The invention discloses a photovoltaic interval prediction method and system based on self-coding and an extreme learning machine. A dual-output extreme learning machine model is used for constructinga photovoltaic interval prediction model, dual outputs are the upper limit and the lower limit of a photovoltaic interval respectively, and due to lack of actual values of the photovoltaic interval,a traditional extreme learning machine training algorithm fails accordingly. According to the method, a heuristic algorithm is adopted to optimize a training extreme learning machine, a correspondinginitialization algorithm is provided for improving the calculation result and efficiency of the heuristic algorithm, and the method comprises: carrying out prediction interval initialization based onlinear regression interval estimation; carrying out inputting weight matrix initialization of an extreme learning machine based on self-coding; and carrying out optimization training of a particle swarm-based double-output extreme learning machine. The photovoltaic prediction interval obtained through the model replaces a traditional photovoltaic point prediction value, and more sufficient information can be provided for day-ahead robust scheduling and intra-day economic scheduling of a power system.