Integrating AI and Statistical Models for Climate Time Series Forecasting

Abstract

Climate change is a pressing global challenge, and predicting its future patterns is essential for mitigation strategies. This study integrates synthetic and real-world climate datasets to develop predictive models. Specifically, we apply Long Short-Term Memory (LSTM) networks alongside ARIMA and SARIMA models to forecast global temperature anomalies. Synthetic data were generated using a Gaussian-based data simulator calibrated on historical NOAA/IPCC data, contributing 30% of the training set. Validation included Kolmogorov-Smirnov tests to ensure distributional similarity to real data. Preprocessing involved interpolation for missing values and stationarity checks using the Augmented Dickey-Fuller (ADF) test (p < 0.05), with differencing of order one applied where necessary. LSTM model architecture included two hidden layers with 64 and 32 units, sequence length of 30 days, and a dropout rate of 0.2 to prevent overfitting. Model performance was evaluated using RMSE, MAE, and MAPE. LSTM achieved the lowest RMSE of 1.8 and MAPE of 6.3%, outperforming ARIMA (RMSE: 2.4, MAPE: 8.2%) and SARIMA (RMSE: 2.0, MAPE: 7.1%). Random Forest and SVR models yielded RMSEs of 2.2 and 2.3, respectively, and were included for benchmarking. A Monte Carlo simulation with 10,000 iterations and normal distribution assumptions estimated prediction uncertainty, aligned with IPCC emission scenarios. Scenario-based forecasting (A: status quo, B: 50% emissions cut, C: net-zero) was validated against past reductions post-Kyoto and Paris agreements. Forecasts indicate a potential 1.5°C rise in temperature by 2050 under Scenario A. Compared to baseline mean anomaly of 14.3°C, this reflects a significant trend.