Abstract

This paper introduces a novel approach to identifying fuzzy prediction intervals using applied statistics concepts. The method is a two-stage process, with the first stage defining the nonlinear deterministic component of the process using a local linear neuro-fuzzy model (LLM). The second stage uses an LLM with respect to inputs to derive the residual model, partitioning it based on the distribution of uncertainties. Based on a finite set of measured data, this approach defines a band that encompasses all future output measurements with lower and upper fuzzy bounds. This technique is well-suited for describing a family of uncertain nonlinear functions with nonnormal noise or uncertainties that do not follow a normal distribution. Based on this advantage, the two-stage fuzzy interval model can be applied to process monitoring, fault detection, and robust control design. For example, the proposed method can forecast wind power, which exhibits highly nonlinear stochastic behavior. Particularly in the example presented in this work, the wind power data originate from an isolated microgrid in Huatacondo, Atacama Desert, Chile, and are used for one-day wind power forecasting to aid the microgrid's energy management system.