Comparative study of seasonal autoregressive integrated moving average and Holt–Winters modeling for forecasting monthly ground-level ozone

Abstract

In this study, ground-level ozone is modeled using Seasonal Autoregressive Integrated Moving Average (SARIMA) and additive Holt–Winters models over the North-Western cluster of Ethiopia using four homogeneous series of more than 13 years of data from the European Center for Medium-Range Weather Forecasts. We split the dataset into training and testing sets. We used the data during the period 2007–2019 for model formulation and parameter estimation and the one year data in 2020 to test model forecasts. More than 60 SARIMA models have been generated for the time series. The goodness of fit of these models has been assessed using the Akaike information criterion and Bayesian information criterion. After rigorous assessment, the seasonal ARIMA(2,0,4)(1,1,2)[12], ARIMA(3,1,0)(2,0,0)[12], ARIMA(0,1,2)(0,0,2)[12], and ARIMA(2,0,0)(2,1,0)[12] models have been identified as best predictive models for Addis Ababa, Ras Dashen, Danakil Depression, and Bahir Dar, respectively. We applied model evaluation metrics, such as root mean square error, mean absolute error, and Mean Absolute Percentage Error (MAPE) to compare the accuracy between seasonal ARIMA and Holt–Winters models. Among the SARIMA and Holt–Winters models, our findings show that the best model for forecasting surface-level ozone is ARIMA(2,0,4)(1,1,2)[12] and ARIMA(3,1,0)(2,0,0)[12] for Addis Ababa and Ras Dashen stations, respectively. However, for Danakil Depression and Bahir Dar stations, the Holt–Winters model with α = 0.346, β = 0.023, γ = 0.36 and α = 0.302, β = 0.019, γ = 0.266, respectively, are found to be best models than the SARIMA. Moreover, the maximum MAPE were found to be less than 7.86% in the study, and hence all the forecasts are acceptable.