TelcoAgent: 5G multi-KPM forecasting with 3GPP explainability

TelcoAgent is a foundation model-based framework for 5G Key Performance Measurement forecasting that combines automated 3GPP knowledge extraction, a time-series foundation model prediction pipeline, and a reasoning layer for domain-grounded explanations. The system was evaluated on a 3-month, city-scale 5G KPM dataset from a U.S.-based network operator and produced high forecasting accuracy for all 7 considered KPMs across 200 cells.

What is TelcoAgent and how does it work?

TelcoAgent is built as a three-component pipeline: an automated three-agent process that constructs a 3GPP knowledge graph from specification documents, a scalable time-series foundation model (TSFM) prediction pipeline that enables zero-shot forecasting across cells, and a reasoning and explanation pipeline that produces domain-grounded diagnostics and actionable instructions. The framework is described by the authors as a single, integrated approach that avoids site-specific training by relying on the extracted 3GPP knowledge graph and TSFM capabilities.

The first component automates knowledge extraction from 3GPP specs into a knowledge graph. The second component applies a TSFM-based prediction pipeline to deliver forecasts without per-site retraining. The third component interprets model outputs to yield explainable insights relevant to network operations.

How was TelcoAgent evaluated and how well did it perform?

The authors evaluated TelcoAgent using a 3-month, real-world, city-scale 5G KPM dataset from a U.S.-based network operator and report high forecasting accuracy for all seven considered KPMs per cell across 200 cells. That empirical test is the framework's cited scale: seven KPMs, 200 cells, and three months of real operator data.

The paper frames those evaluation results as demonstrating accurate, scalable, and explainable forecasting without the need for site-specific training. Specific numerical performance metrics beyond the claim of high accuracy are not provided in the abstract; the evaluation details and figures are contained in the full paper submitted to IEEE GLOBECOM 2026.

Why it matters

TelcoAgent tackles two recurring operational challenges in telecom ML: scalability across many cells and explainability grounded in standards. By extracting a 3GPP knowledge graph directly from specification documents and pairing it with a TSFM that supports zero-shot forecasts, the approach aims to reduce the need for per-site models and to connect predictions to domain knowledge that operators understand. That combination could lower deployment friction for KPM forecasting and make automated diagnostics more actionable for network engineers.

The system also signals a concrete way to combine specification-level knowledge with learned time-series models in a telco setting, which addresses operator concerns about opaque ML outputs and the operational cost of maintaining many localized models.

What to watch

Look for the full paper’s evaluation tables and figures submitted to IEEE GLOBECOM 2026 for granular metrics and error breakdowns across the seven KPMs and 200 cells. The paper includes six figures and six pages of content in the submission; those visuals will be the next place to confirm per-KPM accuracy and the explanations produced by the reasoning pipeline.