The invention belongs to the technical field of battery energy storage and especially relates to an energy storage battery soft fault diagnosis method based on GA-Elman. The method comprises the following steps: carrying out acquisition and analysis on terminal voltage signals and state-of-charge signals under different degrees of attenuation; obtaining two feature vectors of terminal voltage deviation Devk and SOC punishment angle [theta]k; carrying out normalization processing on the feature vectors and serving the processed feature vectors as input of the GA-Elman neural network, and serving residual capacity Qk obtained after battery attenuation as output of the GA-Elman neural network; calculating relative inferiority degree of the Qk; determining four fuzzy subsets and membership functions thereof in a large-capacity battery energy storage system soft fault discourse domain according to the relative inferiority degree; and establishing a soft fault diagnosis model having fuzzy output and realizing fuzzy diagnosis of soft fault of the battery energy storage system. The method has higher accuracy, has battery soft fault level comprehensive judgment capability and points out a new direction for soft fault diagnosis of a large-capacity battery storage system.