A hybrid model based on convolutional neural networks and long short-term memory for ozone concentration prediction

Ozone, which is one of the most crucial pollutants regarding air quality and climate change, negatively impacts on human health, climate, and vegetation; therefore, the prediction of surface ozone concentration is very important in the protection of human health and environment. In this study, a convolutional neural networks and long short-term memory (CNN-LSTM) hybrid model that combines convolutional neural network (CNN), which can efficiently extract the inherent features of huge air quality and meteorological data, and long short-term memory (LSTM), which can sufficiently reflect the long-term historic process of the input time series data, was proposed and used for the ozone predictor to predict the next day's 8-h average ozone concentration in Beijing City. At first, the number of historic data was set as 34 days via optimization, so that the input data suitable for the CNN-LSTM model to ensure the quick and precise prediction of ozone were constructed. In addition, the CNN-LSTM model candidates with different structures were proposed and used to construct the optimal model structure for the proposed ozone predictor. Finally, the performance of the proposed ozone predictor was evaluated and compared with multi-layer perceptron (MLP) and LSTM models; as a result, the performance indexes (RMSE, MAE, and MAPE) were reduced to 83% compared to the MLP model and 35% compared to the LSTM model. In conclusion, it was demonstrated that the proposed CNN-LSTM hybrid model has the satisfactory seasonal stability and the prediction performance superior to MLP and LSTM models.