Abstract: The key of residual-based outlier detection algorithm depends on the accurate prediction. To solve the problem, a prediction algorithm based on the improved RBFN (radial basis function network) is proposed for time series of process industries. In the algorithm the input form of the improved RBFN is changed so as to enable the improved RBFN to be more convenient for introducing the forgetting and penalty factors into it, thus adapting itself to the dynamic performance/robustness required by the residual-based outlier detection. The equivalence between the improved RBFN and conventional RBFN has been proved theoretically, and the comparative test results revealed that the former is simpler than the latter in network architecture with more definite attributes of parameter concerned.