A multiple model self-tuning predictor is presented which does not require the assumption or knowing the parameter vectors in the finite set. The parameters of the submodes are recursively estimated in real-time with a parallel model identification method. The simulation example shows the satisfactory performance of the proposed algorithm. The algorithm is applied to power demand forecasting and gives the significant improved results in comparison with a self-tuning predictor, although the Markov dynamic model assumption is not exactly satisfied.