The causes of long-period error of a small photoelectric encoder and its distribution law were researched and a correction method for the long-period error of the small photoelectric encoder was proposed. A Fourier neural network error correction model was built firstly based on orthogonal trigonometric functions, and the nonlinear optimization problem between the input and output of the encoder was transformed to a linear optimization problem. By taking the output value of the high-accuracy benchmark encoder as the learning reference for the neural network model, an improved differential evaluation algorithm combined with simulated annealing strategy was applied to training of the neural network and to ensuring its global optimization search ability in the initial stage but fast convergence rate and high accuracy in the later period.The method was applied to the long period error correction test of a 16-bit small photoelectric encoder, and experimental results show that the peak errors of the encoder is reduced from 45″—-175″ to 10″—-875″ and the standard deviation of long-period errors is reduced from 203″ to lower than 4″.The results mean that the proposed long-period error correction method improves the accuracy of small photoelectric encoders.