In　order　to　improve　the　accuracy　of　spatial　forecasting　based　on　panel　data,　the　significance　of　spatial　autocorrelation　on　the　panel　data　is　tested　by　Moran　I,　the　first-order　spatial　autoregressive　model　and　the　Kriging　algorithm　model　are　established　from　the　perspective　of　the　cross-sectional　data,　respectively,　and　back-propagation　neural　network　model　trained　by　the　genetic　algorithm　is　established　from　the　perspective　of　the　time-series　data.　Then　a　spatial　combination　forecasting　model　based　on　panel　data　is　established　by　the　three　single　models.　The　weights　are　obtained　by　the　information　entropy　approach.　An　empirical　study　shows　that　the　spatial　combination　forecasting　model　is　the　most　effective　in　the　accuracy　and　robustness.　©2009　IEEE.