The invention belongs to the technical field of new energy power generation consumption and discloses a new energy consumption method based on depth learning. The method is characterized in that a back propagation algorithm of the multi-layered perceptual neural network is utilized for realization, and training is performed through utilizing the improved dynamic recurrent neural network method; the optimization process includes the offline training process of a deep learning model, and the other part is the online optimization process of a deep learning optimization controller. The method is advantaged in that the depth learning-based consumption optimization algorithm can determine optimization controller parameters online according to the content and the quantity of different optimization targets and the content and the quantity of different constraints without the need of artificial adjustment, the method is universally adaptable for application scenes, the dynamic recurrent neuralnetwork model training method based on an adaptive update coefficient can overcome disadvantages that the conventional dynamic recurrent neural network has the slow convergence speed and is easy to fall into a local minimum value, and the training time of an optimization controller model is shortened under the premise of ensuring accuracy.