Short-Term Reservoir Inflow Forecasting Using Long Short-Term Memory with Hydrometeorological Inputs and Wavelet Denoising

Accurate daily reservoir inflow forecasting is essential for hydropower reservoir operation and water resources management, yet inflow dynamics are highly nonlinear and nonstationary. This study compares three long short-term memory (LSTM) based scenarios for one-step-ahead daily reservoir inflow forecasting at the Saguling Hydropower Plant Reservoir: a univariate LSTM using inflow only, a multivariate LSTM using inflow and hydrometeorological variables, and a wavelet-denoised LSTM that applies wavelet denoising to each input window before prediction. Inflow and hydrometeorological data from 2014 to 2023 were divided chronologically into training (2014-2020), validation (2021), and test (2022-2023) sets. The results show that adding hydrometeorological variables improves forecasting accuracy over the univariate LSTM baseline, while wavelet-based preprocessing provides a further gain in accuracy and stability. The wavelet-denoised LSTM achieved the best test performance, with an RMSE of 29.237 m3/s, an MAE of 18.548 m3/s, and an R squared of 0.813. These findings indicate that both input selection and preprocessing strategy materially affect short-term reservoir inflow forecasting and can support more reliable hydropower reservoir operation.