Enhancing Air Temperature Prediction Using Genetic Algorithm-Based Feature Selection and Backpropagation

Extreme temperature changes have significantly impacted human life and the environment, increasing the risk of disease, ecosystem disruption, infrastructure damage, and overall decline in quality of life. Traditionally, communities relied on ethnometeorological knowledge to predict and anticipate temperature changes by observing natural phenomena. Although often effective, this method carried significant risks when predictions were inaccurate. In response, various methods and techniques for air temperature prediction have been developed, utilizing numerous weather features that are considered influential. This study proposed a genetic algorithm-based feature selection to identify the most impactful weather features on air temperature and employed a Backpropagation Neural Network for prediction. These algorithms were integrated to address time complexity and enhance prediction accuracy. The feature selection results indicated that total cloud cover (Nh) was highly relevant to air temperature. Using this feature subset, the model predicted air temperature with a Mean Absolute Error (MAE) of 8.6%, below the 10% threshold, indicating good performance. Additionally, applying this feature subset reduced computational time by up to three times compared to using all-weather features. The findings demonstrate that integrating genetic algorithms with neural networks can effectively optimize predictive models, making them more efficient for real-time applications in weather forecasting.