Semi-AI Approach for RSRP Prediction with Closed-Form Coverage Preservation in Cellular Networks

Abstract

Accurate Reference Signal Received Power (RSRP) prediction is essential for wireless network planning and optimization. Traditional methods rely on drive test data, which are costly and spatially limited. Several studies have presented AI-driven approaches using measurement reports (MRs) from user equipment (UE). These deep neural network (DNNs) were trained using MR and cell configuration data, outperforming the standard propagation model (SPM). However, lightweight models trained using MR data exhibit distortion in coverage due to the limited data in certain areas. To address this, we propose Semi-AI, a transfer learning-based framework that fine-tunes trained models to align predictions with real-world signals. In contrast to typical transfer learning applications, which aim to reduce training time, Semi-AI seeks to preserve the signal coverage generated by the path loss model, maintain high prediction accuracy, and offer flexible coverage planning without extensive drive testing.