Discharge Forecasting in Coastal Rivers Using CNN-LSTM Hybrid Approach

Discharge forecasting in tidal rivers and estuaries represents a challenging task because of the nonlinear interaction between sea dynamics and river flow. Areas of tidally influenced rivers are highly affected by marine factors, more precisely, tides, currents, and waves, as well as riverine factors, mainly river discharge. This study focuses on the lower reach of the Neretva River, located in the Republic of Croatia, up to 25 km upstream of the river mouth. The selected area is specific because the saltwater can penetrate 22 km upstream to the town of Metković. A hydrodynamic model HEC-RAS 6.2, was used to simulate non-stationary water levels and discharges at four stations; however, only a single discharge location was considered for the model construction. A sliding window approach used the previous 24-hour water level and discharge observations as input to estimate the discharge with the selected forecasting horizon up to 12 hours. A hybrid machine learning approach that combines a Convolutional Neural Network (CNN) and a Long Short-Term Memory (LSTM) model was applied. The first scenario included only time series data, the second only spectrograms, and the third combined time series and spectrograms as model inputs. Three metrics were selected for model performance comparison: Nash-Sutcliffe Efficiency (NSE), mean absolute percentage error (MAPE), and root mean squared error (RMSE). The conducted study showed that a spectrogram and the combination of a spectrogram with a time series provided better discharge predictions than using only time series data.