ALTERNATE STRUCTURES FOR ADAPTIVE TIME SERIES MODELLING

Abstract In many signal processing and control applications, the real-time identification of an effective parametric model for an observed time series is of paramount importance. Accurate modelling in a non-stationary environment often requires a large order model. As the order of the model increases, however, the computational demands in terms of throughput and precision increase dramatically. Therefore, it is essential to implement the identification algorithms on efficient structures to optimize the cost of real-time modelling. Some limitations for commonly used model structures such as direct form and lattice are discussed. A number of alternate parallel and cascade structures for adaptive time series modelling are analyzed and compared. The consequences of using these structures with existing algorithms like LMS or weighted least squares are investigated in terms of numerical accuracy, computational complexity, and tracking capability. The parallel structure is identified as a promising alternative for large order time series modelling.