<![CDATA[Microwave transmission networks operate under strict physical-layer capacity constraints, where excessive utilization may reduce service reliability and increase congestion risk. In practical operations, capacity expansion decisions are both time-sensitive and location-sensitive, requiring operators to determine not only when an upgrade should be executed, but also which links require intervention. However, many existing forecasting approaches remain focused on numerical prediction accuracy and do not explicitly support delivery-aware and targeted planning decisions. This study proposes a delivery-aware hybrid smoothing–residual forecasting framework for link-level microwave throughput prediction using hourly Network Management System (NMS) data from 110 links observed over a six-month period. The framework separates the throughput series into a smoothed structural component and a residual component, where primary forecasting is performed on the structural signal, and residual error is modeled using LightGBM. The final forecast is reconstructed by combining both components and evaluated across multiple horizons using WMAPE, MAE, and RMSE. Experimental results show that CatBoost with LightGBM residual correction provides the most consistent performance across all forecasting horizons. Beyond improving prediction accuracy, the proposed framework enables delivery-aware and targeted capacity planning by estimating threshold-crossing timing and identifying high-risk links. This supports proactive congestion avoidance, more efficient upgrade prioritization, and improved resource allocation in microwave transmission networks.]]>
Copyrights © 2026
