水源水质的快速波动会扰乱水处理厂的日常运行。 水源水的同时出现导致了几起水传播疫情 污染和治疗混乱（吴和维森特，2003年；福克斯和莱特尔，1996年）。 根据流域活动了解水源水波动会增加 WTP操作的鲁棒性。该项目的主要目标是:确定 蒙特利尔水处理厂入口处水源水浊度波动的来源 （WTP）；并且，使用此信息提前24小时使用 人工神经网络（ANN）方法。 该项目的第一步是熟悉兴趣现象， 浊度变化。为此，收集了40个月的每日浊度数据 收集和观察，以描述重大事件并定义任何现有模式。对于 同一时期，收集了43个可能与浊度变化有关的变量的数据 基于文献综述。从这个变量列表中，那些呈现显著季节性变化的变量 变异被保存为潜在的自变量。浑浊的主要原因 通过叠加浊度和潜在指标的图表来确定变化。这 运动还可以观察参数之间的时间间隔。作为对 通过图解法，在指标和浊度之间建立了相关矩阵 不同时间间隔的值。 一旦认为浊度波动的主要原因已经确定，人工 选择神经网络作为建模工具对其进行预测。方法论 采用了在该领域工作的研究团队提出的方法 环境和水资源领域（Baxter等人，2002年； 梅尔和丹迪，2000年）。信息技术 包括六个主要步骤:确定需求；绩效标准的选择； 数据库的开发和组织；神经网络模型的构建； 还有，最后的车型选择。包括7个参考文献、表格、图表。

Rapid fluctuations of source water quality can upset routine water treatment plant operations. Several waterborne outbreaks have been caused by the co-occurrence of source water contamination and treatment upsets (Woo and Vicente, 2003; Fox and Lytle, 1996). Understanding source water fluctuations according to watershed activities increases the robustness of WTP operation. The main objectives of this project were to: identify the origins of source water turbidity fluctuations at the inlet of the Montreal water treatment plant (WTP); and, use this information to forecast turbidity peaks 24 hours in advance using an artificial neural network (ANN) methodology. The first step of this project was to become familiar with the phenomenon of interest, turbidity variations. For this purpose, daily turbidity data for a period of 40 months were collected and observed to characterize the major events and define any existing patterns. For the same period, data were collected for 43 variables possibly related to turbidity variations based on a literature review. From this list of variables, those presenting significant seasonal variation were conserved as potential independent variables. The main causes of turbidity variations were identified by superposing graphs of turbidity and potential indicators. This exercise also allowed observing time lags between the parameters. As a complement to the graphical method, a correlation matrix was produced between the indicators and the turbidity values for different time lags. Once it was felt that the main causes of turbidity fluctuations had been identified, artificial neural networks were selected as a modeling tool to forecast them. The methodology employed was elaborated using the approaches proposed by research teams working in the environmental and water resources fields (Baxter et al, 2002; Maier and Dandy, 2000). It includes six main steps: the identification of the needs; the choice of the performance criteria; the development and organization of the database; the construction of neural network models; and, the final model choice. Includes 7 references, tables, figures.