Abstract A pollution source in groundwater may be active at some location for certain periods. There may be multiple potential sources responsible for observed contamination at observation wells. The contamination witnessed in observation wells at different times establishes breakthrough curves (BTCs). These BTCs are usually employed for source identification. In this work, single and multistage artificial neural network (ANN) is employed to identify the potential pollution sources. Temporally varying potential pollution sources are generated using uniform random numbers. These source fluxes are further applied to the simulation of the pollution concentration at observation wells. BTC at an observation well is characterized by statistical parameters and data mining. Characterized BTCs are inputs and source fluxes are outputs of ANN models. Initial stage ANN models are developed at the specified observation well locations, using multilevel BTC characterization. These initial meta models are utilized for the development of intermediate models. Further, the intermediate models are employed for final stage identification. These multi-stage ANN models are found to perform comparatively better than single stage ANN models.

中文翻译：

---

使用人工神经网络 (ANN) 和多级突破曲线 (BTC) 表征识别污染源

摘要 地下水中的污染源可能在某个地点在一定时期内活跃。在观察井中观察到的污染可能有多个潜在的来源。在不同时间在观察井中看到的污染建立了突破曲线 (BTC)。这些 BTC 通常用于来源识别。在这项工作中，采用单级和多级人工神经网络（ANN）来识别潜在的污染源。使用统一的随机数生成随时间变化的潜在污染源。这些源通量进一步应用于观测井污染浓度的模拟。观测井中的 BTC 以统计参数和数据挖掘为特征。表征的 BTC 是输入，源通量是 ANN 模型的输出。初始阶段 ANN 模型是在指定的观测井位置开发的，使用多级 BTC 表征。这些初始元模型用于开发中间模型。此外，中间模型用于最终阶段识别。发现这些多级 ANN 模型的性能比单级 ANN 模型要好。