Estimation of multi-pattern-to-single-pattern functions by combining feedforward neural networks and support vector machines

In many fields there are situations encountered where a function has to be estimated to determine its output under new conditions. Some functions have one output corresponding to differing input patterns. Such types of function are difficult to map using a function approximation technique such as that employed by the multilayer perceptron networks. Hence to reduce this functional mapping to single pattern-to-single pattern type of condition, and then effectively estimate the function, we employ classification techniques such as the support vector machines. This paper describes in detail such a combined technique, which shows excellent results for a practical application in the field of power distribution systems.