The electric power grid uses a set of measuring and switching devices for its operations and control. The data retrieved from the measuring instruments is assumed to be noisy, therefore a state estimator is used to estimate the correct values of state variables on which the system can take control actions. The modern electric power grid is dependent on communication networks for transferring these measurements, which are susceptible to intrusions from hackers. False data injection attacks (FDIA) are one of the most common attack strategies where an intruder tries to trick the underlying control system of the grid to cause disruptions without getting detected by native anomaly detection measures inbuilt in the state estimator. The native anomaly detection mechanism relies on threshold and residual based measure to flag a set of measurements as anomaly. Therefore, if the attack is devised in such a way that the intrusion can be performed without significantly affecting the residual error of state estimation it can go undetected. We propose a data augmented deep learning based solution to detect such attacks in real time. We propose methods of generating realistic random and targeted attack simulations on standard IEEE architectures and methods of detecting them using deep learning models. We propose recurrent neural network (RNN) based architectures to detect and locate FDIAs and devices compromised in real-time. For detection we propose a supervised and an unsupervised method. Similarly, for location we propose a method to find exact devices compromised which is less practical and then move on to a more feasible and practical solution in supervised and unsupervised conditions. Being an intrusion detection system it is critical to detect all attacks which means false negatives should be penalized heavily, whereas false positives can be accommodated. Therefore, we use recall as our primary performance metric and precision recall curve to find an optimal threshold of probability score. In addition, we demonstrate how our approach is better than a residual error and other previous detection models. We also compare the performance of our models with increasing number of devices being compromised. ii