Anomaly Detection in Commercial Load Data Using Bidirectional LSTM and Autoencoders

Abstract

This study proposes an anomaly detection algorithm, employing Bidirectional Long Short-Term Memory (LSTM) and autoencoders, to address data quality issues in commercial load data. The process involves constructing an LSTM autoencoder model for input reconstruction to derive fitted values. Anomalous instances are detected by calculating the average absolute error between these fitted and actual values. Applied to real load datasets from shopping centers and commercial buildings, this method demonstrates superior anomaly detection performance compared to conventional benchmark methods.