Partial discharge (PD) detection and location based on ultra-high frequency (UHF) sensor array have been employed for the power equipment in the whole substation for assessing the insulating condition of electrical equipment and determining a precondition for further implementing insulation diagnosis. At present, one of the widely used localisation methods is time difference of arrival (TDOA) based, localisation result of which is extremely sensitive to time delay estimation and usually time-consuming to solve. Motivated by this, a data-driven method using the deep neural network (DNN) is proposed in this study to significantly speed up the solving process of non-linear TDOA equations and simultaneously guarantee the accuracy of results. It works with sequences of time delay measured from the UHF sensor array as the input of the network and with the corresponding coordinates of PD source as output to train the network. Simulation results demonstrate that the proposed method shows relatively higher accuracy and efficiency in PD location. In addition, many factors such as array shape, error type added to time delay, and detailed structure and parameters of network are taken into consideration for error analysis, laying foundation to more reliable localisation of PD.