Deciphering the black box of deep learning for multi-purpose dam operation modeling via explainable scenarios

Operational rules of a multi-purpose dam are hidden due to missing of the records for decision-making processes. This study aims to assess the explainability of a deep learning model for the multi-purpose dam operation of Seomjin River, Juam, and Juam Control dams in the Seomjin River basin, South Korea. In this study, the Gated Recurrent Unit (GRU) algorithm is employed to predict the hourly water level of the dam reservoirs over 2002-2021. First, the GRU models are trained and validated using the local dam input (precipitation, inflow, and outflow) and output (water level) data to examine similarity/singularity in the operational patterns of these three dams. The hyper-parameters are optimized by the Bayesian algorithm. Secondly, the sensitivity test of the trained GRU model to altered input data (-40%,-20%, +20%, and +40%) is conducted to understand how the GRU models facilitate the input data to simulate the target output data (herein, hourly water level), which is known as explainability scenarios. Results show that the trained GRU models predict the hourly water level well across the three dams (above 0.9 of the Kling-Gupta Efficiency). Results from the explainability scenarios show a linear response to the altered inflow rates, but no response to altered precipitation. Furthermore, the GRU models show a site-specific response to altered outflow rates, depending on whether the observed outflow rate-water level relationship is linear or not. This study hints how to decipher the black box of deep learning in multi-purpose dam operation modeling via explainable scenarios.