Long-term Reservoir Inflow Forecasts: Enhanced Water Supply and Inflow Volume Accuracy Using Deep Learning

Machine-learning algorithms have shown promise for streamflow forecasts, reservoir operations, and scheduling, but have exhibited lower accuracy in predicting extended time horizons for reservoir inflows. Newer deep learning algorithms exhibited improved inflow forecasting accuracy, but existing research has been mostly limited to real-time operation and short-term planning. We propose a new multi-step forecasting approach to improve long-term forecasting accuracy for both water supply and inflow volumes. This approach uses historical snow water equivalent (SWE) and reservoir inflow time-series data to train an Encoder-Decoder algorithm to predict the reservoir inflow of future time-steps during the April-July runoff period. The optimal model and hyperparameters are selected through five-fold time-series cross validation for variations between long shortterm memory (LSTM) and convolutional neural network (CNN) Encoder-Decoder algorithms. We evaluated each algorithm using 30 years of reservoir inflow and SWE data at the Upper Stillwater Reservoir located in Utah. The optimal model was an LSTM Encoder-Decoder algorithm with 16 nodes per layer. Using this algorithm, we investigate the trade-off between model complexity and accuracy for long-term water supply, relative to a process-based Ensemble Streamflow Prediction (ESP) model as a baseline and simpler statistical methods traditionally used in forecasting (SARIMA, VAR, TBATS). Long-term water supply forecasts of the optimal deep learning algorithm proved superior to statistical methods and rivaled those of the ESP 50% exceedance probability forecast (i.e. the most probable forecast) evaluated over five consecutive hold-out periods.