Most of the approaches aiming at the reduction of the power flow computation time approximates the Jacobian matrix. Thus itt degrades the convergence compared to the conventional Newton method. This paper proposes a new approach for reducing the processing time by considering the fact that half of nodes in real power systems arc floating nodes to be able to remove. In the conventional reduced matrix approach where the floating nodes are removed, the sparsity is being lost. The method in this paper does nott remove all the floating nodes but. keeps sortie nodes by using a optimal criteria for keeping the sparsity. The criteria is to indicate the minimum number of elements in the reduced matrix. This method has been applied to a 1000-node test. system. It. was verified the number of elements of the Jacobian has been reduced to about a half against the conventional matrix. And computation time has been remarkably improved without sacrificing the convergence characteristics for the power flow computation.