Smart grid's objective is to enable electricity and information to flow two-way while providing effective, robust, computerized, and decentralized energy delivery. This necessitates the use of state estimation-based techniques and real-time analysis to ensure that effective controls are deployed properly. However, the reliance on communication technologies makes such systems susceptible to sophisticated data integrity attacks imposing serious threats to the overall reliability of smart grid. To detect such attacks, advanced and efficient anomaly detection solutions are needed. In this paper, a two-stage deep learning-based framework is carefully designed by embedding power system's characteristics enabling precise attack detection and localization. First, we encode temporal correlations of the multivariate power system time-series measurements as 2D images using image-based representation approaches such as Gramian Angular Field (GAF) and Recurrence Plot (RP) to obtain the latent data characteristics. These images are then utilized to build a highly reliable and resilient deep Convolutional Neural Network (CNN)-based multi-label classifier capable of learning both low and high level characteristics in the images to detect and discover the exact attack locations without leveraging any prior statistical assumptions. The proposed method is evaluated on the IEEE 57-bus system using real-world load data. Also, a comparative study is carried out. Numerical results indicate that the proposed multi-class cyber-intrusion detection framework outperforms the current conventional and deep learning-based attack detection methods.