A great number of renewable energy sources and nonlinear loads are integrated into the distribution systems through power electronic converters, which leads to serious harmonic distortions. Correctly determining the location of harmonic sources is the primary task to manage the harmonic pollution users and develop a strategy for harmonic suppression on system level. Existing localization methods leverage the time-domain waveforms of voltage and currents across each feeder. Meanwhile, an accurate knowledge of network topology and line impedances are also required. However, these are difficult to meet in engineering practice. Only the RMS values of harmonics are uploaded from the power quality sensors to distribution management system. Besides, a small change of network topology would produce a totally wrong location result. In this letter, a harmonic source localization method based on Dynamic Graph Convolutional Recurrent Network (DGCRN) is proposed. Only the RMS values of harmonics and pre-defined topology of distribution systems are required to capture the spatio-temporal correlations between harmonic sources and harmonic distortion nodes. The effectiveness of the proposed method is tested in the IEEE14 distribution network. By using the same test set, the accuracy of proposed method is improved by 8.42% over the spatio-temporal graph convolutional network (STGCN).