Reinforcing a Monitoring System of a Regulated River Using LSTM

Accurate methods predicting flow in a river require high performance computing systems providing only short-term early warnings. To address these limitations, we propose an optimized Long Short Term Memory layer combined with a Fully connected layer within a six-layer architecture to estimate flows based on readings of three neighboring river gage heights. This study assessed the number of cells in the LSTM and the number of neurons in the Fully connected layer as well as the probability in the Dropout layer to avoid overfitting. Identification of the best settings by means of the RMSE evaluated six different input datasets to estimate flow. Our study proved that relying on the river gage height from neighboring sensors is possible to predict flow which is crucial when a sensor may fail and a reliable neural network at low computational cost can handle that.