East Africa Broadcast Satellite Sizing with ML

This repository combines the engineering sizing model from the Advanced Master's internship report, the accompanying Excel workbook, and MATLAB ensemble models for an NTN/AI-RAN workflow.

Engineering basis

The MATLAB implementation follows the report and workbook values:

• East Africa source grid: 8,239 rows in the report
• usable workbook grid: 8,237 locations, plus two min/max summary rows
• orbital position: 43 degrees East
• DVB-S2 Ku-band link: 14.5 GHz uplink and 12.5 GHz downlink
• TV service across all covered countries
• telemedicine and e-learning limited to Ethiopia
• 99.9% uplink and 99.8% downlink availability, producing 99.70% end-to-end
• 20% modem roll-off and 36 MHz transponder bandwidth
• 16APSK 5/6 for TV and 32APSK 5/6 for e-services
• 4,120 Mbps total aggregate demand
• approximately 42.89 transponder equivalents, rounded to 43 active transponders
• 58 dBW report peak EIRP

The combined carrier quality uses the report equation:

C/(N+I)total =
  [(C/N)up^-1 + (C/I)up^-1 + (C/Im)^-1
   + (C/I)down^-1 + (C/N)down^-1]^-1

Improvements over the original script

• resolves the workbook relative to the repository instead of a hard-coded personal path;
• does not divide broadcast traffic by population;
• prevents e-service samples from being generated outside Ethiopia;
• computes total C/(N+I), link margin, end-to-end availability, and transponder-equivalent demand from report equations;
• cross-checks MATLAB-required EIRP against the workbook EIRP grid;
• splits train/test data by location, preventing the same geographic point from leaking into both sets through different services;
• adds ML models for total C/(N+I), link margin, and required transponders;
• exports reproducible CSV summaries and MATLAB-generated PNG/FIG results.

The supplied original MATLAB script is retained in legacy/.

Run

MATLAB R2026a with Statistics and Machine Learning Toolbox is recommended.

results = run_airbus_ntn_ran_ml_sizing;
verify_airbus_ntn_ran_ml_sizing;

Outputs are written to results/:

• trained models in airbus_ntn_ran_ml_models.mat;
• complete result structure in airbus_ntn_ran_ml_results.mat;
• model metrics and engineering summaries as CSV;
• seven PNG figures and seven editable MATLAB FIG files.

Repository structure

data/       Source sizing workbook
docs/       Internship report
legacy/     Original MATLAB implementation
src/        Workbook, engineering, ML, and plotting functions
results/    Generated models, tables, and figures

The report and workbook remain the authoritative sources for engineering assumptions. The ML models learn deterministic labels produced by those engineering equations; they do not replace link-budget validation.

The report-equation TV EIRP is about 0.58 dB RMSE above the workbook reference. This is expected because the improved model applies the report's 99.8% downlink rain-fade target and the workbook's 0.5 dB potential implementation loss, while the workbook reference column uses its older 99.72% contour formulation.