| _version_ | 1866901077426700288 |
|---|---|
| author | Pasupuleti, Ramakrishna |
| author_facet | Pasupuleti, Ramakrishna |
| contents | <p>This paper presents the K–R framework, a physics-informed residual learning approach for nonlinear battery charging optimisation. The framework decomposes the charging control signal into two orthogonal components: a physics-based K step that inverts the equivalent circuit model to compute baseline charging current, and a data-driven R step that corrects structured residual errors arising from temperature effects, electrochemical ageing, and model mismatch. Unlike conventional deep learning methods, the R step is trained via closed-form least squares from online measurements, requiring no backpropagation, no offline dataset, and no parameter re-identification. Comprehensive simulations calibrated to three internationally recognised battery datasets (NASA Ames Prognostics, CALCE, and Oxford Degradation) across eleven test suites demonstrate three principal findings. First, the K–R controller prevents catastrophic voltage overshoot under extreme cold-start conditions at minus ten degrees Celsius, reducing peak overshoot from 2650 millivolts under conventional constant-current constant-voltage charging to below 363 millivolts. Second, K–R maintains zero voltage overshoot across 500 ageing cycles while conventional charging becomes unsafe after 300 cycles, demonstrating inherent adaptability to ageing-induced model drift without recalibration. Third, the K-step residual exhibits strong temporal autocorrelation of 0.956 at lag-one, empirically validating the theoretical hypothesis that physics model errors are deterministic and learnable rather than stochastic noise. Additional validation confirms that the framework generalises across nickel manganese cobalt, lithium iron phosphate, and high-power cell chemistries without structural modification, maintains closed-loop stability under temperature disturbances and intermittent electric vehicle charging profiles, remains robust under parameter uncertainty up to fifty percent model mismatch, and degrades gracefully under realistic sensor noise and analogue-to-digital converter quantisation. The K–R decomposition establishes a universal, interpretable, and deployment-ready paradigm for intelligent battery management systems, bridging physics-based control and data-driven intelligence through the principled exploitation of structured residual errors.</p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_19225231 |
| institution | Zenodo |
| language | |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | Physics-Informed Residual Learning for Safe and Adaptive Battery Charging Under Extreme Conditions Pasupuleti, Ramakrishna <p>This paper presents the K–R framework, a physics-informed residual learning approach for nonlinear battery charging optimisation. The framework decomposes the charging control signal into two orthogonal components: a physics-based K step that inverts the equivalent circuit model to compute baseline charging current, and a data-driven R step that corrects structured residual errors arising from temperature effects, electrochemical ageing, and model mismatch. Unlike conventional deep learning methods, the R step is trained via closed-form least squares from online measurements, requiring no backpropagation, no offline dataset, and no parameter re-identification. Comprehensive simulations calibrated to three internationally recognised battery datasets (NASA Ames Prognostics, CALCE, and Oxford Degradation) across eleven test suites demonstrate three principal findings. First, the K–R controller prevents catastrophic voltage overshoot under extreme cold-start conditions at minus ten degrees Celsius, reducing peak overshoot from 2650 millivolts under conventional constant-current constant-voltage charging to below 363 millivolts. Second, K–R maintains zero voltage overshoot across 500 ageing cycles while conventional charging becomes unsafe after 300 cycles, demonstrating inherent adaptability to ageing-induced model drift without recalibration. Third, the K-step residual exhibits strong temporal autocorrelation of 0.956 at lag-one, empirically validating the theoretical hypothesis that physics model errors are deterministic and learnable rather than stochastic noise. Additional validation confirms that the framework generalises across nickel manganese cobalt, lithium iron phosphate, and high-power cell chemistries without structural modification, maintains closed-loop stability under temperature disturbances and intermittent electric vehicle charging profiles, remains robust under parameter uncertainty up to fifty percent model mismatch, and degrades gracefully under realistic sensor noise and analogue-to-digital converter quantisation. The K–R decomposition establishes a universal, interpretable, and deployment-ready paradigm for intelligent battery management systems, bridging physics-based control and data-driven intelligence through the principled exploitation of structured residual errors.</p> |
| title | Physics-Informed Residual Learning for Safe and Adaptive Battery Charging Under Extreme Conditions |
| url | https://doi.org/10.5281/zenodo.19225231 |