Flood simulation using parallel genetic algorithm integrated wavelet neural networks

Abstract

The conventional means of flood simulation and prediction using conceptual hydrological model or artificial neural network (ANN) has provided promising results in recent years. However, it is usually difficult to obtain ideal flood reproducing due to the structure of hydrological model. Back propagation (BP) algorithm of ANN may also reach local optimum when training nodal weights. To improve the mapping capability of neural networks, wavelet function was adopted (WANN) to strengthen the non-linear simulation accuracy and generality. In addition, genetic algorithm is integrated with WANN (GAWANN) to avoid reaching local optimum. Meanwhile, Message Passing Interface (MPI) subroutines are introduced for distributed implement considering the time consumption during nodal weights training. The GAWANN was applied in the flood simulation and prediction in arid area. The test results of 4 independent cases were compared to reveal the relations between historical rainfall and runoff under different time lags. The simulation was also carried out with Xinanjiang model to demonstrate the capability of GAWANN. The numerical experiments in this paper indicated that the parallel GAWANN has strong capability of rain-runoff mapping as well as computational efficiency and is suitable for applications of flood simulation in arid areas.

Introduction

Flood simulation and prediction is one of the most active researching areas in surface water hydrology. Flood takes place whenever there is a heavy or a long period of precipitation. An accurate prediction of flood under changeable meteorological and layer conditions can not only help in the water resources management especially in hydropower, but also reduce the loss of lives and property to the minimum in floodplain areas. While with the fast increasing economic activities in floodplain and rivers especially in arid areas in northeast China, further requirements are raised on more accurate flood prediction precision.

The actual rain-runoff (RR) process is such a non-linear problem as until now no explicit formula can describe the process perfectly. During the last decades, great progress in flood prediction has been made by taking several techniques such as empirical model, statistical model, and physical based conceptual or distributed hydrological model into account. Empirical models can provide hydrograph in a special basin but long time series of observations are badly needed before carrying out simulation [1]. Besides, empirical methods are always not capable of generating a runoff hydrograph with complete information on timing of peak. Physical based hydrological model are more popular since the fast development in computer science and technology in the 1940s [2], [3], [4], [5]. Newly developing spatial technology like remote sensing (RS) can also support physical based models by providing large sets of spatial data such as leaf area index (LAI) and land use category, even soil moisture [6], [7], [8]. On the other hand, hydrological models commonly used in flood prediction can be divided into two categories, i.e. conceptual and distributed model. Developed by Zhao [9], [10], [11], Xinanjiang model is a most widely accepted conceptual lumped hydrological model that performs in flood simulation especially suitable in wet areas. The model was conceptualized to divide the runoff into surface and ground flow using Horton's theory [12]. Lin et al. [13], [14] coupled Xinanjiang model and meteorological system in flood prediction of Huaihe River Basin. Compared to distributed hydrological models, less data inputs are needed in conceptual model. Inputs of such conceptual model are mainly about precipitation, evapotranspiration, etc. Researches on distributed hydrological models that started in the 1960s have been applied in many major basins around the world due to its strong power in taking the under-layer viabilities into consideration [15]. Braud et al. analyzed the flash flood event using two distributed hydrological model: CVN and MARINE. Different terms of the two models are discussed in the peak discharges [16]. Results of such models are generally ideal; however many of them involve ecology, atmosphere, human activities, and can also be affected by basic science including physics, bio-process, or chemistry. Others may differ or orient for various simulation purposes [17]. On the other hand, physics based models need a great number of spatial and observation data as inputs to implement simulation and calibration including temperature, wind speed, sunshine hour, etc.

Compared to the aforementioned methods, black-box models, artificial neutral network (ANN), have the advantage of fast data fitting and have became the preferred approach since its development in the 1980s. Ju et al. [18] employ division-based back-propagation neural networks in rainfall–runoff simulation and compared with the Xinanjiang conceptual RR hydrological model. Ahmad et al. [19] adopted ANN model for prediction peak flow in Red River timing and shape of runoff hydrograph together with the employment of meteorological parameters including antecedent precipitation index and melt index.

In addition to the application using ANN issues above, many improvements have also been made to strengthen the performance of ANN. This includes integrating data preprocessing techniques with ANN, such as moving average (MA) and singular spectrum analysis (SSA) [20]. The coupled ANN has the ability to get rid of the effects of white noise that may add errors to the weights training. For another instance, in order to improve the drawbacks of the conventional optimal process, Chen and Chang [21] proposed a novel evolutionary artificial neural network (EANN) for time series forecasting.

In flood prediction, however, a much higher runoff time series forecasting accuracy is often requested. Besides, the forecasting ability, especially the forecasting interval length, is also a significant criterion to be taken into consideration. In this paper, an ANN model hybrid wavelet function was proposed and applied in the simulation and prediction process of flood period in Nen River Basin of China. The nodal weight training of ANN model was improved by the hybrid of parallel genetic algorithm to reach global optimal before carrying out BP at the last iteration. Forecasting ability was then further discussed by taking three different prediction solutions. The objective of this study lies in (1) developing wavelet artificial neural networks and their applications to flood simulation in arid floodplain of Northern China and (2) enhancing the computational efficiency of GAWANN using MPI technique.

The remaining paper outlines a framework for developing a flood prediction system using genetic algorithm hybrid back propagation wavelet neutral networks and can be organized as follows. Section 2 gives a brief introduction to the WANN and the improvements done by integrating genetic algorithm with MPI technology. Section 3 follows the study case with data preprocessing. Results of different prediction solutions were drawn with further discussions and comparisons in Section 4. Meanwhile simulation results of the Xinanjiang model were also used as a comparison to GAWANN. In the last section, conclusion and remarks are summarized.