CLIMATE CHANGE TREND PREDICTION USING MACHINE LEARNING WITH STATISTICAL TIME SERIES ANALYSIS

Abstract

Climate change trend prediction has become a critical task for understanding long-term global temperature shifts and guiding policy decisions. Traditional statistical models, such as SARIMA, have been widely used to model climate anomalies, but they often struggle to capture complex non-linear patterns. This paper introduces a hybrid approach that combines SARIMA with deep learning models, particularly Long Short-Term Memory (LSTM) networks, to improve forecasting accuracy. By leveraging SARIMA to model the linear trend and seasonality, and LSTM to learn from the residuals, the hybrid model significantly outperforms the SARIMA baseline in predicting global temperature anomalies. The model is trained and tested on NASA GISTEMP’s monthly temperature anomaly data spanning from 1880 to 2025. Performance evaluation, based on out-of-sample data from 2016 to 2025, demonstrates an approximate 14% improvement in forecasting accuracy (RMSE and MAPE). The proposed method addresses the complexities of climate data by integrating both linear and non-linear components, offering a more accurate and reliable tool for future climate trend predictions.