Comparing Object Detection Models for Electrical Substation Component Mapping

The gist The research trains and compares three object detection models (YOLOv8, YOLOv11, RF-DETR) to map key electrical substation components in US images, aiming to find the most reliable method for automated mapping.

Introduction and Problem Statement

Electrical substations are pivotal nodes in the electrical grid that step up or step down voltage levels, and their failure can have significant economic and public safety implications The importance of substations and their components goes beyond standard maintenance and regulation because any failure can have significant economic and public safety impacts Traditional manual mapping of substation components is time-consuming and labor-intensive, requiring skilled personnel to inspect each site in person. The study investigates whether state-of-the-art computer vision models can be trained to accurately identify and differentiate between key substation components from high-resolution imagery.

Literature Review of Object Detection Models

The literature review compares how various research papers approach object detection for power infrastructure modeling. Convolutional Neural Networks, or CNN algorithms, are deep learning algorithms used for object detection, segmentation, and image classification. CNNs consist of five layers: a convolution layer, a pooling layer, activation functions, a fully connected layer, and loss functions. Overall CNNs have demonstrated strong performance in detecting substation infrastructure. Many research studies utilize the You Only Look Once (YOLO) family of algorithms utilizing a CNN backbone. YOLO models are often utilized for substation infrastructure detection due to higher efficiency, such as YOLOv8 having a higher inference speed without significant cuts in accuracy and precision. Advanced examples include fusing the YOLOv8 algorithm with an enhanced small-object detection head to improve accuracy in unexpected environments. A newly released model, Roboflow Detection Transformer (RF-DETR), is another focus of this research paper due to its limited number of studies evaluating its performance.

Model Training and Data Preparation

The methodology involves several key steps for model training, including data collection, pre-processing, and augmentation. The dataset was built using open-source imagery of U.S. substations and transmission lines. A dataset of 750 images was collected, with 250 for each component (transformers, circuit breakers, reactors). Preprocessing involved resizing and auto-orienting the images to a fixed size of 640x640 using letterbox padding. Data augmentation included rotations of 15°, 30°, -15°, and-30° degrees, resulting in 875 training images per component. Furthermore, the dataset was expanded by randomly adjusting the hue of the images to create variations in color tint up to 15 degrees.

Model Comparison and Results

The study trained YOLOv8, YOLOv11, and RF-DETR on the prepared dataset. The results showed that machine learning models have a significantly lower average accuracy detecting reactors, over circuit breakers and transformers. Of the three components attempted to detect, circuit breakers were the most accurately detected in each model with the highest average mAP. YOLOv8 achieved the highest average mAP among the three models tested, reaching an average of 0.610. The efficiency comparison showed that YOLOv11 was the most efficient model in terms of training time, but this did not translate into higher average accuracy.

Conclusion and Implications

The research concludes that automated substation component mapping is more effective than a manual approach because it can process many images quickly and save much time. Although the resulting model accuracies would not yet support real-world deployment due to the relatively small dataset, the work demonstrates that object detection models can still begin to recognize substation components under these constraints. The YOLOv8 model is identified as the best model for this scenario due to its higher average mAP, indicating it performed best in terms of overall detection accuracy. This research serves an important purpose in understanding the benefits of utilizing automated substation mapping rather than manual. By continuing to explore these AI-driven approaches, we can enable more informed responses to dangers. The final results showed a count of 7615 circuit breakers, 3132 transformers, and 1133 reactors labeled through the inference mapping done on these images.

Bibliography

Alemazkoor, N., Ayyalasomayajula, A., & Li, M. (2021). Mapping electrical substations using deep learning and overhead imagery (arXiv:2104.06601). arXiv. https://doi.org/10.48550/arXiv.2104.06601

Alzubaidi, L., Zhang, J., Humaidi, A. J., Duan, Y., Santamaría, J., Fadhel, M. A., & Farhan, L. (2021). Review of deep learning: Concepts, CNN architectures, challenges, applications, future directions. Journal of Big Data, 8(53). https://doi.org/10.1186/s40537-021-00444-8

Chai, S., & Lau, V. K. N. (2021). Multi-UAV trajectory and power optimization for cached UAV wireless networks with energy and content recharging: Demand-driven deep learning approach. IEEE Journal on Selected Areas in Communications, 39(10), 3208–3224. https://doi.org/10.1109/JSAC.2021.3088694

Erenoğlu, A. K., Sengor, I., & Erdinç, O. (2024). Power system resiliency: A comprehensive overview from implementation aspects and innovative concepts. Energy Nexus, 15(100311). https://doi.org/10.1016/j.nexus.2024.100311

Faisal, M. A. A., Mecheter, I., Qiblawey, Y., Fernandez, J. H., Chowdhury, M. E., & Kiranyaz, S. (2025). Deep learning in automated power line inspection: A review. Applied Energy, 385(125507). https://doi.org/10.1016/j.apenergy.2025.125507

Florkowski, M., Kuniewski, M., Furgal, J., & Pajak, P. (2017). Investigation of overvoltages in distribution transformers. In Proceedings of the 18th International Scientific Conference on Electric Power Engineering (EPE) (pp. 1–4). https://doi.org/10.1109/EPE.2017.7967288

Gallagher, J. (2024). A multispectral automated transfer technique (MATT) for machine-driven image labeling utilizing the Segment Anything Model (SAM) [Data set]. IEEE DataPort. https://doi.org/10.21227/g06c-yh08

Gallagher, J. E., & Oughton, E. J. (2023). Assessing thermal imagery integration into object detection methods on air-based collection platforms. Scientific Reports, 13(8491). https://doi.org/10.1038/s41598-023-34791-8

Gallagher, J. E., & Oughton, E. J. (2024). VORTEX: A spatial computing framework for optimized drone telemetry extraction from first-person view flight data (arXiv:2412.18505). arXiv. https://arxiv.org/abs/2412.18505

Gallagher, J. E., & Oughton, E. J. (2025). Surveying You Only Look Once (YOLO) multispectral object detection advancements, applications, and challenges. IEEE Access, 13(8654–8684). https://doi.org/10.1109/ACCESS.2025.3526458

Girshick, R. (2015). Fast R-CNN (arXiv:1504.08083). arXiv. https://doi.org/10.48550/arXiv.1504.08083

Girshick, R., Donahue, J., Darrell, T., & Malik, J. (2013). Rich feature hierarchies for accurate object detection and semantic segmentation (arXiv:1311.2524). arXiv. https://doi.org/10.48550/arXiv.1311.2524

Hall, J. W., Tran, M., Hickford, A., & Nicholls, R. J.

1. Bold model selection for component mapping: Select YOLOv8 as the best model to use for this scenario because it achieved the highest mean average precision, or mAP, indicating it performed best in terms of overall detection accuracy.

2. Bold component-specific performance analysis: Identify that circuit breakers were the most accurately detected in each model with the highest average mAP, while noting that reactors performed significantly worse in each model, reaching a maximum mAP of around 0.45 due to their appearance varying.

3. Bold real-world deployment capability: The improved system can be utilized for automated substation component mapping by leveraging the inference mapping done on collected images to provide a concrete number of each component type, such as Circuit Breakers 7615.

4. Bold efficiency optimization strategy: Implement a decision rule that favors accuracy over speed, as the paper found that efficiency is not a major concern compared to accuracy for this use case, concluding that YOLOv8 is the optimal choice despite slower training times compared to YOLOv11.

5. Bold data constraint acknowledgment: The system should incorporate a mechanism to acknowledge current limitations, stating that Because our dataset was relatively small and contained lower-quality images, the resulting model accuracies would not yet support real-world deployment, positioning the output as a proof of concept.

Electrical substations are a significant component of an electrical grid. Indeed, the assets at these substations (e.g., transformers) are vulnerable to hazards such as hurricanes, flooding, earthquakes, and geomagnetically induced currents (GICs). Because failures can have significant economic and public safety implications, identifying key substation components is essential for quantifying vulnerability. Unfortunately, traditional manual mapping of substation infrastructure is time-consuming and labor-intensive. Therefore, an autonomous solution utilizing computer vision models is preferable, as it offers greater convenience and efficiency. In this study, we train and compare 16 models on a manually labeled dataset of US substation images. These models include 12 You Only Look Once (YOLO) models, 2 Roboflow Detection Transformer (RF-DETR) models, and 2 Cascade R-CNN models. RF-DETR-large achieved the highest overall detection performance with mAP@50 and mAP@50:95 scores of 0.881 and 0.632, respectively. Across all models, alternate energy systems were detected most accurately, while transformers and reactors were more difficult to identify due to their smaller size and greater visual variability. Applying our best-performing model to nationwide imagery yielded approximately 22,591 component detections across 11,083 unique substations within the United States. These detections were broken down by state and Federal Energy Regulatory Commission (FERC) regions, with Florida (2,478 detections) and Midcontinent Independent System Operator (MISO; 4,329 detections) having the largest number of detections in their respective categories.