Research on Base Station Energy-Saving Optimization Method Integrating Spatial Features and Temporal Modeling

As mobile communication networks continue to scale, energy-saving optimization for base stations is increasingly constrained by the need for fine-grained, timely decision-making under complex and dynamic operating conditions. In practical deployments, the effectiveness and stability of existing energy-saving strategies are primarily limited by two factors: (i) the lack of high-precision and frequently updated base station scene labels that accurately reflect heterogeneous service environments, and (ii) the prohibitive computational cost of modeling complete annual traffic patterns for large-scale networks, which forces most solutions to rely on short-term or sampled data, thereby weakening their ability to capture long-term traffic regularities. The combination of these issues restricts achievable energy-saving depth and undermines operational stability. To address these challenges, this paper proposes a scenario-aware energy-saving optimization framework that integrates semantic scene perception with long-term spatio-temporal traffic modeling. First, fine-grained and dynamically updated scene labels are generated through the fusion of natural language processing on configuration texts and cloud-based geographic map image recognition, enabling accurate characterization of base station service environments without manual annotation. Second, a GeoAutoformer-Transformer model is developed by embedding geographic-aware graph attention into a decompositionbased Transformer architecture, allowing the model to efficiently learn long-term temporal patterns and spatial dependencies among neighboring base stations while remaining computationally tractable for large-scale deployments. In addition, a masking mechanism is introduced to explicitly distinguish between normal zero-traffic intervals caused by energy-saving shutdowns and abnormal zero values resulting from data loss or transmission failures, thereby improving the robustness and reliability of traffic prediction and subsequent energy-saving decisions. Through jointly extending effective energy-saving durations and recommending more suitable energy-saving strategies for eligible base stations, the proposed framework achieves a 17–19% improvement in overall energy-saving efficiency. At the same time, more accurate scene perception and comprehensive traffic modeling contribute to more stable network excellence rates and reduced negative feedback, indicating improved operational reliability. Extensive experiments conducted on a large-scale real-world dataset comprising approximately 50,000 base stations from a nationwide commercial network demonstrate the effectiveness and general applicability of the proposed framework across diverse geographic regions and service scenarios.