Ultra-low-power (ULP) embedded AI systems increasingly rely on coin-cell batteries, yet accurate modeling of battery discharge behavior under duty-cycled workloads remains a challenge. This research develops a data-driven framework to predict voltage–capacity discharge curves from experimentally collected waveform data. By applying smoothing techniques and piecewise curve fitting, the model decomposes each discharge curve into linear and exponential regions, enabling interpretable analysis of how waveform parameters affect voltage behavior over time. To capture variations across different waveforms, we extract features such as average current and duty cycle, and evaluate multivariate regression models to predict curve shape. This approach supports power-aware optimization in ULP systems by enabling fast, flexible prediction of battery behavior under diverse operating conditions, laying the groundwork for future integration with energy-efficient AI accelerators, allowing for runtime scheduling and architectural co-design that jointly optimize accuracy, performance, and battery life.