The widespread adoption of Battery Energy Storage Systems (BESS) is crucial for integrating intermittent renewable sources like solar and wind into the power grid, thereby advancing the goals of green energy. Deploying BESS underground offers a sustainable solution to land constraints and safety concerns. However, the dynamic and complex thermal environment underground severely challenges the accurate State-of-Charge (SOC) estimation, which is vital for the safety, longevity, and operational efficiency of BESS. Data-driven SOC models often suffer from performance degradation due to data distribution shifts caused by temperature fluctuations, especially when operational data for specific underground temperatures is sparse. To tackle this issue, this paper proposes a transfer learning model based on adversarial domain adaptation. The model utilizes a Gated Recurrent Unit (GRU) network for feature extraction and incorporates a Gradient Reversal Layer (GRL) to learn temperature-invariant features through an adversarial training mechanism. This approach effectively transfers knowledge from a data-rich source domain (standard temperature) to data-sparse target domains (varied underground temperatures). Comprehensive experiments on a public battery dataset covering a wide temperature range (-20 ◦C to 40 ◦C) demonstrate that our method significantly reduces SOC estimation errors under unseen thermal conditions compared to conventional models. The proposed solution enhances the reliability and sustainability of underground BESS, contributing to more resilient and efficient green energy infrastructure.