Hydraulic accidents or abnormal situations, also known as leakages, cause not only water losses, but also service interruptions and other negative effects. In order to facilitate the rapid response of water utilities and reduce water losses caused by undiscovered leakages, a timely detection and isolation method is required. To solve the battle problem in L-town, this study investigates the potential of the combination use of data-driven method and hydraulic modelling and proposes a multistage approach, which comprises three stages: estimation, identification and localization. Firstly, empirical mode decomposition and vector auto regressive are performed to identify the trend in flow time series and the spatial correlation of pressure values at different sensors. The two data-driven estimations of flow and pressure are used to extract leak-induced values from the monitoring data. Secondly, first difference and weekly difference are calculated to reduce the nonstationary of residual series. The duration and size of leakages are identified by analysing residuals though three-sigma SPC method. Meanwhile, the window-size slope variation method is proposed to detect the start time and leak size of incipient leakages. Thirdly, emitters are used to represent and simulate pipe leakages in EPANET2. Localization algorithm, which uses the idea of control variates to find two weeks when the leakage had occurred and not occurred, is employed to represent the impact of the leakage on the pressure of the network by calculating the difference between the pressure data of two weeks. By identifying the gaps between the differences of the observed values and of the simulated values, areas with a high probability of leakage occurring are found. The proposed method is applied to L-town network. Results show that 16 leakages are detected and localized in 2019, which demonstrates the effectiveness of the solution for leak detection and localization in water distribution systems.