Authors: Md Abdullah Al Mehedi, Achira Amur, Jessica Metcalf, Matthew McGauley, Virginia Smith and Bridget Wadzuk – Villanova University
Title: Predicting the Performance of Green Stormwater Infrastructure Using Multivariate Physics-Informed Long Short-Term Memory (LSTM) Neural Network
Abstract: Green Stormwater Infrastructure (GSI) is an increasingly prevalent approach for managing stormwater runoff, utilizing natural and engineered systems, such as green roofs and rain gardens. One of the key challenges in implementing GSI is the ability to accurately predict performance in processing stormwater runoff. Traditional modeling methods often rely on simplified assumptions and limited data, restricting the accuracy of predictions. In this research, we apply a physics-informed LSTM (Long Short-Term Memory) neural network model to predict the performance of GSI systems. The model integrates physical principles, such as water balance and infiltration, into the LSTM architecture to improve the accuracy of predictions. We utilized a five-year dataset of precipitation, air temperature air moisture, and soil moisture with a five-minute temporal resolution to train the model. The proposed physics-informed LSTM model offers a promising approach for predicting the performance of GSI compared to traditional models. The incorporation of physics-based knowledge in the LSTM model improves the accuracy and reliability of the predictions. The results of this study have wide transferability to the design and implementation of GSI systems in urban areas, to optimize their mitigative impact on urban stormwater. The application of these models can be extended to other types of GSI, such as bioretention systems, permeable pavements, and green roofs, to provide a comprehensive understanding of their performance and effectiveness in managing stormwater. Further, these findings open the door to explore the potential of physics-informed machine learning models in predicting the performance of GSI under a range of different environmental conditions.