💻 computer science

A Controlled Benchmark of Raw Sequential Image Based Hybrid and Statistical Models for Representation Aware Anomaly Detection in Household Electricity Consumption Time Series

This study presents a controlled benchmark comparing raw sequential, image-based (GAF), hybrid, and statistical models for anomaly detection in household electricity consumption, finding that direct temporal modeling (Raw 1D-CNN) outperformed other approaches under synthetic perturbation protocols while emphasizing the need for cautious interpretation due to the dataset's single-household scope and lack of real-world fault annotations.

Original authors: Şükrü Mustafa Kaya, alireza esmaili jobani

Published 2026-07-31

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Şükrü Mustafa Kaya, alireza esmaili jobani

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to catch a thief in a giant, bustling city. The city is full of noise: cars honking, people talking, and lights flickering. Your job is to spot the one person doing something weird, like running the wrong way down a one-way street. In the world of computers, this "city" is a stream of data, and the "thief" is an anomaly—a strange event that shouldn't happen. This field is called anomaly detection, and it's crucial for things like keeping power grids safe or spotting fraud.

To catch the thief, you need a good way to look at the clues. Sometimes, you look at the raw sounds of the city (the raw sequential data). Other times, you might try to turn those sounds into a picture, like a sonogram or a heat map, hoping the weird noise will look like a weird shape (this is called image-based representation). There's also the old-school method of just counting how many cars passed by in an hour and checking if the number is weird (this is statistical learning). The big question scientists have been asking is: "Is looking at the picture better than listening to the raw sound, or is the old-school counting method actually the smartest?"

This paper is like a controlled experiment where the researchers set up a fake city with a fake thief to see which detective method works best. They didn't use a real city with real criminals; instead, they took a dataset of electricity usage from a single house and mathematically "injected" fake weird events into it. They then tested four different detective teams: one that listened to the raw sound, one that looked at the picture, one that tried to do both, and one that just did the math.

The results were a bit of a surprise for anyone who loves fancy pictures. The team that just listened to the raw sound (the Raw 1D-CNN) turned out to be the best detective. It caught the most "thieves" while making the fewest mistakes. The team that tried to turn the sound into a picture (the GAF-CNN) actually did the worst of the bunch. It seems that for these specific, tiny, weird glitches, turning the data into an image actually made it harder to see the problem, kind of like how squinting at a blurry photo might make you miss a tiny detail that is obvious when you look at the object directly.

The team that tried to combine the picture and the math (Hybrid Fusion) did better than the picture-only team, but still couldn't beat the raw sound team. And the old-school math team (XGBoost) was actually quite good, proving you don't always need a super-computer to find a problem.

The authors are careful to say this isn't a magic bullet for catching real-world criminals in a real power grid. Because they used a fake city and a single house, they can't promise their method works everywhere. But their experiment strongly suggests that, at least for this specific type of problem, you don't always need to turn your data into a fancy image to find the bad guys. Sometimes, the simplest way of looking at the raw data is actually the sharpest.

Technical Summary: A Controlled Benchmark of Raw Sequential Image Based Hybrid and Statistical Models for Representation Aware Anomaly Detection in Household Electricity Consumption Time Series

Problem Statement

The detection of rare-event anomalies in multivariate electricity consumption time series is influenced not only by model architecture but significantly by the data representation used prior to learning. While deep learning has advanced anomaly detection, there is insufficient controlled comparison between raw sequential modeling, image-based temporal transformations (specifically Gramian Angular Field or GAF), hybrid fusion strategies, and statistical tabular baselines.

Existing literature often evaluates image-based representations without sufficiently strong comparisons against raw sequential baselines or fails to clarify whether performance gains stem from the representation itself, model capacity, or specific anomaly construction protocols. Furthermore, many studies assume that transforming time-series data into images (e.g., GAF) universally improves feature extraction, a claim that requires rigorous validation under imbalanced, rare-event conditions where fine-grained local temporal deviations are critical.

Methodology

The study establishes a controlled, leakage-aware benchmark using the UCI Individual Household Electric Power Consumption dataset. The experimental design prioritizes reproducibility and strict separation of data to prevent information leakage.

1. Data and Preprocessing:

Dataset: A subset of 300,000 chronologically ordered samples (210 days) from a single household, covering seven numerical features (Global active/reactive power, Voltage, Intensity, and three sub-metering loads).
Splitting: Chronological train-validation-test splitting was performed before any anomaly injection to ensure temporal integrity.
Normalization: Min-max normalization parameters were fitted solely on the training set and applied to validation and test sets.

2. Synthetic Anomaly Construction:
Since the dataset lacks natural fault labels, controlled synthetic perturbations were injected post-split:

Point Anomalies: Random spikes/drops on a single feature (ratio: 0.008).
Contextual Anomalies: Local additive shifts within short segments (ratio: 0.008).
Collective Anomalies: Sustained subsequences with spike-block, flatline, or inverse distortions (ratio: 0.015).

3. Windowing and Labeling:

Sliding Windows: Windows of length 60 with a stride of 20 were generated.
Density-Based Labeling: To avoid over-inflating the anomaly class, a window was labeled anomalous only if the anomaly density (ratio of anomalous points within the window) exceeded a threshold of $\tau = 0.15$ . This resulted in a moderately imbalanced dataset (~11% anomaly ratio).

4. Representation Strategies:
Four distinct input representations were derived from the same windowed data:

Raw Sequential: $60 \times 7$ matrix preserving original temporal order.
GAF Image: Each feature transformed into a $60 \times 60$ Gramian Angular Field image, stacked into a $60 \times 60 \times 7$ tensor.
Statistical Descriptors: A 98-dimensional vector per window containing mean, std, percentiles, differences, and inter-feature correlations.

5. Model Pipelines:
Four pipelines were evaluated under a unified protocol (same splits, thresholds, and metrics):

Raw 1D-CNN: Direct convolution on raw sequential windows.
GAF-CNN: 2D convolutional network with residual blocks and attention on GAF images.
Hybrid Fusion: Concatenation of GAF-CNN features and a statistical descriptor branch.
XGBoost: Gradient boosting on statistical descriptors.

6. Evaluation Protocol:

Imbalance Handling: Minority oversampling was applied only to the training set for deep models; XGBoost used class weighting checks.
Threshold Tuning: Model-specific thresholds were tuned on the validation set to maximize F1-score.
Metrics: Emphasis on F1-score, PR-AUC, Precision, and Recall, alongside Accuracy and ROC-AUC.
Robustness Checks: The study included audits for dataset coverage, window-labeling boundary effects (verifying no structural false negatives for length-10 collective anomalies at $\tau=0.15$ ), XGBoost imbalance weighting, and bootstrap confidence intervals.

Key Contributions

Controlled Benchmark: Developed a reproducible representation-comparison benchmark for household electricity time series, isolating the effects of raw sequential, GAF-based, hybrid, and statistical learning under a shared synthetic anomaly protocol.
Leakage-Aware Design: Established an experimental workflow where chronological splitting precedes anomaly construction and windowing, strictly preventing data leakage.
Empirical Evidence on Representation: Provided evidence that, under specific synthetic perturbation and window-labeling designs, GAF-based time-series imaging does not necessarily outperform direct temporal modeling.
Bounded Interpretation: Deliberately framed results as a comparison of representation-model pipelines on a controlled synthetic benchmark, explicitly avoiding claims of superiority for real-world grid-scale fault detection.

Results

Under the F1-optimized validation threshold, the Raw 1D-CNN achieved the strongest overall performance:

Accuracy: 0.9528
Precision: 0.9116
Recall: 0.6471
F1-score: 0.7569
PR-AUC: 0.7500

Ranking of Models:

Raw 1D-CNN: Best overall balance of sensitivity and false-alarm control.
Hybrid Fusion: Ranked second (F1: 0.6908), significantly outperforming standalone GAF and XGBoost, suggesting that fusing image and statistical features mitigates some limitations of image-only learning.
XGBoost: Competitive (F1: 0.6445), confirming that statistical descriptors remain strong baselines.
GAF-CNN: Weakest performance (F1: 0.5410), indicating that standalone image-based transformation was less effective for the specific sparse, localized perturbations in this benchmark.

Key Observations:

The Raw 1D-CNN preserved local fluctuations and short-range dependencies more effectively than the GAF transformation, which may have smoothed or redistributed fine-grained temporal cues.
All models exhibited moderate recall (~60-65%), indicating that approximately 35% of anomalous windows remained undetected, a critical limitation for operational deployment.
The Hybrid Fusion model improved upon the GAF-CNN but did not surpass the Raw 1D-CNN, suggesting that added representational complexity does not guarantee superior performance.

Significance and Claims

The paper claims that representation choice is a critical, often underexplored factor in anomaly detection performance. The study challenges the assumption that time-series-to-image transformations (like GAF) universally improve discriminative capacity. Instead, it demonstrates that for rare, locally distributed anomalies in household electricity data, direct temporal modeling (Raw 1D-CNN) can be more effective than transformed image-based learning.

Modesty of Claims:
The authors explicitly state that these findings are not proof of superiority for real grid-scale anomaly detection. The study is limited by:

The use of synthetic anomaly labels rather than naturally occurring faults.
Data from a single household, lacking grid-scale topology or multi-household dynamics.
The specific window-labeling and perturbation geometry used.

Therefore, the significance of the work lies in providing a cautious empirical comparison of representation strategies. It argues that before claiming architectural superiority, researchers must systematically evaluate representation-model pipelines under explicit, reproducible, and imbalance-aware conditions. The results suggest that increasing representational complexity does not automatically lead to better anomaly detection and that the most suitable strategy must be determined empirically for the specific operational context.