RENFIS-MA: A Recurrent Neuro-Fuzzy Inference System for Time Series Water Level Prediction
Annotatsiya
It is a well-known fact that improving time series prediction accuracy is an important and challenging issue. Exten-sive research has been conducted using soft-computing techniques to improve their prediction accuracy. In this paper, we propose a new methodology to predict time-series water-level, namely, recurrent error-based neuro-fuzzy inference system with moving average (RENFIS-MA). In particular, we used the adaptive network-based fuzzy inference system to build a prediction model. RENFIS-MA realizes that with the lagged variables of a time series feeding back network output-error to the input layer and moving-average of auto-regressive input making accurate predictions. It can identify the system's characteristics quite well and provides a new way to make time-series predictions. A comparative analysis of RENFIS-MA with four soft-computing models is conducted using the prediction of the water level of the river of Jamuna in Bangladesh. The trained networks are used for predicting both single-step and multiple-step ahead. In the simulation, two performance measures [root-mean-square error (RMSE) and prediction accuracy (Acc.)] demonstrate that the proposed method is more effective and accurate for prediction among all the compared models.