Identifying information sources plays a significant role in network science and engineering. However, existing source identification approaches generally focus on static networks without considering the temporal features of networks. To this end, we comprehensively study the problem of identifying single and multiple information sources in time-varying networks. Specifically, we first represent the time-varying networks by time aggregated graph (TAG), and employ a microcosmic susceptible-infected-recovered (SIR) model to characterize the diffusion dynamics of each node. Second, in the case of single-source, we exploit a TAG-based reverse infection (RI-TAG) algorithm to specify a set of suspect nodes, which not only reduces the scope of seeking the source but also ensures the feasibility of path calculation. Then, a novel computationally efficient algorithm is proposed to estimate the information source and diffusion time simultaneously. Subsequently, in the case of multi-source, we design a multi-source estimation algorithm, which divides the set of infected nodes into various partitions, and then runs our single-source estimation algorithm in each partition. Moreover, we present an effective algorithm to estimate the number of sources. Finally, experimental results on various synthetic and empirical time-varying networks demonstrate the effectiveness of the proposed algorithms.