SENSOR ERROR DETECTION SYSTEM ON AUTOMATIC WEATHER STATION USING LONG SHORT-TERM MEMORY ALGORITHM

• Main Author: Santoso, Bayu
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/79607
• Institution: Institut Teknologi Bandung

The weather information in Indonesia plays a crucial role in various sectors due to the country's diverse weather and extreme climate phenomena. In aviation transportation, weather information helps determine new routes or avoid routes with adverse weather conditions, ensuring accurate flight path planning. The Meteorology, Climatology, and Geophysics Agency (BMKG) have the responsibility to disseminate weather information to the public. Currently, BMKG operates 124 Technical Implementation Units (UPT) meteorological stations observing and providing weather information across provinces in Indonesia. Of these, 62 UPT meteorological stations conduct automatic weather observations, while the rest still rely on manual observations. Automatic Weather Stations (AWS) are automated instruments that measure and collect meteorological parameters such as air temperature, relative humidity, air pressure, rainfall, wind speed, wind direction, and solar radiation. BMKG operates 368 AWS across Indonesia, calibrated regularly in compliance with legal requirements. However, periodic maintenance is essential to ensure their operational fitness, as AWS data is pivotal in generating accurate weather information. BMKG currently conducts corrective and preventive maintenance for AWS. Although predictive maintenance is part of BMKG's Strategic Plan for 2020-2024, comprehensive implementation has not occurred. This research aims to develop predictive maintenance by monitoring output values and detecting sensor errors in AWS based on historical sensor data. The Long Short-Term Memory (LSTM) algorithm is utilized to model sensor error detection. The study uses datasets of rainfall, air temperature, relative humidity, and air pressure sensors at AWS Jatiwangi, West Java, from 2017 to 2021, with observations every 10 minutes. Sensor condition determination follows the World Meteorological Organization (WMO) criteria for measurement tolerance, where maximum acceptable values are 5% for rainfall, 0.2°C for air temperature, 3% for relative humidity, and 0.15 hPa for air pressure. The research addresses outlier values in the dataset according to BMKG's data quality control standards. The AWS Jatiwangi dataset is divided into 70% for training, 25% for validation, and 5% for testing. The study yields four models: RR, TT, RH, and PP. Each model has three inputs with two hidden layers and 32 neurons. Validation evaluates the LSTM model's performance on an unseen dataset, while testing is conducted on three datasets: a testing dataset with two schemes, an AWS calibration dataset, and an air pressure sensor malfunction dataset. Model evaluation uses a confusion matrix to measure accuracy, probability of detection (POD), and false alarm rate (FAR). Validation results show high accuracy, above 90% for RR, TT, and PP models, with the RH model reaching over 89%. Additionally, POD values are generally high, exceeding 0.9, with low FAR values, below 0.1, except for the RH model. Initial testing on the first scheme dataset shows model accuracy averaging over 95%. POD and FAR are both 0 because the model detects no errors in all sensor parameters. The second scheme indicates that the RR and PP models can detect errors in over 90 data points, while the TT and RH models detect errors in 76 and 78 data points, respectively. The models effectively identify sensors indicating faults when errors occur for four consecutive hours, complying with WMO sensor condition requirements. The second test shows that the TT and PP models can detect normal values in the AWS Jatiwangi calibration sensor dataset. The third test on the air pressure sensor malfunction dataset reveals that the model accurately detects faults at the exact date and time. Additionally, the model detects several error values in the air pressure sensor, interpreted as pre-fault detection, with three errors on April 3rd and 8th, 2021. The LSTM model's implementation on the web-based interface is successful, effectively detecting reading errors in operational AWS Jatiwangi sensors. The December 15-16, 2023, sample indicates that the predominant pattern of values exceeding WMO tolerance limits is observed in the TT and RH sensors, as shown in the first test. ?