Monthly Temperature Prediction in the Han River Basin, South Korea, Using Long Short-Term Memory (LSTM) and Multiple Linear Regression (MLR) Models

This study compares and evaluates the performance of a statistical model, Multiple Linear Regression (MLR), and a deep learning model, Long Short-Term Memory (LSTM), for predicting monthly mean temperature in the Han River Basin, South Korea. Predictor variables were dynamically selected based on lagged correlation analysis between climate indices and temperature over the past 40 years, identifying the top ten variables with the highest correlations for lag times ranging from 1 to 18 months. The MLR model was developed through stepwise regression with cross-validation, while the LSTM model was constructed using an 18-month input sequence to capture temporal dependencies in the data. Model performance was evaluated using percent bias (PBIAS), Nash–Sutcliffe efficiency (NSE), Pearson’s correlation coefficient (r), and tercile-based probability metrics. Both models reproduced the seasonal variability of monthly temperature with high accuracy (NSE > 0.97, r > 0.98). The LSTM model showed slightly higher predictive skill in several periods but also exhibited larger prediction variance, reflecting the sensitivity of nonlinear architectures to variations in predictor–response relationships. In contrast, the MLR model demonstrated more stable predictive behavior with narrower uncertainty bounds, particularly under low signal-to-noise conditions, owing to its structural simplicity. These findings indicate that the two approaches are complementary; the LSTM model better captures nonlinear temporal dynamics, while the MLR model provides interpretability and robustness. Future work will explore advanced hybrid architectures such as CNN–LSTM and Transformer-based models, as well as multi-model ensemble methods, to further enhance the accuracy and reliability of medium-range temperature prediction.