Direction of arrival (DOA) estimation is one of extremely important techniques in array signal processing and thus used in several applications, such as radar systems, source localization, and wireless channel estimation. In this paper, we present a new solution for enhancing the performance of a deep neural network (DNN) specialized in DOA estimation under very noisy environments. After applying principal component analysis (PCA) to the DNN training dataset whose samples were generated at a high signal-to-noise ratio (SNR), we verified that it is possible to strongly reduce the influence from noise in the test data, especially when this was generated at lower SNRs. We also evaluated the effect of 1) different number of antenna elements in the array and 2) different number of reduced dimensions of the training, validation, and test data on the DNN estimation performance. The results presented here are expected to set a precedent in using PCA prior to training DNNs for DOA estimation.