Identification of Unknown Groundwater Pollution Source Using Linked ANN-Optimization Model

Identification of groundwater pollution sources is a major step in groundwater pollution remediation, particularly for assigning fractions of the remediation cost to different polluters. Identification of unknown groundwater pollution source is an inverse problem which is generally ill-posed. A pollution source is said to be identified completely when its source characteristics (location, strength and release period) are known. In practice, the lag time between the first reading at observation well and the time at which the source becomes active is not known. For such cases, pollution source identification problem becomes more difficult. We propose a linked ANN-Optimization model for complete identification of unknown groundwater pollution sources. Spatial and temporal data of observed and simulated concentration is used to formulate the objective function. An optimization model is then used to minimize the objective function. We define the lag time as the time from the start of the pollutant release to the peak of the breakthrough curve observed at a monitoring well. Lag time for a particular monitoring well is then dependent only on the source location and release period. An ANN model is trained for different source locations and release periods as input data to determine the lag time for breakthrough curve. In the proposed model, the ANN model is linked externally with the optimization model to identify the pollution sources. The main advantage of the proposed model is to identify the unique solution for pollution sources when lag time is not known. The performance of the model is evaluated for a one dimensional case with error-free and erroneous data. The measurement errors incorporated in the data vary from 0% to 10% of the analytically computed values. The results indicate that the proposed linked ANN-Optimization model is able to predict the source parameters quite well for the error-free data. For the observations subjected to random measurement errors, the model was run multiple times to obtain a set of predicted parameters. It was seen that the mean values predicted by the model matched well with the exact values, with an error of about 2% in location, 10% in source strength, and 4% in the release period. The standard deviation of the predicted values shows an increasing trend with increase in the level of measurement errors. The model appears to be robust and may be used to solve the inverse pollution source identification problem efficiently.