Solar radiation is an important factor in forecasting outputs of photovoltaic power systems. A method for solar radiation prediction is proposed based on wavelet transform. The data, including solar radiation and potential input variables, are decomposed into several time-frequency areas via wavelet transform. Given that some variables are relevant in some areas of solar radiation but irrelevant in other areas, we extract input variables from candidates separately in each time-frequency area. We perform principal component analysis (PCA) to extract principal components from input variables and we find their relations to the solar radiation are linear by visualization. Thus, linear regression equation is built to each area respectively, and the final result is the summation of predicted solar radiation from each time-frequency area. A comparison with the method using the same input variables in all time-frequency areas is presented and the results of our proposed method are more accurate.