EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Energy Efficient Edge AI with Measured Energy Saving

Benjamin Agyekum, Agyapong Gloria, Daniel Seffah-Duodu

International Journal of Scientific Research in Science, Engineering and Technology · 25.06.2026 · 13 (3) · 553-579

DOI: 10.32628/ijsrset2613374 ↗

Kurz gefasst

Fachartikel in der Zeitschrift International Journal of Scientific Research in Science, Engineering and Technology, online erschienen am 25.06.2026. Verfasst von Benjamin Agyekum und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The rapid deployment of artificial intelligence at the network edge has created an urgent need for inference systems that can deliver high accuracy while minimizing energy consumption, latency, and hardware overhead. Conventional edge AI models often achieve strong predictive performance but operate with inefficient computational patterns, excessive memory access, and static inference pipelines that fail to adapt to changing workload conditions. This paper proposes an Energy-Aware Adaptive Edge Inference Network (EAAEIN), a novel edge AI optimization framework designed to reduce measured energy consumption while preserving model accuracy across resource-constrained devices. The proposed approach combines dynamic neural pathway pruning, quantization-aware inference, workload-sensitive voltage-frequency scaling, and an energy feedback controller that adjusts computation intensity according to real-time power measurements. Unlike existing techniques such as MobileNetV3, EfficientNet-Lite, TinyML-based static quantization, and early-exit neural networks, EAAEIN integrates model-level compression with device-level energy monitoring to produce measurable and verifiable energy savings during live inference. The study evaluates EAAEIN using benchmark edge AI workloads, including image classification, sensor-based anomaly detection, and lightweight object recognition. Performance is assessed using accuracy, inference latency, throughput, memory footprint, energy per inference, and percentage energy saving. Comparative experiments are conducted against MobileNetV3, EfficientNet-Lite, SqueezeNet, Tiny-YOLO, and a baseline convolutional neural network deployed on edge hardware such as Raspberry Pi, NVIDIA Jetson Nano, and microcontroller-class TinyML platforms. Experimental graphs are used to compare energy consumption, accuracy-energy trade-off, latency reduction, and computational efficiency across all models. The results demonstrate that the proposed EAAEIN model achieves superior energy efficiency by reducing energy consumption per inference while maintaining competitive accuracy and low latency. The measured energy saving is expected to provide practical evidence for sustainable edge intelligence in smart healthcare monitoring, industrial Internet of Things, autonomous sensing, smart agriculture, and low-power surveillance systems. The paper concludes that energy-aware adaptive inference can significantly improve the sustainability and deployability of edge AI systems by aligning algorithmic efficiency with real-time hardware power behavior.

Zitieren

APA

Agyekum, B., Gloria, A., & Seffah-Duodu, D. (2026). Energy Efficient Edge AI with Measured Energy Saving. International Journal of Scientific Research in Science, Engineering and Technology, 13(3), 553-579. https://doi.org/10.32628/ijsrset2613374

Vancouver

Agyekum B, Gloria A, Seffah-Duodu D. Energy Efficient Edge AI with Measured Energy Saving. International Journal of Scientific Research in Science, Engineering and Technology. 2026;13(3):553-579. doi:10.32628/ijsrset2613374

BibTeX
@article{Agyekum2026,
  author = {Agyekum, Benjamin and Gloria, Agyapong and Seffah-Duodu, Daniel},
  title = {Energy Efficient Edge AI with Measured Energy Saving},
  journal = {International Journal of Scientific Research in Science, Engineering and Technology},
  year = {2026},
  volume = {13},
  number = {3},
  pages = {553--579},
  doi = {10.32628/ijsrset2613374}
}
Kostenlos starten →