Prediction of Streamflow using Deep learning and Machine Learning models

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