Multi-step Weekly Average Forecasting of Reservoir Storage Volume Using
Deep Learning
Abstract
Machine-learning algorithms have shown promise for streamflow forecasts,
reservoir operations, and scheduling, but have exhibited lower accuracy
in predicting extended time horizons of peak storage volume (PSV). Deep
learning algorithms exhibited improved inflow forecasting accuracy, but
existing research has been mostly limited to real-time operation and
short-term planning. We evaluate a new approach based on a hybrid
ResCNN-LSTM Encoder-Decoder algorithm, enabling long-term multi-step
reservoir forecasts. The proposed approach provides a three-month,
weekly averaged prediction of reservoir storage volume (RSV) during the
runoff season based on historical snow water equivalent (SWE). The
optimal architecture and hyper-parameters for the model are configured
through five-fold cross validation resulting in a twelve-layered
residual convolutional neural network (ResCNN) as the encoder and a
four-layered long short-term memory (LSTM) neural network as the
decoder. We evaluate the algorithm using 30 years of RSV and SWE data at
the Upper Stillwater Reservoir located in Utah. The most accurate
long-term predictions occurred during periods of large runoff (in excess
of 28,000 ac-ft). The periods where the model performed the worst were
during small runoff and late-season SWE accumulation. We find that the
ResCNN-LSTM consistently outperforms three widely used statistical
models, with an average PSV absolute percent error of 2.66% for the
proposed algorithm compared to SARIMA (14.22%), TBATS (13.82%), and
VAR (18.14%).