A Bi-LSTM Autoencoder Framework for Anomaly Detection -- A Case Study of a Wind Power Dataset

TL;DR

This paper introduces a Bi-LSTM Autoencoder framework for detecting anomalies in multivariate time series data, demonstrating high accuracy and effectiveness on wind farm data, advancing the field of time series anomaly detection.

Contribution

The paper proposes a novel Bi-LSTM Autoencoder model that captures long-term dependencies and establishes an optimal anomaly detection threshold, outperforming existing LSTM Autoencoder approaches.

Findings

Achieved 96.79% classification accuracy.

Outperformed traditional LSTM Autoencoder models.

Effectively detected anomalies in wind farm data.

Abstract

Anomalies refer to data points or events that deviate from normal and homogeneous events, which can include fraudulent activities, network infiltrations, equipment malfunctions, process changes, or other significant but infrequent events. Prompt detection of such events can prevent potential losses in terms of finances, information, and human resources. With the advancement of computational capabilities and the availability of large datasets, anomaly detection has become a major area of research. Among these, anomaly detection in time series has gained more attention recently due to the added complexity imposed by the time dimension. This study presents a novel framework for time series anomaly detection using a combination of Bidirectional Long Short Term Memory (Bi-LSTM) architecture and Autoencoder. The Bi-LSTM network, which comprises two unidirectional LSTM networks, can analyze the time series data from both directions and thus effectively discover the long-term dependencies hidden in the sequential data. Meanwhile, the Autoencoder mechanism helps to establish the optimal threshold beyond which an event can be classified as an anomaly. To demonstrate the effectiveness of the proposed framework, it is applied to a real-world multivariate time series dataset collected from a wind farm. The Bi-LSTM Autoencoder model achieved a classification accuracy of 96.79% and outperformed more commonly used LSTM Autoencoder models.