[Paper Review] DeepAuto: A Hierarchical Deep Learning Framework for Real-Time Prediction in Cellular Networks

Abhijeet Bhorkar, Ke Zhang+1 more|arXiv (Cornell University)|Dec 12, 2019

Advanced MIMO Systems Optimization9 references4 citations

TL;DR

DeepAuto is a hierarchical deep learning framework that uses stacked LSTMs and feedforward networks to predict cellular network KPIs in real time, capturing spatio-temporal patterns and external factors. It achieves up to 32% lower RMSE in long-term cell load prediction and 75% lower KL divergence in radio quality prediction compared to naive baselines.

ABSTRACT

Accurate real-time forecasting of key performance indicators (KPIs) is an essential requirement for various LTE/5G radio access network (RAN) automation. However, an accurate prediction can be very challenging in large-scale cellular environments due to complex spatio-temporal dynamics, network configuration changes and unavailability of real-time network data. In this work, we introduce a reusable analytics framework that enables real-time KPI prediction using a hierarchical deep learning architecture. Our prediction approach, namely DeepAuto, stacks multiple long short-term memory (LSTM) networks horizontally to capture instantaneous, periodic and seasonal patterns in KPI time-series. It further merge with feed-forward networks to learn the impact of network configurations and other external factors. We validate the approach by predicting two important KPIs, including cell load and radio channel quality, using large-scale real network streaming measurement data from the operator. For cell load prediction, DeepAuto model showed up to 15% improvement in Root Mean Square Error (RMSE) compared to naive method of using recent measurements for short-term horizon and up to 32% improvement for longer-term prediction.

Motivation & Objective

• Address the challenge of accurate real-time prediction of key performance indicators (KPIs) in large-scale, dynamic LTE/5G radio access networks.
• Overcome limitations of traditional time-series and machine learning models in handling multi-scale temporal dynamics, spatial correlations, and external factors like network configurations and events.
• Develop a scalable, reusable framework that enables low-latency, high-accuracy prediction across all cells without per-cell model training.
• Integrate real-time streaming data from nationwide eNBs with minimal latency to support proactive network optimization and automation.
• Validate the framework on two critical KPIs: cell load and radio signal quality (RSRQ), demonstrating superior performance over baseline methods.

Proposed method

• Employ a hierarchical deep learning architecture with horizontally stacked Long Short-Term Memory (LSTM) networks to model instantaneous, periodic (daily/weekly), and seasonal (monthly/yearly) temporal patterns in KPI time-series.
• Integrate feedforward neural networks to encode external factors such as network configuration, day of week, hour of day, and minute of day.
• Use real-time Apache Kafka streams for low-latency data ingestion from nationwide eNodeBs, processed at regional centers near network management systems.
• Train the model using a loss function based on Kullback-Leibler (KL) divergence for distributional prediction of RSRQ, and RMSE/MAE/MAPE for cell load forecasting.
• Deploy a single global model per NMS instead of per-cluster or per-cell models to maintain low latency and scalability.
• Regularly retrain the model to adapt to evolving traffic trends and network changes, ensuring sustained prediction accuracy.

Experimental results

Research questions

• RQ1 Can a unified deep learning framework effectively capture multi-scale temporal dynamics (instantaneous, periodic, seasonal) in cellular network KPIs?
• RQ2 How does incorporating external factors such as network configuration, time of day, and holidays affect prediction accuracy for cell load and RSRQ?
• RQ3 To what extent does spatial correlation between cells improve prediction performance in large-scale networks?
• RQ4 How does the proposed hierarchical deep learning model compare to traditional time-series models (e.g., ARIMA) and supervised learning baselines in real-time KPI prediction?
• RQ5 Can a single global model achieve low-latency, high-accuracy predictions across all cells without per-cell model training?

Key findings

• DeepAuto reduced RMSE by up to 15% for short-term (1-minute) cell load prediction and up to 32% for long-term (1-hour) predictions compared to using recent measurements as a baseline.
• For radio signal quality (RSRQ) prediction, DeepAuto achieved a KL divergence of 0.0353, representing a 75% improvement over the naive method of using the previous time point’s value (KL = 0.14).
• Including periodic and seasonal patterns in the model improved RSRQ prediction accuracy, with KL divergence decreasing from 0.038 to 0.0353 when all temporal components were included.
• The use of spatial relationships between cells did not consistently improve performance, leading to the decision to deploy a single global model per NMS for scalability and latency control.
• The framework successfully handled high-dimensional input data and maintained low-latency predictions using real-time streaming data from 1,500+ cells.
• Performance metrics (RMSE, MAE, MAPE) remained stable across different prediction horizons, with RMSE at 0.083 (1 min), 0.066 (15 min), and 0.067 (1 hour) for cell load.