Comparison of CatBoost and LightGBM Models for Air Humidity Prediction

This study uses historical weather data from the Badan Meteorologi, Klimatologi, dan Geofisika (BMKG) to evaluate the performance of two combination machine learning models, LightGBM and CatBoost, in predicting air humidity. Daily weather data including temperature, humidity, rainfall, daylight duration, and wind characteristics are included in the dataset. Among the preprocessing procedures were label encoding, normalization with MinMaxScaler, and managing missing values. Date fields' temporal information was extracted using feature engineering. Both models were optimized using GridSearchCV with three-fold cross-validation after being trained with an 80/20 split. Using R², MAE, and RMSE, the model's performance has been evaluated. CatBoost outperformed LightGBM, which received an R² score of 0.7981, with a better R² score (0.8191) and smaller prediction errors (MAE = 0.0570, RMSE = 0.0744). While feature importance analysis indicated that temperature and seasonal features were important predictors, residual plots validated the models low bias and good generalization. Both models can help with strategic decision-making in climate-sensitive businesses and salt production, according to the results, and are suitable for humidity forecasting.