Feeding Static Values to LSTMs for Seq2Seq Learning for Simultaneous Source Identification and Parameter Estimation In Groundwater
Abstract
Long Short-Term Memory (LSTM) networks are extensively used for learning temporally dependent data, for example time series forecasting. For problems involving temporally dependent data at both input and output i.e., sequence-to-sequence (Seq2Seq) learning, encoder-decoder LSTMs are used. In this technique, the time-dependent data from the input are encoded into an encoded state using an encoder LSTM which is deciphered by a decoder LSTM. This type of deep neural network (DNN) architecture is suitable for inverse modelling problems in groundwater such as the identification of contaminant sources where the contaminant release history data from the source locations is mapped to contaminant observation data at the observation wells. However, many times, hydrogeological parameters such as dispersivity, porosity and hydraulic conductivity etc. may be unknown. As these parameters are static, these can not be directly implemented into encoder decoder LSTMs. Therefore, in this study, we propose a novel DNN architecture named Entity Aware Seq2seq learning using LSTMs (EAS-LSTM) which uses entity aware LSTM (EA-LSTM) as encoder and LSTM as a decoder. The proposed DNN can incorporate both static and dynamic data as input and provide dynamic data as output. It is demonstrated to be used as a surrogate model in conjunction with a stochastic population-based optimizer named Multiverse Optimization (MVO). The performance of the EAS-LSTM-MVO model is compared with other surrogate simulation optimization (SSO) models involving Kriging and Support Vector Regression (SVR) in a case study. It is observed that the proposed SSO model performs better compared to other SSO models due to its higher accuracy in reproducing breakthrough curves at the observation wells for given transport parameters and contaminant release history at source locations. The EAS-LSTM model may also be used in other hydrological applications which involve static and dynamic data as input.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.H33B..05A