Machine learning approaches have become very popular in many fields in this big data age. This paper considers the problem of forecasting realized volatility with machine learning using high-frequency data. Instead of treating the realized volatility as a univariate time series studied by many existing works in literature, we employ panel data analysis to improve forecasting accuracy in the short term. We use six effective machine-learning methods for the realized volatility panel data. We compare our results with the traditional linear-type models under the same panel data framework and with the single time series forecasting via the same machine learning methods. The results show that the panel-data-based machine learning method (PDML) outperforms the other methods.