While BP neural network,classical theory of probability and its derivative on algorithm were used to fire loss prediction,the system structure is complex,the detection data is not stable,and the result is easy to fall into local minimum, etc. To resolve these troubles, a method of reasoning and forecasting the regional fire data was proposed based on adaptive fuzzy Generalized Regression Neural Network (GRNN). The improved fuzzy C-clustering algorithm was used to correct weight for the initial data in network input layer, and it reduced the influence of noise and isolated points on the algorithm, improved the approximation accuracy of the predicted value. The adaptive function optimization of GRNN algorithm was introducd to adjust the expansion speed of the iterative convergence, change the step, and found the global optimal solution. The method was used to resolve the premature convergence problem and improved the search efficiency. While the identified fire loss data is put into the algorithm, the experimental results show that the method can overcome the problem of instable detection data, and has good ability of nonlinear approximation and generalization capability.