Predictive Model for Battery State of Health

Abstract

As battery energy storage systems (BESS) become critical components of grid infrastructure, accurately assessing their State of Health (SoH) is essential for optimizing performance, reducing costs, and ensuring contractual compliance. This thesis investigates the development of accurate, real-time SoH estimation models for utility-scale battery storage sites operated by NextEra Energy. Current SoH measurements—derived from annual capacity tests and Battery Management System (BMS) data—are often inaccurate or infrequent, leading to either over- or under-augmentation and resulting in financial inefficiencies.

To address this gap, four state estimation models were developed and evaluated: an Unscented Kalman Filter (UKF), a Long Short-Term Memory (LSTM) Recurrent Neural Network (RNN), a multitask RNN, and a Delayed Reinforcement Learning (DRL) model. Each model uses operational data—such as voltage, current, temperature, and State of Charge (SoC)—to estimate degradation patterns and predict SoH at the rack, lineup, and site levels. Their outputs were compared against ground-truth capacity test results from a large-scale battery storage site.

The DRL model demonstrated the highest accuracy, achieving a deviation of only 1.6 months compared to capacity test data, significantly outperforming existing BMS readings and the other three models. These findings underscore the value of advanced machine learning techniques in enabling proactive maintenance, optimized augmentation scheduling, and cost-efficient storage site management. This research offers a scalable framework for real-time SoH estimation across large fleets of battery storage assets and contributes to the broader goal of improving grid reliability through smarter energy storage management.