The invention discloses a sparse representation method for the on-line data collection of power, and the method comprises the following steps: (1) building a sparse coding mode, and initializing a zero vector; (2) randomly sampling a data sample for iteration; (3) updating a sparse coding coefficient; (4) calculating an approximation error; (5) judging the algorithm convergence. The method is used for the research of big data of power based on a random gradient descent algorithm, can quickly and effectively process data, greatly improves the solving efficiency, also can effectively solve a sparse coding coefficient, achieves the compression of big-data sparse coding, reduces the storage space, and reduces the storage requirements for hardware.