Condition-Aware ARIMA Cleaning of Substation Equipment Time Series with Local Residual Scaling and Audit Records

Distinguishing data quality defects from genuine operating-condition changes is a central challenge in cleaning substation equipment time series. This study extends the double-loop AO/IO framework of Yan et al. by adding local residual scaling, retain-first state screening, constrained correction, model-conditional prediction intervals, and record-level audit fields; the double-loop structure itself is not claimed as new. Evaluation used six deliberately selected field-derived series (two from each of three data sources) with reproducible synthetic labels. The prior correction cap, which allowed corrections to at most 5% of samples, was disabled. Across 18 representative paired runs and 516 injected points, the proposed method obtained F1 = 0.9647 ± 0.0337, whereas fixed-threshold AO/IO obtained 0.9728 ± 0.0308; the six-series paired comparison did not establish a difference (t-test p = 0.0707; Wilcoxon p = 0.1250). In a constructed heteroscedastic stress test, mean F1 increased from 0.4675 to 0.4941 and low-volatility recall from 0.1458 to 0.2667, but this scenario was designed around variance changes and contains only six independent source series. A second spike-only test retained the naturally varying volatility of measured segments and produced F1 = 0.4835 versus 0.4345, but the series-clustered difference was not significant (p = 0.4489). The evidence therefore supports volatility adaptation and traceable decision output, not universal accuracy superiority or field-proven effectiveness. Prediction intervals are approximate and conditional on the fitted model.