Applicability of neural network models for real-time flood forecasting in dry zone and wet zone river basins, Sri Lanka

Abstract

Flood forecasting is a powerful tool for flood management and early warning, where the anticipated flow values are determined by incorporating basin attributes and climatic factors. In the field, data-driven models offer beneficial solutions compared to comprehensive physical and statistical tools; neural networks have evolved to perform flood forecasting without understanding the physical mechanism. However, forecasting efficiency and reliability are insufficient due to the augmentation of predictive span and improper data handling strategies. In addition, the poor interconnectivity of spatial-temporal resolution influences the accuracy of flood forecasting in a dry zone. Thus, the present study aimed to enhance the flood forecasting ability of neural network models for a 30-day horizon by learning the daily input series of climatic and physiographic factors of the catchment region. Further, the data manipulation strategies were adapted to enhance the learning capabilities. In addition, pretrained

models were developed based on the model performance in the wet zone basin to enhance the predictive quality in the dry zone basin. The NN models were developed for the Kelani River flood forecasting, where significant flood events have frequently destroyed the socio-economic features of the basin. Besides, pretrained

models were tested on the Maduru Basin flood events, which have encountered inundation due to prolonged flood peaks. Thus, climatic and physiographic data were gathered for both basins and improved with hydrological and data science-based data manipulation strategies. On the other hand, the Box-Cox transformation was employed to redistribute the input series into a Gaussian state to enhance the learning ability of NN models. Consecutive windows were proposed to consider 30-day daily input to forecast the next 30day streamflow values while sampling. Thirteen (13) NN models were compiled, fitted, and

tested on the Kelani Basin. In addition, grid analysis was adapted to rank the performance of

models based

on statistical tools, where bidirectional models explicated excellent quality in flood forecasting. Besides, uncertainty analysis was proposed to investigate the impacts of data handling and input combination on flood forecasting. Two hybrid models significantly expounded underperformance without box-cox transformation; none of the models illustrated excellent performance without box-cox transformation. Moreover, scaling/normalization severely influenced the model performance considerably for hybrid models. Besides, sensitivity analysis was applied to verify the applicability of model architecture on model performance. Unlike the types of optimizers, other sensitivity parameters revealed inconclusive results for model performance. None of the modified models delivered more excellent performance than the core models. Further, Bidirectional Gated Recurrent Unit (BiGRU),

Bidirectional Long- and Short-Term Model (Bi-LSTM), and Attention Based BiLSTM

(Att-BiLSTM) expressed 0.98, 0.95, and 0.97 for the wet zone flood forecasting, respectively, which were chosen as pre-trained models delivered a similar performance for the dry basin. In future studies, the consecutive data batches must be determined according to the guiding parameters, such as global warming and climate change. Besides, the loss function should be replaced with other statistical terms to incorporate an optimizer, and autocorrelation must be adapted to control the error propagation. In addition, the core model must be trained for extended periods to effectively perform transfer learning on other basins.