Prediction of ground-level ozone by SOM-NARX hybrid neural network based on the correlation of predictors

Current approaches to ozone prediction using hybrid neural networks are numerous but not perfect. Decomposition algorithms ignore the correlation between predictors and ozone, and feature extraction methods rarely select appropriate predictors in terms of correlation, especially for VOCs. Therefore, this study proposes a hybrid neural network model SOM-NARX based on the correlation of predictors. The model is based on MIC to filter predictors, using SOM to make predictors as feature sequences and using NARX networks to make predictions. Data from the JCDZURI site were used for training, testing, and validation. The results show that the correlation of the predictors, classification numbers of SOM, neuron numbers, and delay steps can affect prediction accuracy. Model comparison shows that the SOM-NARX model has 13.82, 10.60, 6.58% and 12.05, 9.44, 68.14% RMSE, MAE, and MAEP in winter and summer, which is smaller than CNN-LSTM, CNN-BiLSTM, CNN-GRU, SOM-LSTM, SOM-BiLSTM, and SOM-GRU.

Automated summary

This study proposes a hybrid neural network model SOM-NARX based on the correlation of predictors for ozone prediction. The model filters predictors using MIC, transforms them into feature sequences using SOM, and makes predictions using NARX networks. The results show that the correlation of predictors, classification numbers of SOM, neuron numbers, and delay steps can affect prediction accuracy. Model comparison shows that the SOM-NARX model outperforms other models in terms of RMSE, MAE, and MAEP.