Application of Machine Learning Techniques for Battery Bank Health Prediction

SUMMARY & CONCLUSIONSBattery banks are essential to ensure the continuity of operation in telecommunication systems, acting as backup energy storage during interruptions in the primary power supply. These systems, typically composed of Valve Regulated Lead Acid (VRLA) batteries, can be monitored periodically through offline measurements of voltage, temperature, and internal resistance.In large-scale deployments, such as in some Brazilian telecommunication companies operating nationwide, over 200,000 battery banks are monitored regularly. However, failure assessment remains largely manual and conducted on a per-unit basis, making large-scale diagnostics inefficient and error-prone. This approach limits the effectiveness of a Condition-Based Maintenance (CBM) strategy and increases the risk of unexpected failures and associated operational costs.In this context, this paper presents the application and comparative assessment of different supervised machine learning techniques for predicting battery bank health based on historical condition monitoring data. A labeled dataset was used, comprising offline measurements collected between December 2019 and April 2024, along with corresponding failure records. Four techniques were implemented: Decision Tree (DT), Logistic Regression (LR), Support Vector Machine (SVM), and Random Forest (RF). All models were trained and evaluated on the same datasets. LR was trained directly on the full training set without hyperparameter tuning. In contrast, the tree-based and kernel-based models required hyperparameter optimization, which was performed using 10-fold cross-validation within the training set. Performance was assessed using accuracy, precision, recall, F1-score, and Area Under the Curve (AUC) of the Receiver Operating Characteristic (ROC).Results indicate that RF provided the best overall performance, achieving high accuracy and robustness across validation folds. LR, although less accurate, offered greater interpretability through direct analysis of coefficient weights, enabling clearer insights into the contribution of each input variable to failure prediction. While LR did not achieve the top performance, it remains the most promising candidate for initial application due to its simplicity and transparency. SVM also achieved competitive results, though they required careful parameter tuning and were more computationally demanding.From a practical perspective, the implementation of these models can significantly improve condition-based maintenance strategies by enabling early identification of batteries with higher failure likelihood. The adoption of such models allows for rapid screening of large datasets and early fault detection, potentially reducing emergency replacements, contract penalties, and risks of data center shutdowns. This enables maintenance teams to prioritize site visits and replacement actions, reducing the risk of communication outages. These findings reinforce the need for predictive tools that balance accuracy, interpretability, and operational feasibility in large-scale maintenance environments.