Streamflow Prediction Based on Least Squares Support Vector Machines

We used the least squares support vector machine (LS-SVM) algorithm to forecast the future streamflow discharge using the past streamflow data and gage height, which is closely related to regularization networks and Gaussian processes.A Gaussian Radial Basis Function (RBF) kernel framework was built on the data set to optimize the tuning parameters and to obtain the moderated output.The prediction capability with different kernel functions was also explored to evaluate the impact of the streamflow discharges and gage height for long-term prediction of flow rates and its accuracy.The training process of LS-SVM involves the selection of both kernel parameters and regularization constants.The tuning parameters were found by using a combination of coupled simulated annealing (CSA) and a standard simplex method.The CSA finds good starting values and these values were passed to the simplex method in order to fine tune the result.One of the parameters, is the regularization parameter, determining the trade-off between the training error minimization and smoothness.The other parameter, represents the squared bandwidth.The USGS real-time water data at various locations were used as time series input.50% of the data were used for training, and 50% were used for testing.The experimental results showed that the LS-SVM algorithm is a reliable and efficient method for streamflow prediction, which has an important impact to the water resource management field.