A Cost-Effective CNN-LSTM-Based Solution for Predicting Faulty Remote Water Meter Reading Devices in AMI Systems

• Main Authors: Jaeseung Lee, Woojin Choi, Jibum Kim
• Format: Article
• Language: English
• Published: MDPI AG 2021-09-01
• Series: Sensors
• Subjects: machine learning; advanced meter infrastructure (AMI); CNN-LSTM; deep learning; water; fault detection
• Online Access: https://www.mdpi.com/1424-8220/21/18/6229

Automatic meter infrastructure (AMI) systems using remote metering are being widely used to utilize water resources efficiently and minimize non-revenue water. We propose a convolutional neural network-long short-term memory network (CNN-LSTM)-based solution that can predict faulty remote water meter reading (RWMR) devices by analyzing approximately 2,850,000 AMI data collected from 2762 customers over 360 days in a small-sized city in South Korea. The AMI data used in this study is a challenging, highly unbalanced real-world dataset with limited features. First, we perform extensive preprocessing steps and extract meaningful features for handling this challenging dataset with limited features. Next, we select important features that have a higher influence on the classifier using a recursive feature elimination method. Finally, we apply the CNN-LSTM model for predicting faulty RWMR devices. We also propose an efficient training method for ML models to learn the unbalanced real-world AMI dataset. A cost-effective threshold for evaluating the performance of ML models is proposed by considering the mispredictions of ML models as well as the cost. Our experimental results show that an F-measure of 0.82 and MCC of 0.83 are obtained when the CNN-LSTM model is used for prediction.