Abstract:
Monthly streamflow prediction will give us a better idea about for flood warning, hydropower
operation, reservoir operations and environmental protection. The current work look in to
the prediction and evaluation capability of a Deep learning method such as Long Short Term
Memory (LSTM) model, for monthly streamflows of Kidangoor, Pattazhi and Perumannu
stations in Kerala. Prediction accuracy of LSTM method is compared with other Machine
Learning models, i.e. Random Forest (RF) and Support Vector Regression (SVR). Predicted
results of the three stations shows that LSTM model gives better accuracy compared to other
models. For improving prediction accuracy of the models various kernel types are tried and
provide the best results for the stations. The LSTM outperforms the other methods in almost
every stations. It is also found that data preprocessing considerably improves the prediction
accuracy in estimation streamflows. The overall results indicate that the LSTM method
could be successfully used in predicting and estimating monthly streamflow in Kerala. Best
value of R-squared values is shown in Kidangoor data around 0.96 for LSTM. SVR and gives
0.92 and RF gives 0.84 R-squared value. Similarly for every other inputs LSTM provide best
performance. While looking into other stations also LSTM outperforms both SVR and RF
models. It conclude that LSTM is better around these three models
