Wireless sensor network (WSN) nodes depend on small batteries with complex voltage dynamics influenced by internal electrochemical phenomena, which are further pronounced under the inter- mittent and pulse-like loads characteristic of WSN applications. Traditional battery models often fail to capture these nonlinear and time dependent effects. This thesis addresses the problem by developing a machine learning model using XGBoost to predict the full voltage response of lithium titanate (LTO) batteries under realistic intermittent discharge cycles. Extensive experiments were performed using the unique UU CoRe Battery Testbed; this resulted in over 5.8 million voltage mea- surements from 162 discharge cycles with varying current pulses, recovery periods, and duty cycles. The model was trained on the features current, state of charge, relative time within a pulse, and total pulse duration, with the target variable being voltage response. The resulting model predicts voltage with high accuracy, achieving an average RMSE of 0.0134 normalized volts (1.89% error) on unseen test data. Performance decreases toward the end of discharge (final 5%), where the error increases to 0.0325 normalized volts (7.50%), likely due to a lack of training data in that region. By accurately modeling the detailed voltage response to intermittent loads, this work creates a syn- thetic battery model that replicates real battery behavior under realistic operating conditions. This synthetic model provides a powerful tool for optimizing battery lifetime in WSN and IoT applica- tions. Furthermore, this study demonstrates the viability of machine learning for capturing complex battery dynamics and lays a foundation for future research applying machine learning techniques to battery modeling and management across different use cases.