There are some systematic biases in prediction of solar radiation based on numerical models, the AVT method has been verified to reduce the prediction bias effectively. The AVT correction method was used to modify the forecasted solar radiation in two photovoltaic power plants in Hexi area of Gansu Province. The results state: (1) Before correction the forecast bias presented obviously diurnal variation with increase first and then decrease, while it wasn’t obvious after correction. There was a strong linear relationship between forecast deviation of solar radiation and observed values before correction, while it got weak after correction (the correlation coefficient and good fitting degree decreased). (2) There was obviously annual change characteristic of solar radiation with the highest forecast bias in spring, followed by summer, autumn and winter. After correction, its forecast bias was reduced in four seasons, especially in spring and summer.