Multiple strategies for a novel hybrid forecasting algorithm of ozone based on data-driven models

Abstract Ground-level ozone is an air pollutant that has adverse impacts on human health and vegetation growth. The accurate prediction of ozone concentrations is essential for developing strategies for ozone mitigation. To obtain a better forecasting model to predict ozone, this study provides a detailed discussion of the application of three model optimization strategies (i.e., adding decomposition algorithms, adding data and adding factors) to benchmark models, including long short-term memory (LSTM) and support vector regression (SVR), to predict ozone concentrations in Shenzhen. The results showed that adding a decomposition strategy, particularly the wavelet decomposition (WD) algorithm, provided the greatest improvement to the prediction accuracy. Based on this, a novel hybrid forecasting model (WD-LSTMSVR) was further developed that first used the WD algorithm to convert the original data from one dimension to multiple dimensions. Subsequently, each layer of the data set was trained and forecast by the LSTM and SVR models, which involved parameters that were optimized by the autoregressive integrated moving average (ARIMA) partial algorithm and particle swarm optimization (PSO) algorithm. The hybrid forecasting model had the best prediction accuracy performance compared with the benchmark models and optimization models in this study. Our results indicate that the developed hybrid forecasting model is a good technique to provide accurate ozone concentration prediction results.