Water distribution network (WDN) is one of the most essential infrastructures all over the world and ensuring water quality is always a top priority. To this end, water quality sensors are often deployed at multiple points of WDNs for real-time contamination detection and fast contamination source identification (CSI). Specifically, CSI aims to identify the location of the contamination source, together with some other variables such as the starting time and the duration. Such information is important in making an efficient plan to mitigate the contamination event. In the literature, simulation-optimisation methods, which combine simulation tools with evolutionary algorithms (EAs), show great potential in solving CSI problems. However, the application of EAs for CSI is still facing big challenges due to their high computational cost. In this paper, we propose DLGEA, a deep learning guided evolutionary algorithm to improve the efficiency by optimising the search space of EAs. Firstly, based on a large number of simulated contamination events, DLGEA trains a deep neural network (DNN) model to capture the relationship between the time series of sensor data and the contamination source nodes. Secondly, given a contamination event, DLGEA guides the initialisation and optimise the search space of EAs based on the top K contamination nodes predicated by the DNN model. Empirically, based on two benchmark WDNs, we show that DLGEA outperforms the CSI method purely based on EAs in terms of both the average topological distance and the accumulated errors between the predicted and the real contamination events.

供水网络（WDN）是全世界最重要的基础设施之一，确保水质始终是重中之重。为此，通常在WDN的多个点上部署水质传感器，以进行实时污染检测和快速污染源识别（CSI）。具体而言，CSI旨在确定污染源的位置以及其他一些变量，例如开始时间和持续时间。这样的信息对于制定减轻污染事件的有效计划很重要。在文献中，将仿真工具与进化算法（EA）相结合的仿真优化方法在解决CSI问题方面显示出巨大潜力。但是，由于EA的计算成本高，其在CSI中的应用仍然面临着巨大的挑战。在本文中，我们提出了DLGEA，这是一种深度学习指导的进化算法，可通过优化EA的搜索空间来提高效率。首先，基于大量的模拟污染事件，DLGEA训练了一个深度神经网络（DNN）模型，以捕获传感器数据的时间序列与污染源节点之间的关系。其次，在发生污染事件的情况下，DLGEA会根据顶部情况指导初始化和优化EA的搜索空间DNN模型预测的K个污染节点。基于两个基准WDN的经验表明，在平均拓扑距离和预测污染事件与实际污染事件之间的累积误差方面，DLGEA优于单纯基于EA的CSI方法。