Optimizing dam water level prediction through a one-shot neural architecture search

Abstract

This study investigates the effectiveness of path-based and gradient-based one-shot NAS approaches in optimizing models for up to 14-day ahead water level prediction at the Klang Gates Dam. A super-net which incorporated multiple CNN kernels, fusion structures, and activation functions was developed to support both path-based and gradient-based NAS operations. The results indicated that path-based and gradient-based NAS models outperformed LR and RF in 1-day predictions and achieved comparable performance to CNN-GRU for 7- and 14-day predictions. Notably, these optimized models attained the highest NSE values of 0.9694, 0.8685, and 0.7846 for the 1-, 7-, and 14-day predictions, respectively. These values outperformed the conventional random search models, which attained NSE values of 0.9588, 0.8730, and 0.7813, while requiring lower computational cost. Further analysis revealed that super-net, provided multiple GRU activation functions, enabled the one-shot NAS models to predict the lowest observed water level with greater accuracy.