Predicting ozone concentration is an important issue because ozone is a harmful air pollutant that can have negative effects on human health and the ecosystem. However, there are challenges that negatively affect the accuracy of ozone concentration predictions. These challenges include handling non-stationary data, efficiently extracting information, and selecting input features. Thus, our paper proposes a new ozone concentration prediction model called complete ensemble empirical mode decomposition adaptive noise (CEEMDAN)- variational mode decomposition method (VMD)- Gated recurrent unit (GRU)- extreme learning machine model (ELM) to overcome these challenges. This model is created in several steps. First, the best input scenario was chosen using an effective method called Boruta feature selection algorithm (BFSA). The input data included wind speed, relative humidity, solar radiation, temperature, air pressure, NO 2 , SO 2 , PM 2.5 , PM 10 , and CO. In the second step, each submodel was run separately to perform data preprocessing, extract important information, and generate predictions. First, the CEEMDAN method was run to transform non-stationary data into a set of more stationary components called intrinsic mode functions (IMFs). The VMD method was then run to decompose the high-frequency IMFs into modal components, making them suitable for modeling purposes. Afterward, GRU model extracted important information from VMDs and IMFs. Finally, the extracted information was input into the ELM model to forecast ozone concentration in Khorasan province, Iran. The performance of the models was evaluated using several error indices. Specifically, this model achieved mean absolute error (MAE) values of 0.312 and 0.325 during training and testing, respectively. Furthermore, it reduced the testing MAEs of other models by 34% to 55%. It was also observed that the CEEMDAN-VMD-GRU-ELM model improved the R 2 value of other models by 1.02% to 14%. The results also highlighted two significant achievements. Firstly, the CEEMDAN-VMD method effectively addressed the issue of non-stationarity in the input data, resulting in improved prediction accuracy. Secondly, the GRU model improved the feature extraction capabilities of machine learning models. Generally, our paper study demonstrates that the new model is a reliable and effective tool for predicting ozone concentration data. The accurate predictions of this model make a significant contribution to the prediction of environmental pollutants.
Kafi et al. (Sun,) studied this question.