Power transformer failures pose critical risks to electrical grid stability, with dissolved gas analysis (DGA) serving as the foremost non-invasive technique for incipient fault detection in oil-insulated assets. While existing deep learning approaches to DGA gas concentration prediction are constrained by single-transformer training paradigms, absence of physics-informed constraints, and inability to exploit fleet-level monitoring knowledge, this paper proposes CS-Mamba-PAF, a Contrastive Self-Supervised Mamba Network with Physics-Guided AI-Score Fusion addressing all three limitations simultaneously. The framework integrates three complementary innovations: (i) a geomagnetically-aware sub-daily-to-daily pre-processing pipeline incorporating IQR outlier clipping and STL seasonal-trend decomposition, calibrated specifically to real sub-daily online monitoring data; (ii) a physics-guided AI-Score token encoding weighted agreement of four IEC 60599:2022 diagnostic methods, namely the Duval Triangle (DTM), Rogers Ratio (RRM), IEC Ratio (IRM), and Dornenburg Ratio (DRM) as a continuous 3-dimensional IEC-compliant feature vector concatenated with 14-dimensional STL gas features; and (iii) a cross-unit SimCLR contrastive pre-training strategy with hard negative mining across 12 heterogeneous transformer units, feeding a four-layer Mamba Selective State Space Model decoder trained with a composite loss incorporating IEC gas ratio physics constraints. The framework is evaluated on the Lewis et al. (2022) dataset which is the largest open-access multi-transformer DGA archive comprising 316,203 sub-daily records from 13 UK nuclear power station transformers monitored between 2010 and 2015. On the 185-day Transformer M test set, CS-Mamba-PAF achieves
RMSE of 0.06 ppm, mean MAPE of 1.8%, and
of 0.978, representing improvements of 82.9% and 64.7% over the Gaussian Process Regression baseline in
RMSE and mean MAPE, respectively. Leave-one-unit-out (LOUO) evaluation across all 13 transformer units yields a mean RMSE of 0.118 ± 0.022 ppm with a coefficient of variation of 18.6%, confirming robust within-fleet generalisation across the heterogeneous Lewis fleet.