LSTM-Assisted O-RAN RRM for Low-Latency Traffic Bursts
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing radio resource management (RRM) methods in O-RAN networks lack the ability to predict future network parameter values, leading to inefficiencies in managing high data rate and low latency applications like augmented reality, virtual reality, and extended reality, which can result in degraded performance due to sudden data bursts.
Innovation Solution
Implementing a machine-learning-based RRM system using Long Short-Term Memory (LSTM) neural networks to predict per-DRB traffic and buffer occupancy, enhancing RRM policies by deriving optimized scheduling priorities for logical channels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional RRM methods are used without prediction capability, then the system is simpler to implement, but network performance degrades during sudden data bursts
Solution Approach 1:
The system performs preliminary actions by predicting future traffic patterns and buffer occupancy values before actual data bursts occur. The LSTM-based prediction module forecasts network parameters in advance, allowing the RRM policy to proactively allocate resources and prevent performance degradation during sudden traffic increases.
Solution Approach 2:
The prediction module acts as an intermediary between raw network measurements and RRM decision-making. It processes historical buffer occupancy and traffic data to generate predicted future values, which then inform scheduling decisions. This intermediary layer transforms complex historical data into actionable prediction insights without requiring direct complex control logic in the RRM policy.
2Loss of time
If machine learning prediction is implemented, then future traffic patterns can be predicted, but computational complexity increases
Solution Approach 1:
The system performs preliminary actions by pre-training LSTM models offline to capture traffic patterns. During runtime, the pre-trained models quickly predict future buffer occupancy and traffic values based on historical data, enabling fast response times without heavy online computational burden. The complex learning is done beforehand, leaving only inference during operation.
Solution Approach 2:
The LSTM model creates a simplified copy or representation of complex traffic patterns by learning from historical data. Instead of processing all raw historical data during scheduling decisions, the model creates compressed prediction outputs that capture essential patterns, reducing computational complexity while maintaining predictive accuracy.
Data Source
Figure 1a
Figure 1b
Figure 2
AI summary
A method for implementing enhanced radio resource management of open radio access network (O-RAN) based on machine-learning-based technique, includes: sending, from a distributed unit (DU) of the O-RAN to a traffic prediction analytics module, values of at least one network performance parameter comprising buffer occupancy (BO) for a plurality of data radio bearers (DRBs) at 5G Quality of Service Identifier (5QI) level; deploying, at the traffic prediction analytics module, a Long Short-Term Memory (LSTM) neural network comprising at least one LSTM unit for data traffic prediction of one of per-DRB data traffic or per-logical channel (LC) traffic for each one of a plurality of logical channels (LCs) based on the at least one network performance parameter; and deriving, by the traffic prediction analytics module based on at least the data traffic prediction, a set of parameters defining a policy for determining a scheduling priority of each one the LCs.