Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Generative Adversarial Network with Soft-Dynamic Time Warping and Parallel Reconstruction for Energy Time Series Anomaly Detection

Hardik Prabhu, Jayaraman Valadi, Pandarasamy Arjunan

TL;DR

This work tackles sequential anomaly detection in energy time series by training a 1D-DCGAN on normal data and performing gradient-descent-based inversion in the latent space to reconstruct query subsequences. It introduces Soft-DTW as a differentiable reconstruction loss and implements parallel reconstruction to accelerate test-time performance. Anomaly scores combine Soft-DTW reconstruction with a latent-space penalty, while KDE-based analysis converts detected subsequences into timestamped anomalies, enabling robust evaluation with tolerance for sequential events. The approach, evaluated on hourly energy data from 15 LEAD buildings, demonstrates improved performance over autoencoders and traditional methods, with notable speedups from parallelization and strong performance when using Soft-DTW.

Abstract

In this paper, we employ a 1D deep convolutional generative adversarial network (DCGAN) for sequential anomaly detection in energy time series data. Anomaly detection involves gradient descent to reconstruct energy sub-sequences, identifying the noise vector that closely generates them through the generator network. Soft-DTW is used as a differentiable alternative for the reconstruction loss and is found to be superior to Euclidean distance. Combining reconstruction loss and the latent space's prior probability distribution serves as the anomaly score. Our novel method accelerates detection by parallel computation of reconstruction of multiple points and shows promise in identifying anomalous energy consumption in buildings, as evidenced by performing experiments on hourly energy time series from 15 buildings.

Generative Adversarial Network with Soft-Dynamic Time Warping and Parallel Reconstruction for Energy Time Series Anomaly Detection

TL;DR

This work tackles sequential anomaly detection in energy time series by training a 1D-DCGAN on normal data and performing gradient-descent-based inversion in the latent space to reconstruct query subsequences. It introduces Soft-DTW as a differentiable reconstruction loss and implements parallel reconstruction to accelerate test-time performance. Anomaly scores combine Soft-DTW reconstruction with a latent-space penalty, while KDE-based analysis converts detected subsequences into timestamped anomalies, enabling robust evaluation with tolerance for sequential events. The approach, evaluated on hourly energy data from 15 LEAD buildings, demonstrates improved performance over autoencoders and traditional methods, with notable speedups from parallelization and strong performance when using Soft-DTW.

Abstract

In this paper, we employ a 1D deep convolutional generative adversarial network (DCGAN) for sequential anomaly detection in energy time series data. Anomaly detection involves gradient descent to reconstruct energy sub-sequences, identifying the noise vector that closely generates them through the generator network. Soft-DTW is used as a differentiable alternative for the reconstruction loss and is found to be superior to Euclidean distance. Combining reconstruction loss and the latent space's prior probability distribution serves as the anomaly score. Our novel method accelerates detection by parallel computation of reconstruction of multiple points and shows promise in identifying anomalous energy consumption in buildings, as evidenced by performing experiments on hourly energy time series from 15 buildings.
Paper Structure (18 sections, 13 equations, 3 figures, 3 tables, 1 algorithm)

This paper contains 18 sections, 13 equations, 3 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Background and Motivation
  3. Generative Adversarial Networks (GANs)
  4. GAN Training with Wasserstein Loss
  5. Anomaly Detection Using GAN Inversion
  6. Methodology
  7. The LEAD Dataset and Pre-processing
  8. Train-test Segments
  9. Model Input
  10. 1-D DCGAN Architecture
  11. Inverse Mapping
  12. Reconstruction Loss
  13. Parallel Computation
  14. Anomaly Score
  15. Anomaly Detection Using Kernel Density Estimates
  16. ...and 3 more sections

Figures (3)

  • Figure 1: 1D DCGAN architecture for anomaly detection.
  • Figure 2: A test time series segment from a building with timestamps indexed starting from 0 along the x-axis. The orange line represents the scaled actual meter reading, and the blue line represents the scaled KDE. Actual annotated anomalies are indicated by red crosses, while our method marks anomalies with green dots, considering only KDE values above a threshold.
  • Figure 3: Comparison between time taken for reconstruction done using Mean Square Error loss sequentially and in parallel for different batch sizes.