In this paper, in order to determine the optimal radial power system structure rapidly we propose a constrained noise approach with the Hopfield model, which can avoid local minima. When a radial power system has a number of connected feeders, the combinational number for possible system structures become too many. Determination of the optimal system composition from a great number of combinations is a combinational optimization problem, and it has ever been difficult to solve this type of problem quickly with the conventional digital computer techniques so far. However, because of the modelling of excellent parallel processing ability of neurons, it has become possible to carry out the optimization with the Hopfield neural network model. The constrained noise we propose has an ability of satisfying constraints. In addition, we compare the poposed method with a conventional branch-and-bound method which exists in a field of mathematical programming area. Simulations have been carried out for 10 and 21 substation systems successfully.