Застосування моделей часових рядів для прогнозування погодних умов

• Дата: 2025
• Автори: Зуй, Денис, Кузьменко, Ігор, Гевліч, Тарас
• Формат: Стаття
• Мова: Українська
• Опубліковано: Kamianets-Podilskyi National Ivan Ohiienko University 2025
• Онлайн доступ: http://mcm-tech.kpnu.edu.ua/article/view/335209
• Назва журналу: Mathematical and computer modelling. Series: Technical sciences

The study presents an evaluation of time series models ARIMA, SARIMA, and SARIMAX, and assesses their applicability for weather forecasting, which is highly relevant under current climate change conditions. The study presents an evaluation of the time series models ARIMA, SARIMA, and SARIMAX and their applicability for weather forecasting, which is particularly relevant under current climate change conditions. Traditional physical and mathematical models provide high forecasting accuracy but require substantial computational resources and may not promptly adapt to rapidly changing external factors. Therefore, there is a growing need to employ alternative forecasting methods that combine adaptability, computational efficiency, and accuracy. The scientific novelty of this research lies in the first-time comparative analysis of the performance and efficiency of ARIMA, SARIMA, and SARIMAX models for weather prediction based on open datasets from the Kaggle platform, using Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) as evaluation metrics. The purpose of the study is to compare the RMSE metrics of the ARIMA, SARIMA, and SARIMAX models and the computational resources required for their implementation. The system architecture includes modules for data collection, cleaning, processing, and visualization, implemented in the Python programming language using the following libraries: Pandas, NumPy, Statsmodels, Pmdarima, Matplotlib, Seaborn, Plotly, and Streamlit. For model implementation, open-source datasets from the Kaggle platform were processed. Before modeling, the stationarity of the time series was verified using the Augmented Dickey–Fuller (ADF) test at a 10% significance level. The MAE metric was applied to tune the models, while RMSE was used for model evaluation. The results show that the ARIMA model is effective for forecasting non-seasonal weather data, providing an RMSE of approximately 0.2-0.4, although it requires a testing phase of more than 50 iterations. The SARIMA model is recommended for seasonal datasets, achieving an RMSE of around 0.8 after approximately 80 iterations. The SARIMAX model, which incorporates exogenous variables, is also recommended for seasonal data, achieving an RMSE of about 0.75 over 80 iterations. The presented findings demonstrate the potential and applicability of ARIMA, SARIMA, and SARIMAX models for weather forecasting tasks.