This paper proposes a highly accurate data-driven vanadium redox flow battery (VRB) modelling approach for power engineering studies. The proposed approach overcomes the common problem of high model dependency that is encountered by the existing electrochemical principle or equivalent circuit based VRB modelling methods. It directly learns the behavioural relationship between VRB current, flow rate, state-of-charge and voltage through experimentally trained convolutional neural networks (CNN) and thus, avoids the usage of complicated equations for power engineering studies. This contributes to greatly simplify the studies of electrical systems that integrate VRB with improved accuracy. The validity of the proposed approach is verified by experimental results. Noticeably, the performances of both two-dimensional CNN (2D-CNN) and one-dimensional CNN (1D-CNN) on VRB modelling are compared and analysed.