Accurate fuzzy predictive models through complexity reduction based on decision of needed fuzzy rules

Abstract

Abstract Fuzzy predictive models (FPMs) are employed in a wide variety of domains for their high prediction accuracy and good interpretability. For a FPM, the fuzzy rules from data or expert experience are always firstly extracted to establish a relatively complete fuzzy rule base. However, it is found that a large amount of extracted fuzzy rules which are always not used for prediction increase the complexity of the FPM. Moreover, inaccurate fuzzy rules may be used for prediction to degrade the accuracy of the FPM. Therefore, we proposed a method to establish an accurate FPM through complexity reduction based on decision of needed fuzzy rules. To avoid extractions of the rules which are not needed for prediction, reduce the complexity of the FPM, and suppress the influence of some inaccurate fuzzy rules, we add the decision of fuzzy rules needed for prediction based on membership values to predictive modeling before the extraction of fuzzy rules. And then an improved WM method based on optimization of centers of output fuzzy subsets for fuzzy rules (COiWM), which was proven to possess high completeness, robustness, and accuracy in our previous work, is used to generate the needed fuzzy rules directly from historical data. The rules which can not to be generated are extrapolated based on these data-generated fuzzy rules. At last, two fuzzy rule bases are built and used together to fuzzy predictive modeling. Experimental results of a case study on short term daily maximum electric load forecasting and Mackey–Glass chaotic time-series prediction show that our proposed method reduces the complexity of the FPMs and enhances the prediction accuracy.