SwinV2-KAN Multisource Fusion Model for Satellite Solar Array Degradation Estimation

Reliable degradation estimation is required to preserve the power margin of long-life spacecraft exposed to radiation, thermal cycling, contamination, illumination changes, and interconnect ageing. This study proposes a SwinV2-KAN multisource fusion model that combines solar-wing thermal imagery, bus telemetry, I-V descriptors, orbital-environment records, and operation-mode tags to estimate normalized degradation and remaining power capability. The inputs are aligned at the orbital-cycle level, while availability masks and quality indicators prevent missing measurements from being interpreted as nominal observations. Swin Transformer V2 extracts hierarchical thermal tokens, a gated temporal encoder represents electrical and environmental sequences, and a reliability gate suppresses unavailable or low-quality sources before Kolmogorov-Arnold Network fusion. The model returns a point estimate, calibrated interval bounds, source weights, and quality flags for engineering review. Evaluation uses a simulated but engineering-constrained archive of 2,160 orbital cycles covering six representative degradation modes and five source groups. Comparisons include statistical regressors, recurrent and convolutional fusion models, a vanilla Transformer, and SwinV2-MLP fusion. SwinV2-KAN achieves an MAE of 0.018, an RMSE of 0.027, an early-warning lead time of 11.6 days, and 93.8% empirical coverage for a nominal 95% interval. Removing KAN fusion increases MAE by 28.4%, while removing thermal-image tokens reduces hot-string sensitivity by 31.7%. These results indicate that quality-aware spatial-temporal fusion can support power-budget review and maintenance-oriented analysis, although validation with independently labeled flight data remains necessary.