Search

Accurate and reliable dam inflow prediction models are essential for effective reservoir operation and management. This study presents a data-driven model that couples a long short-term memory (LSTM) network with robust input predictor selection, input reconstruction by wavelet transformation, and efficient hyper-parameter optimization by K-fold cross-validation and the random search. First, a robust analysis using a "correlation threshold" for partial autocorrelation and cross-correlation functions is proposed, and only variables greater than this threshold are selected as input predictors and their time lags. This analysis indicates that a model trained on a threshold of 0.4 returns the highest Nash–Sutcliffe efficiency value; as a result, six principal inputs are selected. Second, using additional subseries reconstructed by the wavelet transform improves predictability, particularly for flow peak. The peak error values of LSTM with the transform are approximately one-half to one-quarter the size of those without the transform. Third, for a K of 5 as determined by the Silhouette coefficients and the distortion score, the wavelet-transformed LSTMs require a larger number of hidden units, epochs, dropout, and batch size. This complex configuration is needed because the amount of inputs used by these LSTMs is five times greater than that of other models. Last, an evaluation of accuracy performance reveals that the model proposed in this study, called SWLSTM, provides superior predictions of the daily inflow of the Hwacheon dam in South Korea compared with three other LSTM models by 84%, 78%, and 65%. These results strengthen the potential of data-driven models for efficient and effective reservoir inflow predictions, and should help policy-makers and operators better manage their reservoir operations. [ABSTRACT FROM AUTHOR]