-world applications due to data with non-linear, non-stationary and noisy characteristics. In the past decades, machine learning and statistical techniques have been widely applied for real-world time series forecasting problems. With the new developments in the field of deep learning in recent years, it has received the attention of most researchers in order to propose more effectively and accurately forecasting systems. in this work, it has been proposed an air pollution forecasting approach by using the Convolutional Long Short-Term Memory (ConvLSTM) neural network model recently developed for analyzing and extracting high-level spatiotemporal features. This model presents a forecasting system for multivariate, multi-time step and multi-output real-world time series problems. The hourly data set including the PM2.5 concentrations and the weather information from January 2010 to December 2014 in Beijing, China was selected for the proposed ConvLSTM model. To evaluate the performance and efficiency of the proposed model, it is compared with several state-of-the-art approaches consisting of Decision Tree Regression (DTR) based on machine learning, a single-structure fully connected layer and LSTM neural network with a fully connected layer based on recurrent neural network. Taking into account different performance metrics, the experimental results indicated that the presented model outperforms other approaches in terms of forecasting accuracy.
Anahtar Kelimeler: Air Pollution Forecasting, Convolutional LSTM, Deep Neural Network, Multi-Output Time Series Forecasting