機械学習

[II-212] INVESTIGATION OF MISSING RIVER DISCHARGE DATA IMPUTATION METHOD USING LSTM

〇Toshiharu Kojima¹, Wei lisi², Keisuke Ohashi¹ (1.Gifu University, 2.Graduate school of civil engineering)

Keywords:deep learning, RNN, overtraining, stochastic gradient descent, missing data imputation

n this study, Long Short-Term Memory (LSTM) is applied to complement missing values of river discharge and verified. Stochastic gradient descent (SGD) is generally used as the optimization algorithm for deep learning. SGD can obtain good learning results quickly, but the results are slightly different for each learning by its randomized algorithm. The authors performed 1,000 learning and recognized non-ignorable differences in each learning result. The average hydrograph, which obtained by more than 20 learning results, indicates 0.904 in 95 percentile of Nash-Sutcliffe efficiency. The method to reduce the variation of learning results and obtain sufficient accuracy is proposed.