Abstract
<title>Abstract</title> <p>Valve-regulated lead-acid (VRLA) batteries remain the dominant stationary storage technology across much of the developing world — telecom towers, inverter backup and solar home systems — yet prognostics research is concentrated almost entirely on lithium-ion. A 2023 systematic mapping study of lead-acid prognostics reported that no public dataset could be found, and errors reported on private cohorts of 3–70 cells (0.0005–7.6%) are mutually incomparable. Using the Oxford/BBOXX release of 1,027 field-deployed 12 V/20 Ah VRLA batteries on solar home systems in sub-Saharan Africa (approximately 620 million rows; 400–760 days per unit; 491 observed failures; 536 right-censored units), this work establishes, to our knowledge, the first openly specified and reproducible benchmark for lead-acid field prognostics. The central methodological finding is that, with 52% right censoring, the task is substantially better posed as survival analysis than as regression: a bidirectional Mamba regressor for remaining useful life attains 83.3 d mean absolute error — worse than the one-line heuristic mean lifetime minus age (63.8 d) and worse than a ridge regression (57.2 d) — whereas landmark survival analysis over 6,003 battery-landmark observations with 97 level, trend and volatility features attains AUC 0.848 (95% CI 0.819–0.877) at a four-week horizon and 0.832 (0.808–0.855) at twelve weeks on the at-risk population, with Uno's IPCW concordance of 0.737–0.756 across five estimators (best 0.756, Coxnet). Replicating the protocol and metric of Aitio and Howey (Joule, 2021) on the same dataset, balanced accuracy reaches 84.1% eight weeks before failure (published: 73%) and 89.7% at the point of failure (published: 82%), using only daily aggregates and no state-of-health estimate. A five-model ablation indicates that, within the at-risk population, battery age alone is uninformative (balanced accuracy 52.5–57.5%) while telemetry alone recovers the full model (77–89%). Three negative results are reported in full, and code, features and deterministic splits are released.</p>