Daily inflow forecasting for the Hirakud reservoir on the Mahanadi River basin, India, comparing four neural network architectures on the same gauge and meteorological record.
| Notebook | Architecture |
|---|---|
DNN - (Hirakud Yearly Data), Flood Forecasting.ipynb |
Fully connected network |
1D-CNN - (Hirakud Yearly Data), Flood Forecasting.ipynb |
1D convolutional network |
LSTM - (Hirakud Yearly Data), Flood Forecasting.ipynb |
Long short-term memory |
CNN-LSTM - (Hirakud Yearly Data), Flood Forecasting.ipynb |
Hybrid CNN-LSTM |
Models are evaluated with MAE, Nash-Sutcliffe efficiency (NSE), RSR and volumetric error. The hybrid CNN-LSTM gave the best daily forecast accuracy of the four.
This work was carried out at IIT Kharagpur under Prof. Sudip Misra and formed the basis of:
- Chatterjee, S., Khatun, A., Misra, S., & Ghosh, T. (2022). Daily flood forecasting for a tropical river basin using state-of-the-art deep learning techniques. Proceedings of the American Geophysical Union (AGU) International Conference, Chicago, IL, USA.
- Chatterjee, S., & Khatun, A. (2024). Streamflow forecasting into a large Indian reservoir using standalone and hybrid machine learning models. International Conference on Water Resources, Roorkee Water Conclave, IIT Roorkee & NIH Roorkee, India.
The line of work was later extended into a systematic study of input-variable and time-lag selection, published as:
Khatun, A., Nisha, M.N., Chatterjee, S., & Sridhar, V. (2024). A novel insight on input variable and time lag selection in daily streamflow forecasting using deep learning models. Environmental Modelling & Software, 179, 106126. doi.org/10.1016/j.envsoft.2024.106126
Python, TensorFlow/Keras, scikit-learn, NumPy, Pandas, Matplotlib.
The reservoir inflow and meteorological records are not redistributed here. The notebooks document the modelling approach and can be re-run against equivalent data.