Neural Network EIRP Prediction for Beamforming Control

The patent applies preliminary action by collecting and storing training data from multiple radio access network nodes in advance, before the actual EIRP prediction is needed. This pre-collected data is then used to train the neural network model, which can subsequently perform rapid and accurate EIRP predictions without requiring real-time data collection. This resolves the contradiction by preparing the data infrastructure beforehand, eliminating the time penalty during actual prediction operations.

The patent applies segmentation by dividing the neural network deployment into two parts: the trained neural network model is transmitted to the distributed unit, while the original training data remains at the central network entity. This segmentation allows the distributed unit to perform local real-time inference without requiring continuous transmission of large volumes of training data, thus improving inference speed while minimizing network communication overhead.

The patent applies copying by creating multi-dimensional patterns through interpolation from existing training data samples. Instead of collecting exhaustive real-world data for every possible scenario, the system generates additional training samples by interpolating between known data points in the multi-dimensional space of beamforming weights, angles, and EIRP values. This copying approach completes the dataset with realistic but synthesized examples, improving reliability while avoiding the complexity of collecting and processing vastly more real data.