<![CDATA[Automatic Weather Stations (AWS) continuously collect and record meteorological data in real time and support weather information services. The reliability of AWS observations depends on sensor performance, as sensor failures, equipment degradation, and communication disturbances may produce anomalous data and reduce data quality. This study aims to develop an early-detection approach for AWS sensor anomalies using a Deep Neural Network (DNN) and an Autoencoder, compare their performance, and identify the sensors most frequently associated with anomalies. The dataset consists of 385,237 observations collected from the AWS Ancol station in North Jakarta from January to September 2025, covering nine meteorological and oceanographic parameters. The study involved data preprocessing, Min-Max normalization, model training, and evaluation using accuracy, precision, recall, F1-score, confusion matrix, and ROC-AUC. The DNN achieved 99.6% accuracy, 0.922 precision, 0.996 recall, 0.958 F1-score, and 1.000 AUC. The Autoencoder achieved 95.8% accuracy, 0.578 precision, 0.023 recall, 0.043 F1-score, and 0.584 AUC. The AUC of 1.000 obtained by the DNN should be interpreted within the characteristics of the labeled dataset used in this study and should not be considered evidence of perfect generalization. Sensor analysis identified wind direction as the parameter most frequently associated with anomalies, contributing 46.67% of the detected anomalies in the DNN and 66.02% in the Autoencoder. These findings indicate that the supervised DNN performed better than the Autoencoder for anomaly detection on the labeled AWS dataset used in this study.]]>
Copyrights © 2026
