Application of deep learning algorithms to confluent flow-rate forecast with multivariate decomposed variables

Study region: Song bengue confluent in Cameroon regulates the river flow rate for hydro energy production with input from four upstream reservoirs. Study focus: Deep learning models forecast a day flow rate of the Song bengue confluent. Decomposed time series multivariate variables of flow rate, precipitation, and upstream reservoir inflows, outflows, and precipitation are used. Different windows and horizons for the forecast are analyzed using deep learning models. A comparative study among the models is carried out. Input parameters are decomposed and different partitions are used as scenarios for the best partition. New hydrological insight: A 7-day window and 1-day forecast yield the lowest error. The dense model is the best among the models followed by the Long-short term memory (LSTM) model, and lastly, the one-dimensional convolutional neural network (Conv1D) based on mean absolute error (MAE), mean square error (MSE), root mean squared error (RMSE), and Nash Sutcliff Efficiency (NSE). Using the scenario with all decomposed variables produces the best result with about a 50% difference in error margin. The second-best result is obtained by using only undecomposed data. The remainder component should not be ignored as it contains important hydrological information.