Prediction of time varying composite sources by temporal fuzzy clustering

We present a method for predicting non-stationary signals generated by a time varying composite source. The method is based on the concept of temporal fuzzy clustering. A fuzzy clustering algorithm is applied to the given part (past+present) of the time series and the calculated clusters and membership matrix are then used to estimate a mixture probability distribution function (PDF) underlying the series. In this way a continuous drift in the series distribution expressed as a drift in the clusters' appearance rate can be estimated. A future PDF can then be predicted by fitting a specific model to the estimated past and future PDF values. This also enables the generation of a minimal-mean-squared-error prediction for a future time series element using the estimated mean value of the predicted PDF.