Energy big data abnormal cluster detection method based on redundant convolution codec

Rui Ma, Zhenhua Yan, Jia Liu, Wenni Kang, Dongge ZhuScientific 2024 · DOI

TL;DR

Researchers have developed a new deep-learning framework that uses redundant convolutional codecs and "capsule layers" to detect anomalies in complex energy big data. By analyzing the hidden correlations between electricity, gas, and heat usage, the method achieves over 98.7% detection accuracy, significantly outperforming traditional statistical and standard neural network approaches.

What Problem Does This Paper Address?

As our energy grids become "smarter," they generate massive amounts of data across multiple streams: electricity, natural gas, thermal heating, and solar output. Detecting anomalies in this "energy big data"—such as equipment failures, energy theft, or cyber-attacks—is notoriously difficult.

The primary challenge is that abnormal data points are often "scattered" and don't form obvious clusters. Furthermore, these different energy types are deeply coupled (e.g., a drop in temperature might spike both gas and electricity usage), and traditional 1D time-series analysis often fails to capture these complex, multi-dimensional relationships. High dimensionality and "noisy" data sharing environments further complicate the task.

Methods

The authors propose a multi-stage deep learning architecture to solve these issues:

  1. Copula-Based Quantitative Analysis: The system first uses Copula functions to model the statistical dependence between different energy types (electric, thermal, gas, and light). This allows the model to understand how these variables move together.
  2. Redundant Convolutional Codec: Unlike standard CNNs, this "redundant" codec reassembles and structurally encodes features. This helps the model learn both local and global patterns within the high-dimensional data.
  3. Coupling-Time Capsule Layer: Borrowing from computer vision, the authors use capsule layers to capture "coupling-time" features. This is the "secret sauce" that identifies how different energy loads interact over specific time intervals.
  4. 3D Feature Transformation: The time-series data is transformed into a three-dimensional feature space. This makes the scattered anomalies easier to group into distinct clusters.
  5. Hierarchical Clustering: To speed up the process, the model uses a double-layer iterative training method that combines global and local clustering, ensuring the model converges quickly on the optimal number of user profiles.

Datasets

The study validated its model using a real-world dataset from an integrated energy system in Northern China.

  • Timeframe: Full year of 2020 data.
  • Split: January to August for training; September to December for testing.
  • Variables: Electricity load, thermal load, gas load, and photovoltaic (solar) output.
  • Sampling Interval: 1 hour.
  • Baselines: The model was compared against three existing methods from recent literature (References 8, 9, and 10 in the paper), which include Bayesian filtering and standard deep transfer learning approaches.

Key Insights and Ideas

  • The Power of Coupling: The authors argue that an anomaly isn't just a "weird number" in one stream; it’s often a break in the relationship between streams. By focusing on the "coupling" of gas and electricity, they find patterns that single-stream monitors miss.
  • Redundancy is a Feature, Not a Bug: By using redundant encoding, the model is more robust against the "scattered" nature of abnormal data. It essentially looks at the data from multiple overlapping perspectives to ensure nothing is missed.
  • Dimensionality Flattening: One of the cleverest steps is using a linear layer to transform complex multi-energy "load maps" into a one-dimensional form for final detection, simplifying the computation without losing the underlying 3D feature relationships.

Key Results

The proposed method showed a clear lead over traditional benchmarks:

  • Accuracy: Reached 98.7% when the number of iterations hit 500. For comparison, the baseline methods hovered between 86% and 89.9%.
  • Efficiency: The model reached its "expected detection accuracy" in just 220 iterations, whereas other methods required between 520 and 695 iterations to reach the same level.
  • Reliability: The error rate remained consistently below 0.1, even as the sample size increased, suggesting the model is highly resistant to overfitting.
  • User Profiling: The model successfully distinguished between different user types (e.g., residential vs. commercial) by analyzing the Spearman rank correlation between their various energy loads.

Limitations and Future Work

While the results are impressive, a few areas remain for further exploration:

  • Geographic Specificity: The data is specific to Northern China. Energy usage patterns (especially thermal and solar) vary wildly by climate, so the model would need retraining/validation for different geographic regions.
  • Computational Overhead: While the paper notes fast convergence, the use of capsule layers and redundant codecs is generally more computationally intensive than simple feed-forward networks. The authors do not deeply discuss the hardware requirements for real-time deployment.
  • Data Privacy: The introduction mentions blockchain and attribute encryption as potential solutions for data security, but the core detection model doesn't explicitly integrate these privacy-preserving features into the neural network architecture itself. This remains a promising direction for "secure" anomaly detection.

Summary updated

Abstract

Due to the scattered distribution and poor clustering of abnormal clusters in energy big data, the ability to detect anomalies is poor. Therefore, a high-energy data anomaly clustering detection method based on redundant convolutional encoding is proposed. Quantitative analysis of the coupling characteristics of electrical thermal gas optical time series for multi energy users based on Copula function, and incorporating quantitative values into multi energy feature indicators to extract the energy consumption behavior characteristics of multi energy users. Utilize redundant convolutional codecs to recombine and structurally encode abnormal features of energy big data, and capture multi energy coupling time features using coupling time capsule layers. Then, coupling time features are synthesized through fully connected linear regression layers to generate anomalous clustering feature components, and the energy time series data is then transformed into feature values of the time series in three-dimensional space. Based on this, a comprehensive energy system and massive multi energy user energy big data anomaly clustering analysis are carried out to determine the optimal number of multi energy users. Then, based on linear layers, the electricity heat gas light load characteristic map of multi energy users is transformed into one-dimensional form, and an energy big data anomaly clustering detection model is constructed to complete anomaly detection. The simulation results show that the proposed method has excellent feature clustering performance, detection accuracy above 98.7%, fast convergence speed, and an error rate below 0.1, which has reliable application value.
BibTeX
@inproceedings{ma2024energy,
  author = {Rui Ma and Zhenhua Yan and Jia Liu and Wenni Kang and Dongge Zhu},
  title = {{Energy big data abnormal cluster detection method based on redundant convolution codec}},
  booktitle = {Scientific},
  year = {2024},
  doi = {10.1038/s41598-024-59373-0},
}
Figure