Stream flow prediction using TIGGE ensemble precipitation forecast data for Sabarmati river basin Open Access

TIGGE

THORPEX International Grand Global Ensemble

NWP

Numerical Weather Prediction

ECMWF

European Centre for Medium-Range Weather Forecasts

NCEP

National Centers for Environmental Prediction

UKMO

United Kingdom Meteorological Office

HEC-HMS

Hydrologic Engineering Center – Hydrologic Modeling System

ANN

Artificial Neural Network

EPS

Ensemble Prediction Systems

CWC

Central Water Commission

QPF

Quantitative Precipitation Forecast

DEM

Digital Elevation Model

LULC

Land Use Land Cover

HSG

Hydrological Soil Group

SCS-CN

Soil Conservation Service Curve Number

BMA

Bayesian Model Averaging

EMOS

Ensemble Model Output Statistics

PDF

Probability Density Function

NGR

Nonhomogeneous Gaussian Regression

cNLR

Censored-Non-homogeneous Logistic Regression

There have been a number of natural disasters that have taken place without warning throughout the last several decades. Many people agree that floods are unpredictable and have the ability to ruin both the lives of humans and animals, and property. It is very difficult to predict when and how often severe hydrometeorological events may cause flooding. Early warnings and flood measures to reduce human loss and property damage are based on accurate estimates of runoff volume and flood peaks. For sustainable water resource planning and management, the proper assessment of surface runoff in catchment areas requires an extensive understanding of the flood affecting factors. Understanding the hydrologic response of a region by simulating the rainfall and runoff is a well established method. The most typical use of rainfall–runoff models is to predict flood occurrences, measure water levels for different water conditions and forecast floods (Ahrens & Jaun 2007). Land use, slope, vegetation, and storm attributes such as length, volume and intensity of rainfall have a role in determining the quantity of surface runoff . Asia is the most flood-prone continent, accounting for over half of all flood-related deaths in the twentieth century (Bürger et al. 2009; Hapuarachchi et al. 2022). When it comes to dealing with an emergency caused by significant rains, decision-makers rely heavily on predictions. Precipitation is the most critical piece of information needed to predict floods. Flood forecasting is heavily influenced by two key factors: accuracy and lead time (Jha et al. 2018; Kumari et al. 2019). Forecasting floods using numerical weather prediction (NWP) precipitation products is one of the most effective techniques to increase the lead time of the forecast. The NWP enables the collection of valuable flood data and the dissemination of flood warnings ahead of time (Roulin & Vannitsem 2005). To get around the deterministic forecast, meteorologists and hydrologists are turning their attention to the inherent uncertainty in their respective systems. An ensemble forecast, rather than the usual single deterministic prediction, is increasingly taking its place (Roulin & Vannitsem 2005; Nair & Indu 2017). This is a shift from a deterministic forecast to an ensemble forecast. By taking into account the faulty boundary conditions and assimilation of data in the ensemble prediction, a deterministic forecast of atmospheric variables may be transformed into a fully probabilistic one. Precipitation and runoff forecasting have never been done this way before (Komma et al. 2007). An interactive grand global ensemble system was developed to account for the uncertainty of projections from several global models. As a part of a worldwide research effort, The International Grand Global Ensemble (TIGGE) collects and analyses prediction data from all of the world's main forecasting centers (Thielen et al. 2009). Multiple sources of uncertainty are combined into a probability distribution via TIGGE ensemble prediction.

When the ECMWF and the NCEP first developed ensemble predictions in 1992, they were frequently utilized in inflow forecasts, since they were based on a large number of different models. Flood forecasting, in particular, may benefit from the use of meteorological predictions. The ensemble predictions, which refer to numerous distinct forecasts created using various physical parameterizations or varied beginning circumstances, have made major progress in improving meteorological forecasts (Swinbank et al. 2016). (Saedi et al. 2020). Another use for the TIGGE is to create an early warning system based on operational medium-range ensemble predictions from the different numerical weather centers: the ECMWF, UKMO and NCEP. Each model's climatological probabilistic density function is used to determine the predicted likelihood of occurrence of severe meteorological events. Severe disasters like the 2010 Russian heatwave, the 2010 Pakistan floods, and 2012's Hurricane Sandy were all correctly predicted using this tool (Matsueda & Nakazawa 2015). According to Yang & Yang (2014), data-driven models, such as ANN and autoregressive models, are also employed for rainfall–runoff assessments, as well as predictions. A study by Cheng et al. (2005) found that the ANN model was capable of making accurate predictions about the long-term discharge of a reservoir. The quality and amount of the data, on the other hand, are critical to the success of these models (Wu et al. 2009). If data-driven models are utilized that do not have sufficient data, the prediction uncertainty will be greater. Data-driven models are not included in this research since it is difficult to get high-quality and large amounts of data. When it comes to predicting reservoir inflows, combining NWP with a hydrological model at the catchment size is a simple technique. However, a hydrometeorological prediction system comes with a lot of unknowns, such as boundary and beginning conditions and hydrological model parameters (Liu & Gupta 2007; Hostache et al. 2011; Zappa et al. 2011). The largest source of uncertainty in hydrometeorological predictions is rainfall projections (Rossa et al. 2011; Zappa et al. 2011). The NWP model generates varied rainfall estimates at the same place and time because of insufficient data, approximate forecast models owing to inevitable simplifications, random errors from early atmospheric conditions perturbing, and model parameterizations (Palmer 2001). Rather from being unexpected or devoid of information, differing rainfall estimates, according to Wilks (2006), are not perfectly predictable.

The Ensemble Prediction Systems have evolved over the previous decade and are now used to simulate the impact of observation uncertainty, poor boundary conditions, data assimilation, and other factors on weather forecasts. To anticipate the probability density function beyond the linear error increase in meteorological prediction, an EPS may be considered a system of finite deterministic integrations and regarded as the sole practical technique (Buizza 2008). EPS from each weather station can account for some of the NWP uncertainties that arise from beginning circumstances and stochastic physical processes (Roulin 2007). Grand ensemble (GE) or mixed multi-ensemble (ME) of EPSs from multiple meteorological centers may handle other issues in numerical implementation and data assimilation (Goswami et al. 2007). The ensemble predictions probabilistic character is better preserved when each model participating in the EPS at various weather centers is merged (He et al. 2009, 2010; Bao et al. 2011). As some of the uncertainties may be defined, ensemble weather prediction products can be utilized for hydrometeorological, hydrological (and geological disaster-related) weather forecast, and early flood warning (Bao & Zhao 2012).

When used to estimate streamflow, NWP models QPF are an important input to hydrological models (Coulibaly 2003; Cuo et al. 2011; Liu & Coulibaly 2011; Ahmed et al. 2014). Uncertainty stems from inaccuracies in the NWP model's beginning circumstances and in the atmospheric processes that are approximated, as well as in the NWP model's ability to predict the weather (Palmer et al. 2005). One estimate of streamflow with poor or high-quality precipitation forecasts would have an enormous influence on decision support, such as the management of water infrastructure, sending warnings of upcoming flood or drought, or scheduling reservoir operations. Many people are now interested in probabilistic forecasts that can be used to estimate the likelihood of any future weather event occurring, which will allow water management agencies and emergency services to prepare for the risks associated with low- or high-flow events several days or weeks in advance (Palmer 2002; Thirel et al. 2014; Tao et al. 2015). In order to provide accurate precipitation predictions, the NWP model must be perturbed and physically parameterized, both of which are technically demanding and computationally intensive processes. A post-processing step is required before QPFs (either ensemble or deterministic) provide valid estimates of any data (e.g., streamflow). Several post-processing techniques based on statistical models have been presented in the recent decade. Based on observations and NWP forecasts, a statistical model may be developed by exploiting the relationship between the two, estimating the model parameters using historical data, and reproducing post-processed ensemble forecasts of the future (Velázquez et al. 2010; Jinyin et al. 2016; Jha et al. 2018).

More than 5,500 major dams exist in India. For reservoirs to be effective in reducing flood damage, accurate inflow projections are essential. CWC gives inflow projections for over 150 places in India right now (Jain et al. 2022). An increase in this quantity is urgently required in order to provide projections for all big and minor dams, as well as key cities. It is necessary to employ a better flood forecasting model to give predictions at critical locations and sites, since we have a poorer flood forecasting system in India. According to Sudheer et al. (2019) and other researchers, accurate precipitation and inflow projections are critical to reservoir efficiency. Using ensemble predictions is critical, as Nanditha & Mishra (2021) explained in detail the present state of flood forecasting in India. Floods are becoming increasingly common as the frequency and severity of extreme precipitation events rises throughout most of India and in the world (Jain et al. 2022). The number of cloud-burst incidents is also increasing. Indian scientist lack the resources and expertise to accurately predict these catastrophes.

Using TIGGE data for early flood predictions and computer simulations of rainfall–runoff processes, this research intends to construct an atmospheric–hydrologic flood forecasting model for use in the future. ECMWF, UKMO, NCEP (ensemble) numerical models are used to assess the predictions of the TIGGE database over the Sabarmati basin. Deterministic, dichotomous, and probabilistic evaluation methodologies were used to evaluate the models’ abilities for the time period 2014–2020. The main novelties of the research are to predict stream flow using ensemble precipitation for Sabarmati river basin. This work focuses on statistical analysis of the ECMWF, UKMO and NCEP ensemble precipitation data for Sabarmati basin. The key feature of this research work is to develop the coupled Hydrological and Meteorological model for the accurate flood prediction with statistical post-processing of are ensemble model and its verification for 5-day lead time.

By the middle of June, the southwest monsoon has arrived and will go by the first week of October. The southwest monsoon has a major impact on rainfall. The majority of the basin receives rainfall ranging from 600 to 800 mm. In the basin, a good network of hydrological and meteorological stations has been established. Wireless stations have been set up at different sites to transmit information on rainfall and discharge to a central control center, where it is utilized to make regulatory choices.

According to this research, the datasets that were utilized included observed daily precipitation and ensemble projections of daily rainfall. The Central Water Commission, Gandhinagar and the Gujarat State Water Statistics Center, Gandhinagar provided daily precipitation and inflow data for the last 30 years. As part of this study's ensemble precipitation prediction (Table 1), we used data from the ECMWF, NCEP, and UKMO, under TIGGE. The perturbed predictions are formed by altered beginning circumstances in the CF, which is built by a data-assimilation process (Ye et al. 2014). The forecasting lead period for all of these predictions is 1–15 days. It is possible to obtain these data for free at http://apps.ecmwf.int. In order to establish an inflow prediction, the research uses ensemble precipitation forecasts with a lead period of 1–5 days for the years 2014–2010. After interpolating prediction data to six rain gauges, the Thiessen polygon technique was used to calculate area precipitation.

The Indian Space Research Organization's Cartosat-1 Digital Elevation Model (CartoDEM) is a national DEM (ISRO). It is based on the Cartosat-1 stereo payload, which was launched in May 2005. The DEM used for this research is purchased from NRSC, Hyderabad for the finest resolution of 2.5 m × 2.5 m.

Forecasting scenarios (also known as ensemble members) based on beginning circumstances and physical parameterizations that have been significantly altered in the NWP system are used to create probabilistic predictions. Unfortunately, such ensemble predictions are unable to reflect the whole forecasting uncertainty since it is impossible to precisely and consistently represent all sources of error (Buizza 2018). As a result, ensemble projections tend to be skewed and too optimistic (Wilks 2018). In order to calibrate ensemble predictions, statistical post-processing might be performed. For risk assessment and decision making in business, agriculture, and finance, reliable weather predictions are essential. A good example of this is flood forecasting, which relies on accurate precipitation predictions to determine future streamflow (e.g. Aminyavari & Saghafian 2019). To eliminate systematic mistakes in future forecasts, statistical post-processing looks for structure in previous forecast–observation combinations. For example, Wilks (2018) lists a selection of post-processing ensemble prediction algorithms that have been developed in the recent several years.

Bayesian Model Averaging (BMA) and Ensemble Model Output Statistics (EMOS) are the two basic parametric techniques listed by Raftery et al. (2005) for the class of parametric methods (Gneiting et al. 2005). In the BMA technique, the single ensemble member predictions are weighted averaged to provide a predictive probability density function (PDF). These methods have found several uses, such as in Sloughter et al. (2007) and Schmeits & Kok's (2010) investigations on ensemble precipitation post-processing. The EMOS technique uses a parametric distribution whose parameters are dependent on the ensemble prediction to generate a predictive PDF. The Nonhomogeneous Gaussian Regression (NGR) technique is a popular EMOS model (Gneiting et al. 2005). As opposed to a homogeneous regression model, an inhomogeneous method expresses the predictive distribution's variance as a function of the ensemble's variance. Many studies such as Baran et al. (2013) and Hemri et al. (2014) have used the NGR model, which is based on the assumption of a Gaussian predictive distribution, to make postprocess precipitation projections. It is common to use EMOS with a left-censoring of the prediction distribution at zero for precipitation since it is a non-negative variable. Over the last several years, new methods for post-processing at hitherto undetected locations have emerged. A local model's post-processing parameters are geostatistically interpolated before being used in the interpolation approach. It is via the use of Geostatistical Model Averaging that Kleiber et al. (2011) first introduced geostatistical interpolation as part of a BMA post-processing framework. Precipitation predictions may now be predicted using the same methods used for normally distributed precipitation forecasts by Kleiber et al. (2011) and Pappenberger et al. (2008).

It is first necessary to calibrate and evaluate each model's inflow forecasts using datasets. ECMWF's 1–5 day ensemble predictions from 2007 to 2012 are then included into the operational model to give ensemble inflow forecasts. HEC-HMS models were used to simulate Dharoi reservoir's inflows during calibration and verification periods. The panels indicated that simulated inflows compared to observed inflows. With NSEs of 0.92 and 0.89 for the whole series, the combined rainfall–runoff model performs well in the calibration and verification phases. In this case, we can see that the combined model may provide strong results and be put to good use throughout the prediction period:

The suggested approach's robustness, reliability, and efficacy are evaluated using deterministic and probabilistic verification measures. Accordingly, the produced ensembles are evaluated using deterministic measure and probabilistic measures. Normalized measurements are used to compare data in a straightforward manner. During both calibration and verification, the created ensembles are evaluated to ensure adequate evaluation of the findings (Khajehei & Moradkhani 2017).

Deterministic techniques are used to examine the variance in the mean ensemble precipitation prediction. In addition, a deterministic framework would be useful for studying the link between the observation and the raw prediction.

Deterministic measurements may be skewed by too or underly confident projections, hence probabilistic methods must be used to analyze the resulting forecast. Ensemble forecast reliability may be evaluated using probabilistic methods (DeChant & Moradkhani 2014, 2015). The Continuous Ranked Probability Skill Score (CRPSS) is used to evaluate the forecasting capacity of the created ensembles. Brier Score has been extended to include all conceivable thresholds using this normalized form of Continuous Ranked Probability Score (CRPS) (Hersbach 2000).

CDF (cumulative distribution function), F_t^f(x) is the predicted probability CDF for the t_th forecast scenario, F_t⁰(x) is the observed probability CDF, and T is the number of forecasts. The lower the CRPS number, the better, because CRPS tends to rise when prediction bias increases (whether in a positive or negative way). Mean absolute error is substituted for the CRPS in deterministic forecasts because it is the limiting value of the spread of CRPS when the prediction spread approaches zero. Percentage of daily observations is used to show the relative CRPS. Errors are normalized using relative CRPS, making it possible to compare results across catchments.

An ensemble precipitation's systematic bias may be assessed using the correlation coefficient (CC), RMSE, and relative bias (BIAS). The CC measures the degree of linear correlation between satellite-based precipitation data and gauge measurements.

There are a number of metrics used to evaluate the performance of an individual, such as the probability of detection (POD), false alarm ratio (FAR), miss rate (MISS), and critical success index (CSI). Table 2 shows the most often used contingency table (CSI).

Table 3

Ensemble model verification measures and their formula

Sr. no. .	Verification measures .	Formula .	Perfect/No skill .
1	Root Mean Square Error		0/
2	Relative Root Mean Square Error		0/
3	Probability of Detection (Hit rate)		1/0
4	False Alarm Ratio		0/1
5	Frequency Bias		1/
6	Brier Score		0/1
7	Brier Skill Score		1/ < =0
8	Critical Index Ratio

Sr. no. .	Verification measures .	Formula .	Perfect/No skill .
1	Root Mean Square Error		0/
2	Relative Root Mean Square Error		0/
3	Probability of Detection (Hit rate)		1/0
4	False Alarm Ratio		0/1
5	Frequency Bias		1/
6	Brier Score		0/1
7	Brier Skill Score		1/ < =0
8	Critical Index Ratio

Note: There are N forecast–observation samples, with F denoting the number of forecast–observation pairs, O denoting the observation matching to the prediction, P_F denoting probability of precipitation, and P_O denoting the probability of occurrence of observation. Like and signify the forecast average and observation average, BS_ref is normally the Brier Score of the reference probability prediction, typically the likelihood of event occurrence from climate data. H, M, F and T are generated from contingency table (Table 3).

View Large

Figure 7 shows the various plot for the ensemble data. Scatter plot shown in the figure represent the ensemble and observed data variation, While the spread skill plot shown in the figure shows the absolute error in the raw ensemble data. Rank histogram shows the frequency of the precipitation for the particular data sets.

In the context of discrete event forecasting, the relative operating characteristics (ROC) assess the accuracy of the prediction based on a threshold. The ROC curve assesses the quality of a choice based on a prediction probability for a probability forecast. It shows the trade-off between the POD and the probability of false detection (POFD). Higher POD values and lower POFD values correlate to improved results. The Area under the ROC curve reflects the performance forecasting ability, with more Area Under the Curve (AUC) indicating greater skill.

This is referred to as a Hit if the observed occurrence is accurately predicted (H). A False Alarm (F) occurs when a simulation indicates the occurrence of an event that has not been observed, while Miss (M) marks an event that was observed but not predicted in simulation, and No Alarm implies an event that was neither forecast nor occurred in real time. Figure 13 shows that the models shows the good accuracy for the peak flood time which is 29 July 2015. Models shows 80–90% predicted data using the ensemble prediction technique having a close relationship with the observed datasets.

An atmospheric–hydrologic flood forecast model powered by TIGGE ensemble predictions in upper Sabarmati watershed during 2015 flood event is put up to examine possible advantages of utilizing the TIGGE database in flood forecasting. In order to predict the rainfall–runoff process, the semidistributed HEC-HMS model is used. The findings show that the HEC-HMS model is capable of simulating and predicting floods in the Sabarmati watershed with high accuracy and precision. The TIGGE archive dataset is a potential technique for delivering a pretty accurate warning 5 days in advance with similar discharge projections. Longer forecast lead time may help to increase the predictability, which is good for flood mitigation and preparation.

Multi-model forecasting techniques need to be developed. The usage of many EPS inputs should be done cautiously since each has a separate error structure and cannot be readily mixed. The GE's performance may be improved by assigning various weight coefficients to distinct weather predictions. The semidistributed hydrologic model HEC-HMS in conjunction with the TIGGE ensemble precipitation data is used in this study for flood forecasting. HEC-HMS is driven by TIGGE data from ECMWF, NCEP, and UKMO ensemble prediction products to create flood forecasts for the Sabarmati river watershed. Coupled model simulations reveal that the discharge during the occurrences may be accurately predicted by this semidistributed model. There are some gaps that exist amongst the model predicted results, which makes it difficult to accurately estimate when and where a peak discharge will occur. The amount of precipitation is a crucial factor in flood predictions. Although precipitation directly affects flood prediction precision, input data uncertainty into hydrological models have a significant effect as well. It is possible to improve hydrological forecasts by using ensemble precipitation prediction to generate more quantitative data. The study utilizes the 2015 floods in Sabarmati basin in North Gujarat, India, as an example and evaluates the influence on flood prediction results using the HEC-HMS model forecast and its ensemble forecast results as precipitation in the forecast period. When the precipitation decreases, the future precipitation has less influence on the flood prediction outcome. It is possible to improve the input information for a hydrological forecast by using the precipitation ensemble prediction in conjunction with a definite forecast. To improve forecasting accuracy, it is now able to accurately predict the shape and time of the flood peak. Many hydrologists and meteorologists have highlighted that anticipated precipitation has a significant influence on flood forecasts, and contemporary weather forecasting technology gives solid scientific support for predicting the precipitation in the forecast period.

The authors are thankful to the Civil Engineering Department, Sardar Vallabhbhai National Institute of Technology, Surat for providing an opportunity to do research work. The authors are also thankful to CWC, Gandhinagar and State Water Data center, Gandhinagar for their valuable support in data provision as well as guidance in this project. We would also like to expand our deepest gratitude to all those who have directly and indirectly funneled us in this research work. The authors are thankful to the Civil Engineering Department, Institute of Technology, Nirma University for sponsoring the funded research project with the sum of Rs. 1 lakh for conducting the research.

All relevant data are available from an online repository or repositories. Available from: https://drive.google.com/drive/folders/1k1YATti2aEHPiCKANQFb8fQIHH7YlUuw.

The authors declare that, they have conflict of interest with Nirma University, Institute of Technology, Ahmedabad for providing research funding with the sum of Rs. 1 Lakh to carry out this research work.