To develop ensemble or multi-modeling technologies for optimization of forecasting model parameters. The proposed forecasting model in the form of a recurrent penalty spline (P-spline) has several adjustable parameters, which ensures the adaptability of the model for predicting the behavior of a time series based on its retrospective values. Creating an ensemble of models in the class of variable P-spline parameters is the initial information for longitudinal clustering of time series. Such an approach allows us to estimate the centers of clusters along the time axis, which correspond to the optimal values ​​of the model parameters on the selected time series segment in real time. This made it possible to increase the efficiency of the recurrent P-spline as a real-time forecasting model, reduce computational costs and increase the performance of forecasting algorithms.