The invention relates to a Prony algorithm and ill-conditioned data analysis-based power grid low-frequency oscillation on-line detection algorithm. According to the algorithm, signals obtained after preprocessing are inputted into a buffer window of 6 to 20 points; a least squares matrix coefficient is calculated; obtained matrix eigenvalues are compared with each other; whether a matrix is an ill-conditioned matrix is judged according to relationships between the eigenvalues; if the matrix is an ill-conditioned matrix, the data are filtered out, otherwise, the least squares method is adopted to obtain a matrix, and a generalized solution is further solved; the parameter of low-frequency oscillation signals is solved according to the generalized solution; and the amplitude, phase angle and attenuation coefficient of the low-frequency oscillation of a power grid can be estimated accurately in real time; and signals estimated by using the parameters can be applied to the linear prediction and suppression of low-frequency oscillation. The method has the advantages of high operability, insensitivity to signal noises, high corresponding speed and economic and practical value.