Transfer Learning for Cellular Base Station Traffic Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current radio communication systems face challenges in accurately predicting future traffic demand, leading to inefficient spectral bandwidth allocation in base stations, where either insufficient or excessive spectrum is allocated, and historical data from one base station may not be relevant for another, causing prediction errors and overtraining issues.
Innovation Solution
A server-based system that uses transfer learning to manage traffic prediction models among base stations by selecting a source base station with abundant historical data, training a similarity network, and adjusting importance scores to generate a target prediction model for each base station, ensuring accurate radio system parameter prediction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If historical data from one base station is shared with another base station for training, then the destination base station can improve prediction accuracy, but the history data may not be relevant for the destination BS leading to overtraining on irrelevant data
Solution Approach 1:
The patent applies local quality by selectively transferring only the most relevant parameters from the source base station's model to the target base station, rather than transferring all parameters uniformly. This is achieved through computing an importance score matrix that identifies which parameters are most important for each target base station, thereby adapting the model transfer to local characteristics of each destination BS and avoiding overtraining on irrelevant data.
2Reliability
If a base station is allocated more spectral bandwidth to meet future demand, then the base station can provide requested bandwidth from user equipment devices, but the base station may be inefficiently allocated more spectrum than necessary
Solution Approach 1:
The patent applies preliminary action by using the transfer learning model to predict future traffic demand at base stations before the demand actually occurs. The system computes prediction models that forecast upcoming traffic patterns, allowing network operators to proactively allocate spectral bandwidth in advance, thereby ensuring bandwidth availability when needed while avoiding inefficient allocation by basing decisions on accurate predictions rather than excessive provisioning.
3Productivity
If transfer learning is used to share prediction models among base stations, then data and bandwidth efficiency is improved, but the model may require adjustment to respect the importance of parameters for each target base station
Solution Approach 1:
The patent applies parameter changes by dynamically adjusting the importance scores of model parameters for each target base station based on their specific characteristics and similarity to the source base station. The system computes a scaled importance score matrix that modifies the original importance scores according to the degree of similarity between source and target stations, thereby automatically adapting the transferred model to local conditions without requiring manual customization or complex post-processing.
Data Source
AI summary
Transfer learning based on prediction determines a similarity between a source base station and a target base station. Importance of parameters is determined and training is adjusted to respect the importance of parameters. A lack of historical data is compensated by selecting a base station as source base station which has a larger amount of historical data.