Short term river water level forecasting using artificial neural network

Abstract

An artificial neural network (ANN) is a flexible mathematical structure, which is capable of modelling the complex, non-linear, and dynamic relationships between the input and output data scts. ANN models have bcen found useful and efficient, , particularly in problems for which the characteristics of the processes are difficult to describe using physical equations. Moreover, artificial neural networks (ANNs) belong to a class of data-driven modelling, therefore, the quality and consistency of water level data is an important consideration for time series modelling using ANNs. This study highlights the usc of ANN in real-time forecasting of water level at a given site continuously throughout the year based on the water level time history data at the same site. An error back propagation algorithm with gradient descent optimisation technique and adaptive learning rate for the construction as well as validation and verification of the neural network model has been used. Computer software named 'W ALF' is also developed in MicrosoftTM Visual Basic™ based on the error back propagation methodology as the part of this study. The main objective of the study is to explore the sensitivity of the developed model to the calibration season and forecasting lead-time. The non-Iincar watcr level records of 18 years at Meghna Ferry Ghat of Meghna Rivcr havc bccn used. In addition, the optimisation procedure is further extended to the determination of the network control parameters for the supervised feed forward network. Moreover, detail methodology tor the selection of representative calibration, validation, and verification sets is also presented. This study also presents the important aspects of the validation and verification of ANN models including the selection of performance measures and analysis of residual errors for supervised lcarning nctworks. Thc gcneralization error of the ANN forecasting model have been estimated using the traditional (modified) train and test methodology. The pcrror1nance mcasurcs, which arc uscd in this study, are root mean squared crror, mcan squarcd crror, mean absolutc crror, coefficicnt of efficiency, and correlation coefficient. This study reveals that the ANN does provide a viable and effective approach for developing input-output simulation and forecasting models in situation that do not require modelling of thc internal structure of the river basin. The consistency and quality checking of the selected stations and that of the water level records provides the quality input-output data for constructing the ANN model. Moreover, the ANNs are purely empirical models; therefore, validation is critical to its operational success. While validating several neural networks using the traditional (modified) train and test methodology, the calibrated or trained model shows good generalization performance on verification sets. The analysis of residuals of the calibration and test sets can warn the user of the phenomena of overlearning or overfitting and provide an overall distribution of performance measures. To explore the choice of appropriate performance measures, it is found that correlation coefficient as a single correlation index between actual and predicted output shows spurious correlation with the increasing training tolerance due to clustering of forecasted and observed data. The coefficient of efficiency is a better choice in this respect, which can represent the true performance of the forecasting model. The optimisation technique used in this study for the optimisation of ANN topology, parameters (weights), and network control parameters significantly reduces the operational cost of the water Icvcl forecasting model in terms of required amount of information and execution time. An optimised three-layered neural network topology is obtained which is designed as 2 input neurons in the input layer, and 4 and 12 hidden neurons in the first and second hidden layers respectively. To minimize the erroneous effect of the initial conditions in the process of model construction (i.e. initial random valucs of ANN parametcrs), scvcral numbers of individual runs are considered with different initial conditions and the best learned network is considered for watcr level forceasting. Sensitivity analysis of thc model to thc calibration season demonstrates that the choice of content rather than length of the calibration and validation period does have a significant effcct on the forecasting ability ofthe ANNs. It is revealed from the study that the ANN models should be calibrated and validated for a common season. In case of sparse data situation, wet period trained model can be used for the real time forecasting of any hydrologic year data but dry period calibration model should only be used for the prediction of dry period data. The forecasted values of water levels obtained show a very high degree of accuracy up to 7-lead-days within the :1:0.5 m error bound but as the lead-time increases, such as in case of 10 or IS-lead-days, the performance of ANN forecasting model also decreases rapidly. Moreover, plots of residual errors for the predicted and observed water level do not show any systematic distributions although there arc noticeable large errors for the rising and falling limb between the predicted and observed water level hydrographs. This result signifies that the large error will be induced in case of the prediction of watcr levels at beginning and ending of the monsoon period in Bangladesh. Thc foregoing studies shows that short term forecasting of river water levels in realtime sense is possible through the use of neural networks. Moreover, it also significantly reduces the unnecessary data collection and operational time.