For the single input and single output linear system, a new method of approximately aggregated model reduction is presented by using the Moore-Penrose generalized inverse of a partial block controllability matrix as an aggregation matrix. Two aggregated order reduction models are presented firstly under the condition that the system is controllable and uncontrollable, and then a unique least error reduced model is integrated no matter whether the system is controllable or not. A simple method is deduced to compute the errors of all order reduced models, which are the distances from some known vectors to some certain subspaces. The errors of all order reduced models can be used as a model reduced order selection criterion and according to this the best order reduced model can be chosen easily. Some examples are shown to verify the validity and feasibility of this method.