The invention provides an RSA method combining service prediction and spectrum conversion in an elastic optical network, which is used for solving the problems of dynamic routing and spectrum distribution in the elastic optical network and comprises the following implementation steps of: firstly, selecting K candidate paths for all node pairs of the network; predicting service request change information on the link, and calculating the comprehensive weight of each candidate path; secondly, judging whether a service request can be processed or not according to a frequency spectrum constraint condition; and finally, when all the candidate paths do not meet the spectrum constraint condition, adding a spectrum converter to process the service request. According to the invention, a radial basisfunction neural network RBFNN prediction technology is used; during routing, the path length and the spectrum occupation condition are considered, the competition degree of the service request duration time on the candidate path is also considered, the network performance can be effectively improved in combination with spectrum conversion, and an efficient resource optimization allocation schemeis provided for requirements of operators or data center interconnection and the like.