River Flow Prediction Using Dynamic Method for Selecting and Prioritizing K-Nearest Neighbors Based on Data Features

Abstract

River flow prediction is an important aspect of robust water resources planning and flood warning systems operation. Data-driven approaches have been found efficient to this end. K-nearest neighbors (KNN) is a lazy learning method that can be used for this purpose. In this study, a new method for selecting neighbors named dynamic number of k-nearest neighbor (DKNN) is introduced which uses an optimized distance to select a different number of neighbors for each instance of predictors instead of using a fixed k number as in the classic method. The particle swarm optimization (PSO) algorithm is used for the optimization process to improve the results. Three techniques for prioritizing the contributing neighbors are applied: (1) using the pattern of the predictor data, (2) considering the date of the predictor data, and (3) using both of these features in the prediction procedure. The performance of the proposed method and techniques is tested using 2 years of the daily inflow to the Gheshlagh reservoir in Iran and is compared with the results of classic KNN, artificial neural networks (ANN), random forest regression (RFR), and support vector machines (SVM). The results indicate that the proposed method increased the accuracy of prediction by 4.9% by reducing the root-mean-square error (RMSE) compared to the classic KNN. Using the recorded date of the predictor gives the best performances out of the three proposed techniques and performs better than classic KNN, ANN, RFR, and SVM by showing 49%, 38%, 31%, and 24% improvement in RMSE, respectively. Considering the pattern of the predictor and the combined technique also resulted in 12% and 35% reduction in RMSE, respectively, compared to classic KNN.