Aiming at the problems of poor accuracy and large calculation amount in the existing battery bank SOC (State of Charge) estimation, the invention provides a battery bank SOC online estimation method based on feature parameter selection and Gaussian process regression (GPR). The method comprises the following steps: before a power battery bank leaves a factory, carrying out charging and dischargingtesting on the power battery bank, acquiring measured data in the charging and discharging processes, and calculating an actual SOC curve of the battery pack by an ampere hour integration method; calculating the correlation coefficients of current, total voltage, monomer voltage, monomer temperature and SOC respectively, removing data with the correlation coefficients lower than a certain threshold, and constructing an input data set; carrying out principal component analysis on the input data set, and only keeping k principal components; selecting an exponential square function to serve as akernel function, and carrying out model training by using the obtained sample data; and importing the acquired data into a trained GPR model for battery pack SOC estimation during use of a power battery system.