Time series denoising is an inevitable problem in many processes of time series analysis and forecasting. In this paper, a novel denoising method based on empirical mode decomposition (EMD) and dictionary learning is proposed. With the method of EMD, the original signals are adaptively decomposed into a finite number of intrinsic mode functions (IMF). Then we apply the wavelet thresholding method to distinguish the noise components and effective components of the original signals for subsequent dictionary learning. Next, the K-SVD dictionary training algorithm is used to learn inner structures and properties of effective compositions and noise respectively. Finally, the noise of time series can be removed using this combined method of EMD and dictionary learning. To demonstrate the effect of combination strategy, we compare the performance of our proposed method to two state-of-the-art algorithms, i.e., wavelet thresholding algorithm (WTA) and EMD thresholding algorithm (EMD-TA) with two commonly used evaluating metrics: signal-to-noise ratio (SNR) and correlation coefficient (CC). The results show that the proposed denoising method provides a competitive performance to the traditional denoising techniques.