Hybrid model to improve the river streamflow forecasting utilizing multi-layer perceptron-based intelligent water drop optimization algorithm

Abstract

Artificial intelligence (AI) models have been effectively applied to predict/forecast certain variable in several engineering applications, in particular, where this variable is highly stochastic in nature and complex to identify utilizing classical mathematical model, such as river streamflow. However, the existing AI models, such as multi-layer perceptron neural network (MLP-NN), are basically incomprehensible and facing problem when applied for time series prediction or forecasting. One of the main drawbacks of the MLP-NN model is the ability of the used default optimization algorithm [gradient decent algorithm (GDA)] to search for the optimal weight and bias values associated with each neuron within the MLP-NN architecture. In fact, GDA is a first-order iteration algorithm that usually trapped in local minima, especially when the time series is highly stochastic as in the river streamflow historical records. As a result, the overall performance of the MLP-NN model experienced inaccurate prediction or forecasting for the desired output. Moreover, due to the possibility of overfitting with MLP model which may lead to poor performance of prediction of the unseen input pattern, there is need to introduce new augmented algorithm capable of identifying the complexity of streamflow data and improve the prediction accuracy. Therefore, in this study, a replacement for the GDA with advanced optimization algorithm, namely intelligent water drop (IWD), is proposed to enhance the searching procedure for the global optima. The new proposed forecasting model is, namely MLP-IWD. Two different historical rivers streamflow data have been collected from Nong Son and Thanh My stations on the Vu Gia Thu Bon river basin for period between (1978 and 2016) in order to examine the performance of the proposed MLP-IWD model. In addition, in order to evaluate the performance of the proposed MLP-IWD model under different conditions, four different scenarios for the model input–output architecture have been investigated. Results showed that the proposed MLP-IWD model outperformed the classical MLP-NN model and significantly improve the forecasting accuracy for the river streamflow. Finally, the proposed model could be generalized and applied in different rivers worldwide.