When regression is carried out in a prediction purpose, one of the main assumptions is the absence of evolution in the modeled phenomenon between the training and the prediction stages. Unfortunately, this assumption turns out to be often false in practical situations. The present work investigates the estimation of regression mixtures when population has changed between the training and the prediction stages. The main idea of this work is to link the regression mixture of the prediction population with the known regression mixture of the training population. For this, two approaches are suggested. On the one hand, a parametric approach modeling the relationship between dependent variables of both populations is presented and the EM algorithm is used for the parameters estimation. On the other hand, a Bayesian approach is also proposed in which the priors on the prediction population depend on the mixture regression parameters of the training population. In this latter case, a MCMC procedure is used for inference. The relevance of both the parametric and the Bayesian approaches is illustrated on simulations and then compared to classical strategies on an environmental dataset.