Building complex digital circuit power models is a popular approach for estimating the average power consumption without detailed circuit information. In the literature, most power models must increase in complexity to meet the accuracy requirement. The authors propose a novel power model for complementary metal-oxide-semiconductor sequential circuits using recurrent neural networks to learn the relationship between the input/output signal statistics and the corresponding average power dissipation. The complexity of our neural power model has almost no relationship to the circuit size and the number of inputs, outputs and flip-flops such that this power model can be kept very small, even for complex circuits. Using such a simple structure, the neural power models can still have high accuracy because they can automatically consider the non-linear power distribution characteristics and temporal correlation of the input sequences. The experimental results have shown that the estimations are still accurate with smaller variations even for short sequences with only 50 pattern pairs.