Abstract: Power consumption in embedded systems is becoming a hot issue that attracts more attention. Different assemble instruction set, software algorithm, and high-level software architecture can significantly affect the system energy consumption. In this paper, we firstly analyze the relations between software energy consumption and some software characteristics on algorithm level. Through measuring three software characteristics, i.e., average time complexity, space complexity, and input scale, we propose a BP neural network software power model based on algorithm complexity. Then, we design and train a kind of BP neural network to accomplish energy consumption function approximation. Simulation experiment results show that the error between the estimation value of this energy consumption function and the real energy consumption value is below 10%. Therefore, it could quickly estimate the energy consumption of software in some input scale, which is an important fundament to explore the energy consumption optimization in the future.