1. Application of satellite-based rainfall and medium range meteorological forecast in real-time flood forecasting in the Upper Mahanadi River basin
Trushnamayee Nanda1* Harsh Beria1, Bhabagrahi Sahoo1, Chandranath Chatterjee1
IIT Kharagpur, India
EGU
Abstract
Objectives
Spatial Variation of Rainfall Bias
Performance Evaluation
To evaluate the real-time satellite precipitation product such as Tropical Rainfall Measurement Mission (TRMM) Multi-satellite Precipitation Analysis (TMPA) and numerical weather forecasts such as European Centre for Medium-range Weather Forecast (ECMWF) against gauged-based rainfall of India Meteorological Department (IMD).
To assess the utility of the real-time TRMM (TRMM_RT) and rainfall forecasts of ECMWF.
To develop real-time inflow forecasting system using neural network-based models.
Increasing frequency of hydrologic extremes in a warming climate call for the development of reliable flood forecasting systems. The unavailability of meteorological parameters in real-time, especially in the developing parts of the world, makes it a challenging task to accurately predict flood, even at short lead times. The satellite-based Tropical Rainfall Measuring Mission (TRMM) provides an alternative to the real-time precipitation data scarcity. Moreover, rainfall forecasts by the numerical weather prediction models such as the medium term forecasts issued by the European Center for Medium range Weather Forecasts (ECMWF) are promising for multistep-ahead flow forecasts. We systematically evaluate these rainfall products over a large catchment in Eastern India (Mahanadi River basin). We found spatially coherent trends, with both the real-time TRMM rainfall and ECMWF rainfall forecast products overestimating low rainfall events and underestimating high rainfall events.
However, no significant bias was found for the medium rainfall events. Another key finding was that these rainfall products captured the phase of the storms pretty well, but suffered from consistent under-prediction. The utility of the real-time TRMM and ECMWF forecast products are evaluated by rainfall-runoff modeling using different artificial neural network (ANN)-based models up to 3-days ahead.
Design of neural networks
ECMWF rainfall forecasts
Fig. Performance (i.e., Nash Sutcliffe efficiency, NSE) of lead time inflow forecasting by static and dynamic neural networks.
The peak flow events are better captured by the NARX compared to the ANN model.
The improved performance of the NARX model may be attributed to the dynamic feedback inputs of the network.
Overall performance of NARX (NSE≥80%) is better than the ANN model (NSE≤60%) for inflow forecasting.
The use of rainfall forecasts of ECMWF increases flow forecast efficiency by ~5% over the TRMM_RT only.
Introduction
Five major flood events occurred in the Mahanadi River Delta between impacting the livelihood of more than 20 million people.
Inflow forecasting to a reservoir helps in effective dam operations and downstream flood management
Unavailability of real time rainfall data acts as a major hindrance in accurate flood forecasting.
Existing flood forecasting system is based on gauge to gauge correlation, and has a maximum lead time of up to 24 hours.
Real time satellite precipitation products and numerical weather forecasts may aid in increasing the lead times.
Fig (a). Static Artificial neural network (ANN) (b) Dynamic Non-linear autoregressive network with exogenous inputs (NARX)
Evaluation of TRMM_RT and ECMWF
Discussion
Neural network runs
TRMM-RT, which is available on a real-time basis, does not capture the peak rainfall events. However, it tends to capture the rainfall better than ECMWF forecasts spatially over the Mahanadi River Basin.
TRMM-RT efficiently predicts the inflow to the reservoir with the dynamic neural network model in comparison to the static neural network model.
The rainfall forecasts of ECMWF is efficient in increasing the lead time of inflow forecast up to 3 days using the dynamic neural network model.
The NARX model explicitly accounts for seasonality and persistence in flow time series.
It is expected that with the improved Global Precipitation Measurement Mission (GPM)-based rainfall products, the flow forecast capability of the NARX model would be further enhanced.
Basin characteristics
Data period:
Calibration:
Validation:
Static NN (ANN)
2-days ahead inflow forecast
Precipitation (mm/day)
Dynamic NN (NARX)
ECMWF
Conclusion
3-days ahead inflow forecast
Fig. Mahanadi River Basin.
Empirical models like neural network (NN) tend to be more parsimonious and give reasonable results with coarse resolution satellite forcings.
Dynamic NN-based model suggests auto-bias correction by hidden neurons of NN.
Numerical weather forecasts (ECMWF) can be used to issue flood warnings using data-driven models.
Area: 141,000 sq. km (drains into Bay of Bengal)
Annual Rainfall: 1400 mm (90% during the monsoon)
Climate: Tropical Monsoon
Min/Max Temperature: 42.7⁰C / 8.4⁰C
Dominant Landuse: Agriculture(56%), Forest (26%)
Major Occupation: Agriculture
Precipitation (mm/day)
Fig. Skill corrected probability of detection (SPOD), false alarm ratio (FAR) and equitable threat score (ETS) for TRMM_RT and rainfall forecasts of ECMWF at different precipitation thresholds.
Fig. Reproduction of Observed discharge hydrograph using the Static and Dynamic NN-based models during validation
* Indian Institute of Technology Kharagpur