For the problem of low accuracy of photovoltaic power low prediction, this paper puts forward the algorithm of forecasting active power of wavelet neural network based on Markov modification. This paper divides the historical data into sunny, cloudy weather types, trains them respectively, and establishes network model, and then according to the prediction of daily weather forecast, uses different model prediction algorithms. Large-scale grid-connected photovoltaic power will cause great disturbance on the balance of power system, so the accurate prediction of dispatching department by photovoltaic power and advanced control on the automatic generation control can effectively reduce the frequency fluctuation and reduce the regulating pressure of automatic generation control unit. Practical examples are presented to validate the algorithm and frequency recovery characteristics under active disturbance are simulated, which show that ahead control of automatic generation control based on photovoltaic power generation prediction is effective.