A study of optimal model lag and spatial inputs to artificial neural network for rainfall forecasting

Abstract

Artificial neural networks (ANNs), which emulate the parallel distributed processing of the human nervous system, have proven to be very successful in dealing with complicated problems, such as function approximation and pattern recognition. Due to their powerful capability and functionality, ANNs provide an alternative approach for many engineering problems that are difficult to solve by conventional approaches. Rainfall forecasting has been a difficult subject in hydrology due to the complexity of the physical processes involved and the variability of rainfall in space and time. In this study, ANNs were adopted to forecast short-term rainfall for an urban catchment. The ANNs were trained to recognise historical rainfall patterns as recorded from a number of gauges in the study catchment for reproduction of relevant patterns for new rainstorm events. The primary objective of this paper is to investigate the effect of temporal and spatial information on short-term rainfall forecasting. To achieve this aim, a comparison test on the forecast accuracy was made among the ANNs configured with different orders of lag and different numbers of spatial inputs. In developing the ANNs with alternative configurations, the ANNs were trained to an optimal level to achieve good generalisation of data. It was found in this study that the ANNs provided the most accurate predictions when an optimum number of spatial inputs was included into the network, and that the network with lower lag consistently produced better performance.

Introduction

Flash flooding is a life-threatening phenomenon. One way to reduce the risk to life and to alleviate economic losses is to provide advance warnings to people likely to be affected by flash floods. A flood warning system consists of three interlinked modules:

• a data collection module which collects rainfall data at strategic locations within the catchment;

• a forecasting module that forecasts the rainfall based on observed rainfall and other relevant hydro-meteorological factors; and

• a rainfall–runoff module that translates the observed and forecast rainfalls into corresponding flow values.

While a number of studies have been conducted to ascertain the utility of alternative model approaches in the third (rainfall–runoff) module, studies investigating alternative approaches for the second (rainfall forecasting) module are few.

Rainfall forecasts at short (5–30 min) intervals are of added importance in the case of small, urbanised catchments (Lettenmaier and Wood, 1993). Such catchments are characterised by a fast hydrologic response due to their size and a relatively high fraction of effective impervious surface. Developing a flood warning system for such catchments is not a simple task. While the process of converting rainfall to runoff poses a series of difficulties in itself, forecasting future rainfall on a small spatial scale is fraught with errors of a considerably higher magnitude. Although a physically based approach for rainfall forecasting has several advantages, given the short time scale, the small catchment area, and the massive costs associated with collecting the required meteorological data, it is not a feasible alternative in most cases. A statistically based approach that attempts to model the pattern of the underlying physical attributes manifested in the observed rainfall data is an efficient alternative. Proposed herein is one such alternative to forecast the future spatial distribution of rainfall over a catchment using an Artificial Neural Network (ANN).

The rainfall forecasting approach presented herein has been discussed previously by Luk et al. (1999) who described the use and implementation of alternate forms of ANNs to the rainfall forecasting problem. The approach presented by Luk et al. (1999) is used in this study to investigate the necessary spatial and temporal data for the ANN.

The ANN approach has evolved as a branch of artificial intelligence and is now a recognised tool to model underlying complexities in any artificial or physical system. The ANN is based on a model of the human neurological system which consists of a series of basic computing elements (called neurons) interconnected together to allow recognition of incidents that have had a similar pattern to the current input. With this parallel-distributed processing architecture, ANNs have proven to be very powerful computational tools that excel in pattern recognition and function approximation. As shown by Hornik et al. (1989), an ANN with sufficient complexity is capable of approximating any smooth function to any desired degree of accuracy. In addition, ANNs are computationally robust, having the ability to learn and to generalise from examples to produce meaningful solutions to problems even when the input data contain errors or are incomplete. The application of an ANN, however, involves a complicated development process. If carelessly used, an ANN can easily “learn” irrelevant information (noises) in the system, with the resulted ANN model being able to predict past incidents but unable to predict future events.

The ANN modelling framework has been used increasingly in various aspects of science and engineering because of its ability to model both linear and nonlinear systems without the need to make any assumptions as are implicit in most traditional statistical approaches. Some of the hydrologic problems ANNs have been used for include rainfall–runoff modelling (Hsu et al., 1995, Smith and Eli, 1995, Achela et al., 1998), scheduling of hydroelectric power systems (Saad and Bigras 1996), and river flow prediction (Karunanithi et al., 1994, Zhu and Fujita, 1994). The ANN methodology has been applied also to forecast rainfall; for example, French et al. (1992) used synthetically generated rainfall storms to calibrate an ANN model and then generated plausible rainfall sequences that could occur over a catchment using a physically based rainfall to validate the ANN. While similar in spirit to the French et al. (1992) paper, this study focuses on the problems faced in developing an ANN based rainfall forecasting model using observed rainfall records in both space and time. As such, this paper aims at identifying an optimal set of spatio-temporal inputs for an ANN rainfall forecasting model.