Modern battery management systems have a critical need for highly accurate battery terminal voltage models, which are a key component of algorithms that estimate or predict power capability, range, temperature, and other factors. While electro-chemical and equivalent circuit models are widely used for this purpose, they typically struggle to model efficiently the complex, non-linear dynamics inherent in real-world battery operation. This study proposes a robust, data-driven approach for terminal voltage estimation using a feed-forward neural network (FNN) machine learning model. Characterization and drive cycle tests were performed on a 60 Ah prismatic cell from a Fiat 500e at temperatures ranging from -20 °C to 40 °C. The collected data was used to train and test the models, with model error reported for HWFET, UDDS, US06, and LA92 cycles. Model size was swept between around 100 and 35,000 trainable parameters for an FNN with three inputs – unfiltered power, state of charge, and temperature - to select the best size for the baseline model. Next, low pass filters were applied to measured power and used as additional inputs. The two filter frequencies (ranging from 0.1–200 mHz) resulting in the lowest error were selected. This five-input model was shown to have 42% lower error compared to the baseline (no filters) model, with an average error of just 9.8 mV which is 30 to 50% lower than values reported in literature for equivalent circuit and other machine learning battery voltage models.