Prediction of Flood Using Hybrid ANFIS-FFA Approaches in Barak River Basin

Introduction

Global flood occurrences have time and again raised questions on predicting flood and measures on its control. Precise and well-timed forecast of imminent flood is a very serious issue for diverting and distributing water, preventing drought, human protection, and sustainable ecology. Consistent flood forecast models are very essential in India because most Indian states are affected by flood events, causing major damage to public property and threat to human and animal life. Recently, artificial intelligence methods are more progressively utilized to forecast flood events. Several scholars have efficaciously used ANN to forecast flood at altered time lead (Sudheer et al„ 2002; Chang et ah, 2007). Similarly, neuro-fuzzy is an additional field of study that is effectively being used to forecast flood events (Nayak et ah, 2005; Singh, 2007).

Che et ah (2011) applied BPNN and GA for training FFNN for coping with weighing adjustment problem and compared their performance considering measurement indicators and experimental data. Outcomes showed that BPNN is superior to GA and had faster training speed than GA. Dar et ah (2015) presented applicability of ANN method to develop flood prediction model for a large-size catchment of Jhelum River in J&K. India. Based on statistical parameters, ANN proved to be a potential model in flood discharge prediction with reliability and accuracy in proposed area of study. Shiau and Hsu (2016) applied FFBPNN and radial basis function network (RBFN) linked with different time-lagged stream flow and precipitation inputs for extending short stream flow record at Lilin gauging site situated in Gaoping River basin, Taiwan. Findings revealed that proposed techniques performed suitably forextending daily stream flow records at the selected gauge station. Pandey and Srinivas (2015) investigated the potential of Auto Regressive Integrated Moving Average (ARIMA), FFBPNN, and RBFN for forecasting daily stream flow at Basantpur gauge station of River Mahanadi, Odisha. India. Results indicated that FFBPNN model performed better compared to other proposed models in daily stream flow forecasts. Wei and Hsu (2008) employed FFBPNN for estimating downstream water levels of River Tanshui located in Taiwan. Results from optimized model were compared with historical proceedings and observed that FFBPNN effectively helped to solve problems in controlling flood events. Arulmurugan and Anandakumar (2018) applied FFBPNN and wavelet transform classifier for recognizing and classifying lung nodules. Results indicated that FFBPNN method produced more accuracy in detecting lung cancer at early stage. Syahrullah and Sinaga (2016) explored the usability of FFBPNN for optimizing and predicting potential of motorcycle fuel injection system of gasoline. Investigation showed that proposed model accurately predicted the potential of motorcycle injection system. Mehr et al. (2015) studied usability of FFBPNN. generalized regression NN, and RBF to predict stream flow' in poor rain gauging stations considering monthly stream flow data of two succeeding stations on Coruh River, Turkey. Findings of present study revealed that RBFN performed better than FFBPNN and GRNN in proposed study area. Mehr et al. (2014) investigated efficacy of linear genetic programming (LGP), FFBPNN, GRNN, and RBFN techniques for stream flow prediction. Results specified that LGP technique provided more accurate results for successive station monthly stream flow' prediction compared to other ANN techniques. Konate et al. (2015) analysed and compared GRNN and FFBP to model porosity on four wells of Zhenjing oilfield data, China. Findings proved that GRNN made more precise and reliable porosity constraint estimation compared to FFBPNN.

Nowadays ANN is a successive tool commonly applied for prediction of various hydrologic parameters in different watersheds (Mohanta et al, 2020a; Samantaray and Ghose, 2019, 2020; Samantaray and Sahoo, 2020d; Samantaray et al., 2019a, 2019b, 2020b 2020c). Mukerji et al. (2009) used ANN, ANGIS, and ANFIS models for forecasting flood at Jamtara gauge station of River Ajay, Jharkhand, India, and comparative performance of proposed models was conducted. Results revealed that ANGIS predicted flood occurrences w'ith extreme accurateness followed by ANFIS and ANN models. Shu and Ouarda (2008) proposed ANFIS technique for providing regional flood estimation of Quebec province, Canada, and contrasted to ANN, NLR, and NLR with regionalization (NLR-R). Results indicated that ANFIS model generalized input parameters with much better capability compared to NLR and NLR-R models. Sehgal et al. (2014) developed wavelet ANFIS-split data and wavelet ANFIS-modified time series models for forecasting river w'ater levels of Kamla and Kosi River basins in India. Proposed models helped in forecasting river water levels precisely and when the two models w'ere compared, WANFIS-SD gave better performance compared to WANFIS-MS for high river stages. Chau et al. (2005) applied hybrid genetic algorithm based ANN model (ANN-GA) and ANFIS model to forecast flood events in a channel reach of River Yangtze, China. Rezaeianzadeh et al. (2014) used ANN, ANFIS. MLR, and MNLR to forecast peak flows on daily basis at exit of Khosrow Shirin catchment, situated in Fars State, Iran. Model performances were evaluated and outcomes revealed that considering area prejudiced rainfall as input for ANNs and MNLR and spatially dispersed rainfall as input for ANFIS and MLR gave more precise prediction performances. Kisi et al. (2012) evaluated accuracy of several data-driven methods, i.e. ANN, SVM. and ANFIS, to forecast daily intermittent stream flow's. Gong et al. (2016) applied ANN, SVM, and ANFIS for groundwater level prediction for w'ells near Lake Okeechobee in Florida, considering interface amid surface w'ater and groundwater. Results demonstrated that ANFIS and SVM models predicted groundwater level more accurately than ANN model and lake level variations w'ere found to be the key driving factor in groundw'ater level prediction. Dastorani et al. (2010) used data from bordering sites, ANN and ANFIS models to find missing data of selected gauge stations from different parts of Iran. Findings demonstrated that ANFIS presented a superior ability to predict missing flow data. Mehr et al. (2019) developed an amalgam model integrating SVR and FFA for one-month ahead precipitation forecasting at Tabriz and Urmia sites in northwest Iran. Results of hybrid model were compared with that of regular SVR and genetic programming models and found that SVR-FFA model outperformed other models in forecasting rainfall more accurately in the desired stations. Hussain and Khan (2020) explored potential of MLR SVR. and random forest (RF) to forecast river flow in Hunza region of Pakistan utilizing in situ dataset. Results showed that RF gave best test results compared to MLP and SVR and proved useful to forecast river flow with great accuracy. Shamseldin (2010) explored applicability of ANN to forecast Blue Nile river flows in Sudan. Results showed that ANN has substantial ability to forecast river flow in emerging countries. Yaseen et al. (2018) proposed a novel hybridized model called ANFIS-FFA to forecast monthly precipitation of Pahang River situated in Malaysia. Projected conjoint model is equated with regular ANFIS model, and results demonstrated that the hybrid model is ascertained to be a prudent modelling approach to simulate monthly precipitation in proposed study site. Zhou et al. (2019) applied ANN, SVM, ANFIS, hybridized ANFIS-FFA, and ANFIS-GA for predicting particle size distribution of muck piling after blast at different timescales in two pit mines located in Iran. On the basis of evaluation criteria, both ANFIS-FFA and ANFIS-GA performed agreeably, but ANFIS-GA model slightly gave better prediction results, making it a potential prediction model. Soft computing is a widely used technique for prediction of flood at a gauge station (Sahoo et al., 2019; Sahoo et al., 2020a, b). Shafiei et al. (2020) used ANFIS and hybrid ANFIS-FFA techniques for modelling discharge coefficient of labyrinth weirs and employed Monte Carlo simulations for enhancing capability of proposed models. ANFIS models were compared with that of computational fluid dynamics model and comparison results revealed that ANFIS-FFA conjoint model is significantly more accurate. Riahi-Madvar et al. (2020) proposed an ANFIS model hybridized with FFA for predicting contaminant dispersal coefficient in river beds using Subcategory Assortment by Maximum Distinction. On basis of results, projected amalgam model exhibited substantial developments than standard ANFIS in dispersion coefficient prediction. Tien-Bui et al. (2018) projected three novel amalgam models integrating ANFIS with bees (ANFIS-BA), invasive weed optimization (ANFIS-IWO), and cultural (ANFIS-CA) to map flood-susceptible areas in Haraz catchment, Iran. Results revealed that ANFIS-BA showed superior prediction ability followed by ANFIS-IWO and ANFIS-CA. Termeh et al. (2018) integrated ANFIS with GA, ant colony optimization (ACO), and particle swarm optimization (PSO) for flood vulnerability mapping over Jahrom Township in Fars Province, Iran, and compared their accurateness. ANFIS-PSO performed best as most practical model to produce extremely engrossed flood-prone maps.

The objective of this study is to utilize hybrid ANN modelling technique for flood forecasting at BP ghat and Dholai gauge stations of Barak River flowing through the Barak valley region in Assam, India. A comprehensive study is carried out to compare relative performances of ANN, ANFIS, and ANFIS-FFA models at proposed study area. Findings of present research could be useful for local and national governing bodies to plan for future and develop suitable new infrastructure to protect the lives and property.