Physics-Guided, Temperature-Coupled Universal Differential Equations for Grid-Connected Battery Energy Storage

A grid-connected battery system is a closed loop: a parameter that drifts as hardware ages corrupts all of it. Equivalent-circuit models stay interpretable but cannot track that drift; neural surrogates degrade outside the training regime and return no physical parameter. We embed two bounded-output networks in the full eight-state coupled physics and identify nine physical parameters jointly with the weights, scoring every model beyond its training horizon. The hybrid cuts forecast error 9.5× over nominal physics and 43× over a Neural ODE, in 657 parameters at 36 ms per forecast on one CPU core. Five of nine parameters are recovered to within 13.1% of truth at cross-seed CV below 4%; re-fitted from scratch, the same architecture also improves on a second, structurally different plant; and cell-temperature RMSE falls 91% once temperature enters as a ninth state. All results are synthetic.