Machine Learning-Driven Energy Optimization for Sustainable Telecommunications: A Multi-Algorithm Benchmark

Abstract

The telecommunications sector consumes approximately 3% of global energy, with mobile networks accounting for 76% of operator energy costs. This paper presents the first comprehensive multi-algorithm comparison framework for green telecom energy optimization, integrating AutoGluon-based machine learning prediction models with six diverse optimization approaches: NSGA-II genetic algorithm, Particle Swarm Optimization (PSO), Dynamic Programming, Greedy heuristics, Mixed-Integer Linear Programming (MILP), and Simulated Annealing. We evaluate our framework on the International Telecommunication Union (ITU) AI for Good competition dataset comprising 10 telecom sites with hybrid energy systems combining solar, grid, diesel, and battery storage. Our AutoGluon ensemble models achieve strong prediction performance with R² scores of 0.92 for solar generation and 0.96 for load forecasting. Comprehensive evaluation reveals that domain-specific heuristics (Smart Conservative strategy achieving score 17.9) outperform sophisticated metaheuristics (NSGA-II with score 48,341.1) by a factor of 2,700×, while greedy algorithms provide near-optimal solutions in under 2 seconds. We achieve 100% feasibility across all sites and demonstrate that forecast accuracy is critical: a forecast error of 20% degrades performance by 800%. Our competitively validated implementation achieves a score of 6674.4, providing actionable insights for algorithm selection in production telecommunications energy management systems and contributing to sustainable network operations aligned with net-zero climate commitments.