Predictive Maintenance Automatic Weather Station Sensor Error Detection using Long Short-Term Memory

Abstract

Weather information plays a crucial role in various sectors due to Indonesia's wide range of weather and extreme climate phenomena. Automatic Weather Stations (AWS) are automated equipment designed to measure and collect meteorological parameters such as atmospheric pressure, rainfall, relative humidity, atmospheric temperature, wind speed, and wind direction. Occasionally, AWS sensors may produce erroneous values without the technicians' awareness. This study aims to develop sensors error detection system for predictive maintenance on AWS using the Long Short-Term Memory (LSTM) model. The AWS dataset from Jatiwangi, West Java, covering the period from 2017 to 2021, will be utilized in the study. The study revolves around developing and testing four distinct LSTM models dedicated to each sensor: RR, TT, RH, and PP. The research methodology involves a phased approach, encompassing model training on 70% of the available dataset, subsequent validation using 25% of the data, and finally, testing on 5% of the dataset alongside the calibration dataset. Research outcomes demonstrate notably high accuracy, exceeding 90% for the RR, TT, and PP models, while the RH model achieves above 85%. Moreover, the research highlights that Probability of Detection (POD) values generally trend high, surpassing 0.8, with a low False Alarm Rate (FAR), typically below 0.1, except for the RH model. Sensor condition requirements will adhere to the rules set by World Meteorological Organization (WMO) and adhere to the permitted tolerance limits for each weather parameter.