This paper starts at a software architecture level, considers the functional relation between software characteristic quantities and embedded software energy as nonlinear (linear functional relation can be considered as a special nonlinear functional relation). Next, the paper presents an energy model at architecture level by using BP neural network. The energy model measures 5 software characteristic quantities at architecture level and uses BP neural network to fit the functional relation between software characteristic quantities and embedded software energy. Experimental results show that this model is effective.