Embedding Relevance Vector Machine in Fuzzy Inference System for Energy Consumption Forecasting

This paper introduces a Bayesian framework to gain sparse solutions in forecasting tasks. In this study, a local identification method based on the Takagi-Sugeno relevance vector machine (TS-RVM) for nonlinear time series forecasting is introduced. The core idea is applying a set of nonlinear models, i.e. local RVM models, as the consequent part of the fuzzy rules. In this method, at first, the fuzzy rules are created based on clustering techniques and the parameters of the rules' premise are tuned. Then, the parameters of each local RVM model are determined in a learning process in a Bayesian framework. It is shown that by utilizing a probabilistic Bayesian learning framework, we can achieve very accurate prediction models. One of the benefits of this model is that it typically uses fewer basis functions in comparison with support and relevance vector machines. Also, it includes automatic estimation of nuisance parameters, and the ability to use arbitrary basis functions (e.g. non mercer kernels). Second, in comparison with global RVM, proposed model have better result in terms of prediction error. Third, it adds in the benefits of probabilistic predictions, providing a probability distribution for predictions, while other function approximators like SVM and ANN are just point estimators.