Ablation Analysis and Optimization of a Dual-Branch Time-Frequency Network for Regional Ionospheric TEC

We run ablation experiments on a dual-branch time-frequency network for regional TEC forecasting over China. Hourly TEC from 260 grid points (2009-2015) is used. The time branch represents the TEC history via patch MLP; external variables are embedded separately and then mixed by multi-head attention across variables. The spectral branch works on the 48-h lag residual of TEC, and RevIN is applied only to TEC. We compare MV-time and PatchTST with and without covariates. With covariates, MV-time wins at 195 of 260 sites; without covariates, the split is 132 vs 128. Thus the gain comes from covariate fusion. The input sweep selects five solar/seasonal variables and keeps Dst/Kp only for sites outside the mid-latitude band. A spatial residual MLP, trained on validation errors, corrects each forecast using neighbour statistics, location, and lead time. The full pipeline reduces mean MAE from 3.146 to 3.085 TECU and improves 247 out of 260 sites over the 9-variable baseline. Integrated Gradients is kept as a post-hoc consistency check.