The combination of the Inertial Measurement Unit (IMU) and Global Navigation Satellite System (GNSS) is widely used in the localization of autonomous cars. However, GNSS is highly dependent on external conditions and has a low sampling rate. In order to make the localization of autonomous cars more reliable in various external conditions, a virtual sensing system using Error-state Kalman Filter (ESKF) and Diagonal Recurrent Neural Network (DRNN) approach is proposed in this paper. In this proposed system, DRNN served as a predictor for the location of an autonomous car. DRNN is applied due to its independency against the external conditions, the ability to learn, and also its faster sampling rate compared to the global navigation system. Implementation and testing of this new approach using the CARLA Simulator show that the proposed system could correct the deviation caused by the absence of absolute position measurement. With further developments and improvements, this alternative sensing method would be able to replace the existing GNSS and unlock the possibility for offline localization.