EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Efficient Deep Learning for O-RAN: An Interpretable Framework for KPI Prediction Using Hybrid Reservoir Computing

Md Fahad Monir, Raymond Zhao, Nishith D. Tripathi, Lingjia Liu, Jeffrey H. Reed

IEEE Access · 2026 · 14 · 101205-101218

DOI: 10.1109/access.2026.3709224 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Md Fahad Monir und 4 weiteren Autor:innen. Die Arbeit nennt 44 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Public Wireless Supply Chain Innovation Fund.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Efficient KPI prediction is becoming increasingly important in Open Radio Access Network (O-RAN), where data-driven monitoring and optimization rely on large volumes of network measurements. However, processing a large number of key performance indicators (KPIs) can increase computational cost and latency in the network. Moreover, high dimensionality and redundancy of KPIs make the overall prediction task more challenging for deep learning models, especially when limited training samples are available and fast model adaptation is desired. In this paper, we address these challenges by proposing a framework that uses explainable AI (XAI)-based feature selection techniques, specifically LIME and SHAP, combined with a lightweight hybrid reservoir computing-based Echo State Network-MLP (ESN-MLP) predictor. LIME and SHAP identify the most influential KPIs for Spectral Efficiency (SE) prediction, while ESN-MLP is trained on these features to enable efficient time-series prediction of network metrics, particularly when training samples are limited. Evaluated on O-RAN testbed data, our framework maintains competitive prediction performance with XAI-selected 8-KPI subsets from 13 KPIs compared with baseline models such as MLP, LSTM, and GRU. For example, under the 5% training-data condition, the proposed framework achieves MSE reductions of 20.95%, 24.19%, and 14.43% over the traditional ESN for the full, LIME-selected, and SHAP-selected KPI sets, respectively. Compared with LSTM, the corresponding MSE reductions are 28.66%, 29.40%, and 21.51%, respectively. These results show that XAI-guided KPI selection with ESN-MLP can provide efficient and interpretable SE prediction under limited training-data conditions.

Förderung

  • Public Wireless Supply Chain Innovation Fund

Zitieren

APA

Monir, M. F., Zhao, R., Tripathi, N. D., Liu, L., & Reed, J. H. (2026). Efficient Deep Learning for O-RAN: An Interpretable Framework for KPI Prediction Using Hybrid Reservoir Computing. IEEE Access, 14, 101205-101218. https://doi.org/10.1109/access.2026.3709224

Vancouver

Monir MF, Zhao R, Tripathi ND, Liu L, Reed JH. Efficient Deep Learning for O-RAN: An Interpretable Framework for KPI Prediction Using Hybrid Reservoir Computing. IEEE Access. 2026;14:101205-101218. doi:10.1109/access.2026.3709224

BibTeX
@article{Monir2026,
  author = {Monir, Md Fahad and Zhao, Raymond and Tripathi, Nishith D. and Liu, Lingjia and Reed, Jeffrey H.},
  title = {Efficient Deep Learning for O-RAN: An Interpretable Framework for KPI Prediction Using Hybrid Reservoir Computing},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {101205--101218},
  doi = {10.1109/access.2026.3709224}
}
Kostenlos starten →