Aluminum electrolysis full-process energy-saving scheduling method fusing deep reinforcement learning algorithm

Abstract

The present application relates to the field of computer systems based on specific calculation model, disclose a kind of aluminum electrolysis full-process energy-saving scheduling method of fusion deep reinforcement learning algorithm, comprising: obtaining data stream from data buffer and constructing space-time associated state tensor;Noise is eliminated by sliding window space-time covariance matrix to generate filter state tensor;Input policy network, forward inference outputs control instruction to register;By endogenous check program, the action characteristic distribution variance is calculated according to discrete control quantity and returned in situ to be used as time window trigger condition;According to global energy efficiency slope, calculate the difference reward value and introduce gradient step length attenuation operator to modify network parameters, the present application breaks the rigid binding of calculation time sequence and sampling, avoids parameter divergence caused by high-frequency noise, improves resource scheduling efficiency and convergence stability.

Images