Comparison of evolutionary algorithm to neuro-fuzzy and fuzzy clustering for electrical load time series forecasting

Abstract EN

The Neuro-Fuzzy Network and Fuzzy Clustering system are known as good algorithm in time series forecasting application but the parameters need to be optimized.

As the Evolutionary Algorithm (EA) grows, it is interesting to implement EA to optimize the parameters of Neuro-Fuzzy and Fuzzy Clustering.

The time series data is the electrical load of East Java-Bali, Indonesia, in 2005-2007.

The results of unoptimized parameters for both algorithms were not satisfied.

Therefore, the EA is implemented to optimize their parameters.

The EA used Real Code Genetic Algorithm.

The parameters optimized with EA are mean and variance of the Gaussian MF with pc = 0.6, pm = 0.1 and 20 chromosomes per population.

The experiments without optimization showed that MSE LTF for Fuzzy Clustering is 2.9x10-3 and Neuro-Fuzzy is 2.0 x 10-3.

MSE of long time forecasting (LTF) for optimized Fuzzy Clustering is 2.42 x 10-3 while for Neuro-Fuzzy Network is 1.9 x 10-3.

These results indicated that the EA has no effect for Neuro-Fuzzy because the difference is small.

For Fuzzy Clustering, the result is interesting.

But it is still not very satisfying.

This project still needs improvements in order to get more satisfying results.