Wi-Fi-based　fingerprint　localization　plays　a　crucial　role　in　indoor　localization　services,　where　the　collection　of　fingerprint　localization　information　by　mobile　clients　poses　privacy　risks　during　data　transmission.　Recently,　Federated　Learning　(FL)　has　been　employed　for　training　fingerprint　localization　models　without　data　sharing.　However,　the　non-independent　identical　distribution　characteristics　of　Wi-Fi　fingerprint　data　result　in　suboptimal　performance　of　the　aggregation　model　in　FL.　In　this　letter,　we　propose　a　transformer-based　Coef-Feature　Alignment　Federated　Learning　(TCFAFed)　method　to　enhance　FL　performance.　An　adaptive　aggregation　approach　is　devised　to　dynamically　obtain　the　optimal　model　and　address　client　drift　issues.　The　coefficient　of　variation　is　utilized　to　improve　the　calculation　accuracy　of　the　aggregation　coefficient,　while　feature　alignment　constraints　are　imposed　to　restrict　local　feature　representation.　Extensive　experiments　demonstrate　that　our　proposed　method　achieves　superior　global　localization　accuracy　compared　to　classical　FL　aggregation　methods.