The invention discloses an extreme learning machine-based power grid harmonic voltage signal detection method. The method comprises the following steps that: 1, distorted voltage signals are chosen randomly, and the amplitude of n-th harmonics (n is an odd number) is increased sequentially in a proportion of 10% from zero until the amplitude of the n-th harmonics achieves 50% of the amplitude of fundamental waves, and therefore, a reasonable training sample can be obtained; and step 2, voltage to be detected in distorted voltage is selected as the input quantity of an extreme learning machine model, and the Fourier coefficient of the distorted voltage signals is adopted as the output quantity of the model, wherein the number of input layer nodes of the above extreme learning machine model is the number of the categories of voltage signals to be detected, and the number of output layer nodes is four, and after training is completed, an extreme learning machine detection model of harmonic wave voltage signals to be detected can be obtained; and step 3, the above prediction model is utilized to detect power grid harmonic wave voltage signals. According to the extreme learning machine-based power grid harmonic voltage signal detection method of the invention, the feature of nonlinear function approximation ability of the extreme learning machine is utilized to detect the power grid harmonic voltage signals, and therefore, detection accuracy and a real-time property can be improved.