Encoder-Decoder Neural Network for Battery Voltage Simulation

The patent replaces conventional physical modeling approaches with machine learning models (LSTM, GRU, Transformer architectures). These data-driven models learn patterns directly from historical battery data without requiring explicit physical equations, thereby improving prediction accuracy while eliminating the need for complex initial state assumptions and physical parameter definitions.

The patent transforms the modeling approach by changing from fixed physical parameters to adaptive learned parameters. The machine learning models automatically adjust their internal parameters during training to capture complex battery behaviors, allowing accurate future behavior simulation without relying on predetermined physical model structures or initial state guesses.

The patent segments the battery system into individual storage units, each with its own dedicated machine learning model. This segmentation allows each model to be trained on specific unit data, improving adaptability while keeping individual model computations manageable. The modular architecture enables parallel processing and reduces overall computational burden compared to a single large model.

The patent performs preliminary training of machine learning models during offline phases using historical battery data. This preliminary action allows the models to learn complex patterns and adapt to specific battery characteristics before deployment. During real-time operation, the pre-trained models require only inference computations, significantly reducing online computational complexity while maintaining high adaptability.