Entity aware sequence to sequence learning using LSTMs for estimation of groundwater contamination release history and transport parameters

Abstract

Groundwater contaminant sources identification and parameter estimation using the simulation–optimization (S/O) approach require numerous runs of the computationally expensive simulation model through the optimization algorithm. The computational cost can be effectively reduced by using a surrogate model which can accurately approximate the simulation model. With the advent of deep learning, Long Short-Term Memory (LSTM) networks, which are suitable to learn sequential data, are being increasingly applied to regression problems involving time dependencies. However, for the simultaneous contaminant source identification and parameter estimation problem, the surrogate model requires to establish a relationship between release histories at the source locations to concentration measurements at observation points subject to given values of aquifer parameters. In this study, a novel deep neural network framework Entity Aware Sequence to sequence learning using Long Short-Term Memory (EAS-LSTM) is proposed as a surrogate model which takes both sequential i.e., release histories at the source locations at different stress periods and static variables i.e., transport parameters as inputs to predict breakthrough curves (BTCs) at observation points. The proposed surrogate model is applied to a heterogeneous field scale aquifer with 4 zones. The Mean Squared Error (MSE) using EAS-LSTM is 0.0033 ppm2 which is significantly better than that of Kriging and Support Vector Regression (SVR) which are 0.048 ppm2 and 1.302 ppm2 respectively. Comparison of EAS-LSTM model performance with Kriging and SVR based models demonstrates its higher accuracy. Further, the optimization algorithms for inverse modelling, the performances of Multiverse Optimizer (MVO), Grey wolf optimization (GWO) and Particle swarm optimization (PSO) are investigated. It is observed that the combination of EAS-LSTM and MVO provides better results in comparison to other surrogate simulation optimization (SSO) models.