Forecasting Energy Consumption Using Fuzzy Transform and Local Linear Neuro Fuzzy Models

Abstract

This paper proposes a hybrid approach based on the local linear neuro-fuzzy (LLNF) model and fuzzy transform (F-transform), termed FT-LLNF, for the prediction of energy consumption. LLNF models are powerful in modeling and forecasting highly nonlinear and complex time series. Starting from an optimal linear least square model, they add nonlinear neurons to the initial model as long as the model's accuracy is improved. Trained by the local linear model tree learning (LOLIMOT) algorithm, the LLNF models provide maximum generalizability as well as outstanding performance. Besides, the recently introduced technique of fuzzy transform (F-transform) is employed as a time series pre-processing method. The technique of F-transform established based on the concept of fuzzy partitions, eliminates noisy variations of the original time series, and results in a well-behaved series that can be predicted with higher accuracy. The proposed hybrid method of FT-LLNF is applied to the prediction of energy consumption in the United States and Canada. The prediction results and comparison to optimized multi-layer perceptron (MLP) models and the LLNF itself, revealed the promising performance of the proposed approach for energy consumption prediction and its potential usage for real-world applications.