Hybrid CNN-BiLSTM-Attention Modeling for Air Temperature Prediction Using LOF-iForest-RFE Multi-Stage Preprocessing on Soekarno-Hatta Meteorological Station AWS Data (2021–2025)

Nur Fuat, Muhamad
Anggai, Sajarwo
Handayani, Murni

JOURNAL OF APPLIED INFORMATICS AND COMPUTING
Vol. 10 No. 5 (2026): October 2026

Air temperature is a critical meteorological parameter that dynamically affects density altitude and Maximum Take-Off Weight (MTOW) calculations in aviation operations. Although the 10-minute resolution Automatic Weather Station (AWS) data at Soekarno-Hatta International Airport effectively captures micrometeorological fluctuations, it is chaotic and prone to anomalies, often causing prediction failures in conventional models. This study aims to propose an artificial intelligence-based short-term framework to predict air temperature for the next 10-minute horizon by integrating multi-level preprocessing with a hybrid CNN-BiLSTM-AM (Convolutional Neural Network, Bidirectional Long Short-Term Memory, Attention Mechanism) architecture. Utilizing a Data-Centric AI approach on observational data from 2021–2025, the methodology begins with local and global anomaly purification via Local Outlier Factor (LOF) and Isolation Forest (iForest) algorithms. Multicollinearity is then mitigated through hybrid feature selection using Pearson Correlation and Recursive Feature Elimination (RFE). In the modeling phase, the 1D-CNN layer extracts spatial correlations among variables, the BiLSTM evaluates bidirectional temporal dependencies, and the attention mechanism adaptively distributes probability weights to prevent lag response during extreme weather oscillations. The findings demonstrate that data purification is strictly necessary to avoid model generalization failures, and RFE successfully isolated the four most optimal physical predictors: maximum temperature, relative humidity, air pressure, and solar radiation. Comparatively, the proposed CNN-BiLSTM-AM model proved to be the most superior, achieving the highest accuracy with an RMSE of 0.8151 °C, MAE of 0.6329 °C, and R² of 0.9106. The addition of attention and convolution mechanisms was shown to reduce the root mean square error (RMSE) by 16.44% and the mean absolute error (MAE) by 13.64% compared to the pure BiLSTM baseline model. In conclusion, achieving an extreme error deviation (RMSE) consistently below the 1.0 °C operational tolerance confirms that this architecture is highly feasible, robust, and has strong potential as a highly reliable meteorological reference to support Density Altitude and MTOW estimations in equatorial climates.