Channel World Model · RSRP Prediction

About the Project

A Channel World Model entry for the 2026 Universal Cup Summer Online Challenge sponsored by Huawei. The goal is to predict reference signal received power (RSRP) at measurement locations for new transmitters with hidden labels, supporting coverage prediction.

Data and features: official training data includes transmitter specifications, measured RSRP samples, and 3D point clouds. The pipeline creates distance, azimuth, antenna downtilt, and frequency-band features, and rasterizes point clouds to describe elevation, density, and obstruction along propagation paths.

Implemented approach: a LightGBM base model combined with transfer across sectors at the same site and residual learning based on relationships between receiver locations and transmitters. PyTorch CNN/U-Net models and prediction ensembles are comparative experimental branches. NumPy, pandas, and SciPy support data processing, geometry, and spatial-neighbor search; scikit-learn supports modeling and evaluation.

Evaluation and reproducibility: candidates were tested on held-out physical sites, using MAE in dB. Submission records show that the runner reproduced 19 CSV files totaling 501,433 prediction rows, with SHA-256 hashes matching the accepted archive.

Competition result: SmallChicken ranked 10th and won a Bronze Medal with 4,584,784 points, according to the official award list.

Follow-up research: cross-transmitter residual transfer for cold-start prediction is a separate experimental branch using complete physical-site splits, out-of-fold predictions, and data-leakage checks. Baselines include propagation models, interpolation, local Gaussian processes, and tabular models. A gating experiment did not achieve the intended reduction in negative transfer; point-cloud effects and robustness under sparse data are still being evaluated. The competition leaderboard score is kept separate from this research branch's experimental evidence.

The project's source code is private.