Neural-Network-based Dynamic Area Optimization Algorithm for High-Altitude Platform Station

High-altitude platform stations (HAPSs) have attracted significant attention as promising communication platforms that offer extensive communication services directly from the stratosphere to ground-based smartphones. In the HAPS system, multicell configurations are used to enhance the coverage and capacity. To optimize multicell configurations, including beam directions and beamwidths, in the HAPS service area, the application of a genetic algorithm (GA) as a dynamic area optimization technique has been investigated. However, search methods such as GA have high computational costs because they need to search for optimal values for multicell configurations whenever the user distribution changes. Consequently, it is difficult for them to promptly adapt to abrupt changes caused by large-scale disasters or events. Herein, we propose a dynamic area optimization method that employs a pre-trained neural network (NN) using GA optimization results as training data. The pre-trained NN model enables immediate adaptation to unknown user distributions. However, it is difficult for NN to learn complex intercell relationships. To address this, we incorporate classifier chains (CC) and an ensemble technique to enhance the system’s ability to model adjacent-cell interactions and improve the overall performance. When applied to user distribution data in Japan, the simulation results show that under multicell conditions, the cumulative distribution function of the total throughput achieved by beam estimations using the proposed method closely matches that of the GA.