PREDICTION OF MULTIVARIATE TIME SERIES DATA USING ECHO STATE NETWORK AND HARMONY SEARCH

Abstract

Multivariate time series data prediction is widely applied in various fields such as industry, health, and economics. Several methods can form prediction models, such as Artificial Neural Network (ANN) and Recurrent Neural Network (RNN). However, this method has an error value more significant than the development method of RNN, namely the Echo State Network (ESN). The ESN method has several global parameters, such as the number of reservoirs and the leaking rate. The determination of parameter values dramatically affects the performance of the resulting prediction model. The Harmony Search (HS) optimization method is proposed to provide a solution for determining the parameters of the ESN method. The HS method was chosen because it is easier to implement, and based on other research, the HS method gets the optimum value better than other meta-heuristic methods. The methods compared in this study are RNN, ESN, and ESN-HS. Root Mean Square Error (RMSE) and Mean Absolute Percent Error (MAPE) are used to measure the error rate of forecasting results. ESN got a smaller error value than RNN, and ESN-HS produced a minor error value among the other trials, namely 0.782e-5 for RMSE and 0.28% for MAPE. The HS optimization method has successfully obtained the appropriate global parameters for the ESN prediction model.