Optimized neural network for daily-scale ozone prediction based on transfer learning

Abstract

Tropospheric ozone (O₃) pollution is worsening in China, and an accurate forecast is a prerequisite to lower the O₃ peak level. In recent years, machine learning techniques have attracted increasing attention in O₃ prediction owing to their high efficiency and simple operation. However, the accuracy of predicting the daily O₃ level is low. This study proposed a novel model by coupling long short-term memory neural network with transfer learning (TL-LSTM), with meteorology and pollutant concentration information as the model input. L2 regularization was applied to reduce the risk of overfitting and to improve the accuracy and generalization ability of the model prediction. Our results indicated that by transferring the knowledge in the model configuration from the hourly LSTM module, TL-LSTM greatly improves the predictability of the daily maximum 8 h average (MDA8) of O₃ in Hong Kong. The coefficient of determination (R²) increased from 0.684 to 0.783 and the mean square error (MSE) reduced from 1.36 × 10^-2 to 1.05 × 10^-2. Furthermore, R² and MSE were the highest in summer, indicating an under-prediction of peak O₃ levels. This was a result of the limited number of high O₃ days, which did not provide sufficient knowledge for the model to make an accurate prediction. Sobol analysis indicated that wind speed was the most sensitive factor in O₃ prediction, largely due to the development of land-sea breeze circulation which effectively traps pollutants and expedites O₃ formation. The results clearly demonstrate the effectiveness of the TL-LSTM in predicting the daily O₃ concentration in Hong Kong. Thus, TL-LSTM can be promulgated into other photochemically active regions to assist in O₃ pollution forecasting and management.

Keywords: Hong Kong; L2 regularization; Long short-term memory; Ozone pollution; Transfer learning.