Monthly Mean Streamflow Prediction Based on Bat Algorithm-Support Vector Machine

摘要

Accurate and reliable prediction of runoff generation is necessary for flood control scheduling, water supply planning, and hydropower generation. Support vector machine (SVM), which is at the forefront of current research of regression and classification, was used in this paper to conduct monthly mean streamflow prediction. A novel heuristic optimization named bat algorithm (BA) was introduced to determine the parameters of SVM [penalty parameter (C) and kernel parameter (η)], in which the initial fitness was supposed to be equal to the initial loudness for all bats. In order to evaluate the effectiveness of the proposed approach, monthly mean streamflow from 1952 to 2011 of Yichang station in the middle reaches of the Yangtze River were trained and tested. In the meantime, the given data set was also modeled using artificial neural networks (ANN) and cross validation-based SVM. The comparison results indicate that the proposed model (bat algorithm-based SVM) is more accurate compared with both ANN and cross validation-based SVM. However, two main shortages exist, i.e., time-consuming and relatively low accuracy in the break points of continued dry (wet) years. To relieve these shortages, local optimization algorithms [e.g., differential evolution (DE) algorithm, immune algorithm (IA), and genetic algorithm (GA)] were suggested to be combined with the bat algorithm to produce the initial population. Modifications of the stochastic term of the local search were also useful.