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 DL algorithms have improved the accuracy of inflow forecasts, yet most studies emphasize short‑term applications or real‑time operations. This study proposes a novel multi‑step forecasting framework to enhance long‑term predictions of reservoir inflow and water supply. Using snow‑water equivalent (SWE) and historical inflow data, we trained a DL model built on a convolutional neural network (CNN)–long short‑term memory (LSTM) encoder–decoder architecture to forecast inflows during the critical March–August runoff period. Model architecture and hyper‑parameters were tuned via multi‑fold cross‑validation of the time series, examining various CNN‑ and LSTM‑based encoder–decoder adaptations. The methodology was applied to 40 years of SWE and inflow data from Jordanelle Reservoir, Utah. The optimal configuration—an LSTM encoder–decoder with 16 nodes per layer—achieved substantial improvements in long‑term forecast accuracy. We also assessed the trade‑off between model complexity and performance by benchmarking against a process‑driven ensemble streamflow prediction (ESP) model and classical statistical methods (SARIMA, VAR, TBATS). The DL approach outperformed the statistical models for long‑term water‑supply forecasts and achieved accuracy comparable to the ESP model’s 50 % exceedance‑probability forecast. Overall, these results highlight the promise of advanced DL methods for enhancing long‑term hydrological forecasting and water‑resource management.