Skip to main content
Back to timeline
PLoS ONESource publication:

Battery voltage prediction stops restarting each cycle: SeqBattNet pairs discrete states with aging adaptation to reach RMSE 0.0257 and 0.0253 on two NASA cells

Synopsis

SeqBattNet turns terminal-voltage prediction into a discrete-state model advanced step by step, with an encoder that estimates aging-adaptation parameters such as ohmic resistance, RC time constants and state of health from a short opening segment of each discharge cycle, and a decoder that computes voltage from an equivalent circuit model. Earlier physics-informed models either needed many hard-to-identify physical parameters or advanced continuous differential equations whose errors accumulate and which struggle to stop at the cutoff voltage. On the public TRI, RT-Batt and NASA datasets, trained on a single cell and tested on unseen cells, it reaches RMSE of 0.0257 and 0.0253 on NASA RW4 and RW5, below PDB-Net's 0.0382 and 0.0406.

Source-provided article image: SeqBattNet: A discrete-state physics-informed neural network with aging adaptation for battery modeling.
PubMed

Interpretation

Terminal-voltage prediction can now take only a short current and voltage segment at the start of each discharge cycle, derive cycle-specific aging parameters, and roll out the full discharge voltage step by step. Earlier physics-informed battery models either embedded physical constraints in the loss or combined neural networks with equivalent circuit models; the continuous-state variant of the latter needs a predefined prediction horizon, can let the voltage trajectory run past the cutoff voltage, and accumulates error over long horizons. On the NASA dataset, SeqBattNet reaches RMSE of 0.0257 and 0.0253 on RW4 and RW5, where PDB-Net reaches 0.0382 and 0.0406 and BattNN 0.0504 and 0.0524; training uses only RW3, and testing uses RW4 and RW5, which are not trained on.

The model has two parts: an encoder that estimates the aging state of the current cycle and a decoder that converts current into voltage. The encoder is a Hierarchical Reasoning-GRU, which replaces the Transformer in the hierarchical reasoning model with a gated recurrent unit to cut computation; the decoder follows a first-order equivalent circuit model and uses feedforward networks to predict open-circuit voltage and RC branch resistances. The encoder input is an opening segment of discharge current and voltage, 80 samples for TRI and RT-Batt and 30 for NASA; RMSE is 0.0280, 0.0416, 0.0346 and 0.0491 on TRI-b1c4, TRI-b1c2, RT-Batt-1-2 and RT-Batt-1-3, and 0.0257 and 0.0253 on NASA RW4 and RW5.

Removing cycle-by-cycle aging adaptation makes voltage error clearly larger under the same decoder and training settings. Without adaptation, the encoder's parameters are replaced by fixed values chosen by grid search on initial discharge cycles, while decoder architecture, training data and optimization settings stay unchanged. RMSE rises from 0.0280 to 0.0747 on TRI-b1c4, from 0.0346 to 0.1088 on RT-Batt-1-2, and from 0.0253 to 0.0395 on NASA-RW5; differences in parameter fitting prevent this contribution from being fully isolated.

Perspective

Evaluation is limited to the discharge phase of the public TRI, RT-Batt and NASA datasets, with different cells for training, validation and testing and with RMSE, MAE and MAPE as metrics; the model takes only current and voltage as inputs and does not consider temperature, and the authors plan to add ambient and cell temperature features and evaluate on real-world operating data.

The encoder's physical parameters contribute unevenly to voltage prediction, and their physical accuracy and identifiability have not been independently validated; in the no-adaptation comparison the parameter fitting differs, so the contribution of aging adaptation cannot be fully isolated; on the NASA dataset some simpler encoder variants reach slightly lower RMSE, so encoder effectiveness depends on the dataset.