Improving Reservoir Water Supply and Inflow Volume Predictions with Encoder-Decoder Deep Learning Models

Abstract

Recently, deep learning (DL) models have shown tremendous potential for hydrological prediction, reservoir management, and operational planning. However, their effectiveness in predicting reservoir inflows over extended time horizons remains limited. Recent advancements in deep learning algorithms have enhanced the accuracy of inflow forecasts, but most studies have focused on short-term applications or real-time operations. This study introduces a novel multi-step forecasting framework aimed at improving long-term predictions of reservoir inflow and water supply. Using the snow water equivalent metric and reservoir inflow data, we trained a deep learning model using a Convolution Neural Network (CNN.) - Long Short-Term Memory (LSTM.) encoder- decoder model to forecast inflows to predict future time series of steps during the critical March-August runoff period. The model’s architecture and hyper-parameters were fine-tuned using multi-fold cross-validation of the timeseries, analyzing different adaptations of Encoder-Decoder architectures based on CNNs and LSTMs. The proposed methodology was applied to 40-year time series of SWE and inflow data from the Jor- danelle Reservoir in Utah. The optimal configuration a 16-node-per- layer LSTM Encoder-Decoder model demonstrated significant improvements in long-term forecast accuracy. We further examined the balance between model complexity and performance by benchmarking against a A process-driven Ensemble Streamflow Forecasting model and traditional statistical methods, including SARIMA, VAR, and TBATS. The deep learning approach outperformed statistical models in long-term water supply forecasts and achieved comparable accuracy to the ESP model’s 50% exceedance probability forecast. These findings indicate the potential of enhanced DL methods to improve the long-term hydrological fore- casting and resource management.