Recurrent Neural Network for forecasting water quality parameters: Sensitivity Analysis

Abstract. Recurrent neural network is important for forecasting of dynamic time series. One of the challenging research problems is to extend the black-box modeling into white-box modeling in order to gain insights into the physical processes. The main objective of this study was to perform sensitivity analysis of recurrent neural network in order to identify parameters that are important for predicting water quality constituents.

The study site was windrow compost pad located at UGA Bioconversion center, Athens, GA, USA. Runoff from windrow composting pad was collected in order to prevent the discharge of organic pollutants. We used nine-years time series data of precipitation, temperature, pond volume, material volume on the pad, total suspended solids (TSS), biological oxygen demand (BOD) and Nitrate (NO₃) concentration levels of stored runoff in the collection pond.

Previously, we applied recurrent neural network for predicting water quality constituents and performed auto-correlation and cross-correlation analysis. We used first eight years of data for building the model (from January 2001 to December 2008) and last year of data for testing the model (from January 2009 to December 2009). In this study, sensitivity analysis of recurrent neural network allowed better understanding of dynamics of the complex system by ranking the input parameters and quantifying their influence on water quality prediction.