Learning Approaches for Weather Forecasting: A Comparative Study

This study investigates daily average air temperature forecasting using meteorological data collected in Istanbul, Türkiye, between January 2015 and December 2024. The dataset consists of 3,653 daily observations including temperature, precipitation, and wind-related variables. To improve forecasting performance, several preprocessing and feature engineering techniques were applied, including seasonal encoding, lagged temperature features, moving averages, and Min–Max normalization. The forecasting task was formulated as a multivariate time-series regression problem, and the temporal structure of the data was preserved through chronological train–test splitting. Six prediction models representing statistical, machine learning, and deep learning approaches were evaluated: Linear Regression, Random Forest, Support Vector Regression (SVR), XGBoost, LightGBM, and Long Short-Term Memory (LSTM). Model performance was assessed using Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and the coefficient of determination (R²). Experimental results showed that Linear Regression achieved the best overall performance with an RMSE of 0.0442 and an R² value of 0.9543. Random Forest produced comparable results with an RMSE of 0.0452 and an R² value of 0.9523. In contrast, SVR, XGBoost, LightGBM, and LSTM exhibited substantially lower predictive performance and negative R² values on the test dataset. The findings indicate that daily temperature data in Istanbul exhibit strong temporal continuity and seasonality. Models capable of effectively utilizing lagged temperature information and seasonal patterns achieved the highest forecasting accuracy. The results also demonstrate that increasing model complexity does not necessarily improve predictive performance when working with relatively limited local meteorological datasets. Overall, the study highlights the effectiveness of simple and interpretable models for short-term temperature forecasting and provides a comparative evaluation of different machine learning approaches under real-world forecasting conditions.