Data-Driven State-of-Charge Assessment for Batteries in Radio Base Stations

Sammanfattning

A radio base station (RBS) is a vital component in enabling mobile communication. Although these systems are designed to handle high traffic loads, most stations are not located in high-traffic areas, resulting in low average resource utilisation. This opens the opportunity to use the batteries in these stations as local energy storage systems to support electricity grid services, for which accurate information and prediction of the battery's operational status is of great importance. One particularly important battery parameter is state of charge (SOC), which indicates the remaining battery capacity as a percentage of its total capacity. As SOC cannot be measured directly and reporting from the battery management system (BMS) has been found inconsistent, this work investigates an alternative approach to estimate SOC from battery quantities reported by the RBS power infrastructure, hereafter referred to as virtual state-of-charge (vSOC). vSOC was calculated based on the current delivered from the RBS to the battery, with the RBS-to-battery charge efficiency estimated and incorporated to account for operational characteristics between the RBS and its battery. The results show that the estimated vSOC produced values consistent with those reported by the BMS. Building on this, long short-term memory (LSTM) and gated recurrent unit (GRU) models were developed to predict SOC over different horizons ranging from 0.5 h to 2.0 h. To mitigate data scarcity during real-world deployment, the physics-informed machine learning (PIML) framework was investigated to enable efficient training with less data, physically consistent predictions, and improved generalisation to unseen data. The results indicates that the PIML models can produce more accurate and less volatile predictions compared to purely data-driven models, while also requiring fewer computational resources during inference. The energy use and carbon emissions of all models were estimated, and a framework for investigating how minimising training data affects prediction accuracy was developed. Overall, the findings demonstrate the feasibility of estimating and predicting SOC from RBS infrastructure measurements, and that embedding physical knowledge into the models during training provides a strong foundation for further development and validation.