Intermittent Streamflow Forecasting and Extreme Event Modelling using Wavelet based Artificial Neural Networks

Forecasting of intermittent stream flow is necessary for water resource planning and management at catchment scale. Forecasting of extreme events and events outside the range of training data used for artificial neural network (ANN) model development has been a major bottleneck in their generalization capabilities till date. Despite of several studies using wavelet analysis in water resource modelling, no study has yet been conducted to explore capabilities of hybrid ANN modelling techniques for extreme events outside the training range. In this study a wavelet based ANN model (WANN) is proposed for intermittent streamflow forecasting and extreme event modelling. This study is carried out in a watershed in semi arid middle region of Gujarat, India. 6 years of hydro-climatic data are used in this study. 4 years of data are used for model training, 1 year for cross-validation and remaining 1 year data are used to evaluate the effectiveness of the WANN model. Two different approaches of data arrangement are considered in this study, in one approach testing data are within the range of training dataset, whereas in another approach testing data are outside the training dataset range. Performance of four different training algorithms and different types of wavelet functions are also evaluated for WANN model development. In this study it is found that WANN model performed significantly better than standard ANN models. It is observed in this study that different wavelet functions have different role in modelling complexities of normal and extreme events. WANN model simulated peak values very well and it shows that WANN model has the potential to be applied successfully for intermittent streamflow forecasting even for the data outside the training range and for extreme events.     

Wavelet Bootstrap Multiple Linear Regression Based Hybrid Modeling for Daily River Discharge Forecasting

A new hybrid model, the wavelet–bootstrap–multiple linear regression (WBMLR) is proposed to explore the potential of wavelet analysis and bootstrap resampling techniques for daily discharge forecasting. The performance of the developed WBMLR model is also compared with five more models: multiple linear regression (MLR), artificial neural network (ANN), wavelet-based MLR (WMLR), wavelet-based ANN (WANN) and wavelet–bootstrap–ANN (WBANN) models. Seven years of discharge data from seven gauging stations in the middle reaches of Mahanadi river basin in India are applied in this study. Significant input vectors are decomposed into discrete wavelet components (DWCs) using discrete wavelet transformation (DWT) to generate wavelet sub time series that are used as inputs to the MLR and ANN models to develop the WMLR and WANN models, respectively. Effective wavelets are selected by considering several types of wavelets with different vanishing moments. WBMLR and WBANN models are developed as ensemble of different WMLR and WANN models, respectively, developed using different realizations of the training dataset generated using bootstrap resampling technique. The results show that the wavelet bootstrap hybrid models (i.e. WBMLR and WBANN) produce significantly better results than the traditional MLR and ANN models. Hybrid models based on MLR (WMLR, WBMLR) perform better than the ANN based hybrid models (WBANN, WANN). The WBMLR and WMLR models simulate the peak discharges better than the WBANN, WANN, MLR and ANN models, whereas the overall performance of WBMLR model is found to be more accurate and reliable than the remaining five models.     

Wavelet Regression Model as an Alternative to Neural Networks for River Stage Forecasting

River stage forecasting is an important issue in water resources management and real-time prediction of extreme floods. The present study investigates the performance of the wavelet regression (WR) technique in daily river stage forecasting. The WR model was improved combining two methods, discrete wavelet transform and a linear regression model. Two different WR models were developed using the stage sub-time series, and these were compared with each other. The data from two stations on the Schuylkill River in Philadelphia were used. The root mean square errors (RMSE), mean absolute errors (MAE) and correlation coefficient (R) statistics were used for evaluating the accuracy of the WR models. The accuracy of the WR models was then compared with those of the artificial neural networks (ANN) models. Based on a comparison of these results, the WR models were found to perform better than the ANN models. For the upstream and downstream stations, it was found that the WR models with upstream readings of with RMSE = 0.070, MAE = 0.027, R = 0.937 and with downstream readings of RMSE = 0.048, MAE = 0.024, R = 0.969 in the validation stage performed better in forecasting daily river stages than the best accurate ANN models with upstream readings of RMSE = 0.168, MAE = 0.052, R = 0.802 and with downstream readings of RMSE = 0.115, MAE = 0.051, R = 0.807, respectively.     

Lake Level Prediction using Feed Forward and Recurrent Neural Networks

The protection of high quality fresh water in times of global climate changes is of tremendous importance since it is the key factor of local demographic and economic development. One such fresh water source is Vrana Lake, located on the completely karstified Island of Cres in Croatia. Over the last few decades a severe and dangerous decrease of the lake level has been documented. In order to develop a reliable lake level prediction, the application of the artificial neural networks (ANN) was used for the first time. The paper proposes time-series forecasting models based on the monthly measurements of the lake level during the last 38 years, capable to predict 6 or 12 months ahead. In order to gain the best possible model performance, the forecasting models were built using two types of ANN: the Long-Short Term Memory (LSTM) recurrent neural network (RNN), and the feed forward neural network (FFNN). Instead of classic lagged data set, the proposed models were trained with the set of sequences with different length created from the time series data. The models were trained with the same set of the training parameters in order to establish the same conditions for the performance analysis. Based on root mean squared error (RMSE) and correlation coefficient (R) the performance analysis shown that both model types can achieve satisfactory results. The analysis also revealed that regardless of the model types, they outperform classic ANN models based on datasets with fixed number of features and one month the prediction period. Analysis also revealed that the proposed models outperform classic time series forecasting models based on ARIMA and other similar methods .

     

Coupling Singular Spectrum Analysis with Least Square Support Vector Machine to Improve Accuracy of SPI Drought Forecasting

The study proposed a Standardized Precipitation Index (SPI) drought forecasting model based on singular spectrum analysis (SSA) and single least square support vector machine (LSSVM) with a twofold investigation: (Beguería et al. Int J Climatol, 34(10): 3001–3023, 2014) the forecasting performance of the LSSVM-based model with or without coupling SSA and (Belayneh et al. J Hydrol, 508: 418–429, 2014) the model performances by using different inputs (i.e., antecedent SPIs and antecedent accumulated monthly rainfall) preprocessed by SSA. For the first part investigation, the LSSVM-based model using antecedent SPI as input (LSSVM1) and the LSSVM-based model coupling with SSA using antecedent SPI as input (SSA-LSSVM2) were developed. For the second part of investigation, the SSA-LSSVM-based model using antecedent accumulated monthly rainfall as input (SSA-LSSVM3) was developed and compared to SSA-LSSVM2. The drought indices (SPI3 and SPI6) were chosen as the outputs of the SPI drought forecasting models. The Tseng-Wen reservoir catchment in southern Taiwan was selected to test the aforementioned models. The results show that the forecasting performance of SSA-LSSVM2 is better than that of LSSVM1, which means the input data preprocessed by SSA can significantly increase the accuracy of the SPI drought forecasting. In addition, the performance comparison between SSA-LSSVM2 and SSA-LSSVM3 indicates that using antecedent accumulated monthly rainfalls (i.e., 3-month and 6-month accumulated rainfalls) as input of SSA-LSSVM3 are much better than using antecedent SPIs (i.e., SPI3 and SPI6) as input of SSA-LSSVM2. SSA-LSSVM3 is found to be the most appropriate model for SPI drought forecasting in the case study.

     

Forecasting Urban Water Demand Via Wavelet-Denoising and Neural Network Models. Case Study: City of Syracuse, Italy

Forecasting urban water demand can be of use in the management of water utilities. For example, activities such as water-budgeting, operation and maintenance of pumps, wells, reservoirs, and mains require quantitative estimations of water resources at specified future dates. In this study, we tackle the problem of forecasting urban water demand by means of back-propagation artificial neural networks (ANNs) coupled with wavelet-denoising. In addition, non-coupled ANN and Linear Multiple Regression were used as comparison models. We considered the case of the municipality of Syracuse, Italy; for this purpose, we used a 7?year-long time series of water demand without additional predictors. Six forecasting horizons were considered, from 1 to 6?months ahead. The main objective was to implement a forecasting model that may be readily used for municipal water budgeting. An additional objective was to explore the impact of wavelet-denoising on ANN generalization. For this purpose, we measured the impact of five different wavelet filter-banks (namely, Haar and Daubechies of type db2, db3, db4, and db5) on a single neural network. Empirical results show that neural networks coupled with Haar and Daubechies?? filter-banks of type db2 and db3 outperformed all of the following: non-coupled ANN, Multiple Linear Regression and ANN models coupled with Daubechies filters of type db4 and db5. The results of this study suggest that reduced variance in the training-set (by means of denoising) may improve forecasting accuracy; on the other hand, an oversimplification of the input-matrix may deteriorate forecasting accuracy and induce network instability.     

River Flow Forecasting using Recurrent Neural Networks

Forecasting a hydrologic time series has been one of the most complicated tasks owing to the wide range of data, the uncertainties in the parameters influencing the time series and also due to the non availability of adequate data. Recently, Artificial Neural Networks (ANNs) have become quite popular in time series forecasting in various fields. This paper demonstrates the use of ANNs to forecast monthly river flows. Two different networks, namely the feed forward network and the recurrent neural network, have been chosen. The feed forward network is trained using the conventional back propagation algorithm with many improvements and the recurrent neural network is trained using the method of ordered partial derivatives. The selection of architecture and the training procedure for both the networks are presented. The selected ANN models were used to train and forecast the monthly flows of a river in India, with a catchment area of 5189 km² up to the gauging site. The trained networks are used for both single step ahead and multiple step ahead forecasting. A comparative study of both networks indicates that the recurrent neural networks performed better than the feed forward networks. In addition, the size of the architecture and the training time required were less for the recurrent neural networks. The recurrent neural network gave better results for both single step ahead and multiple step ahead forecasting. Hence recurrent neural networks are recommended as a tool for river flow forecasting.     

Daily Mean Streamflow Prediction in Perennial and Non-Perennial Rivers Using Four Data Driven Techniques

This study examines and compares the performance of four new attractive artificial intelligence techniques including artificial neural network (ANN), hybrid wavelet-artificial neural network (WANN), Genetic expression programming (GEP), and hybrid wavelet-genetic expression programming (WGEP) for daily mean streamflow prediction of perennial and non-perennial rivers located in semi-arid region of Zagros mountains in Iran. For this purpose, data of daily mean streamflow of the Behesht-Abad (perennial) and Joneghan (non-perennial) rivers as well as precipitation information of 17 meteorological stations for the period 1999–2008 were used. Coefficient of determination (R²) and root mean square error (RMSE) were used for evaluating the applicability of developed models. This study showed that although the GEP model was the most accurate in predicting peak flows, but in overall among the four mentioned models in both perennial and non-perennial rivers, WANN had the best performance. Among input patterns, flow based and coupled precipitation-flow based patterns with negligible difference to each other were determined to be the best patterns. Also this study confirmed that combining wavelet method with ANN and GEP and developing WANN and WGEP methods results in improving the performance of ANN and GEP models.     

Monthly Streamflow Forecasting Using Convolutional Neural Network

Monthly streamflow forecasting is vital for managing water resources. Recently, numerous studies have explored and evidenced the potential of artificial intelligence (AI) models in hydrological forecasting. In this study, the feasibility of the convolutional neural network (CNN), a deep learning method, is explored for monthly streamflow forecasting. CNN can automatically extract critical features from numerous inputs with its convolution–pooling mechanism, which is a distinct advantage compared with other AI models. Hydrological and large-scale atmospheric circulation variables, including rainfall, streamflow, and atmospheric circulation factors are used to establish models and forecast streamflow for Huanren Reservoir and Xiangjiaba Hydropower Station, China. The artificial neural network (ANN) and extreme learning machine (ELM) with inputs identified based on cross-correlation and mutual information analyses are established for comparative analyses. The performances of these models are assessed with several statistical metrics and graphical evaluation methods. The results show that CNN outperforms ANN and ELM in all statistical measures. Moreover, CNN shows better stability in forecasting accuracy.

     

Predicting Monsoon Floods in Rivers Embedding Wavelet Transform,Genetic Algorithm and Neural Network

Monsoon floods are recurring hazards in most countries of South-East Asia. In this paper, a wavelet transform-genetic algorithm-neural network model (WAGANN) is proposed for forecasting 1-day-ahead monsoon river flows which are difficult to model as they are characterized by irregularly spaced spiky large events and sustained flows of varying duration. Discrete wavelet transform (DWT) is employed for preprocessing the time series and genetic algorithm (GA) for optimizing the initial parameters of an artificial neural network (ANN) prior to the network training. Depending on different inputs, four WAGANN models are developed and evaluated for predicting flows in two Indian Rivers, the Kosi and the Gandak. These rivers are infamous for carrying large flows during monsoon (June to Sept), making the entire North Bihar of India unsafe for habitation or cultivation. When compared, WAGANN models are found to be better than autoregression models (ARs) and GA-optimized ANN models (GANNs) which use original flow time series (OFTS) for inputs, in simulating river flows during monsoon. In addition, WAGANN models predicted relatively reasonable estimates for the extreme flows, showing little bias for underprediction or overprediction.     

A Comparative Study of Various Hybrid Wavelet Feedforward Neural Network Models for Runoff Forecasting

Considering network topologies and structures of the artificial neural network (ANN) used in the field of hydrology, one can categorize them into two different generic types: feedforward and feedback (recurrent) networks. Different types of feedforward and recurrent ANNs are available, but multilayer perceptron type of feedforward ANN is most commonly used in hydrology for the development of wavelet coupled neural network (WNN) models. This study is conducted to compare performance of the various wavelet based feedforward artificial neural network (ANN) models. The feedforward ANN types used in the study include the multilayer perceptron neural network (MLPNN), generalized feedforward neural network (GFFNN), radial basis function neural network (RBFNN), modular neural network (MNN) and neuro-fuzzy neural network (NFNN) models. The rainfall-runoff data of four catchments located in different hydro-climatic regions of the world is used in the study. The discrete wavelet transformation (DWT) is used in the present study to decompose input rainfall data using db8 wavelet function. A total of 220 models are developed in this study to evaluate the performance of various feedforward neural network models. Performance of the developed WNN models is compared with their counterpart simple models developed without applying wavelet transformation (WT). The results of the study are further compared with - multiple linear regression (MLR) model which suggest that the WNN models outperformed their counterpart simple models. The hybrid wavelet models developed using MLPNN, the GFFNN and the MNN models performed best among the six selected data driven models explored in the study. Moreover, performance of the three best models is found to be similar and thus the hybrid wavelet GFFNN and the MNN models can be considered as an alternative to the most commonly used hybrid WNN models developed using MLPNN. The study further reveals that the wavelet coupled models outperformed their counterpart simple models only with the parsimonious input vector.     

Hydrologic Data Exploration and River Flow Forecasting of a Humid Tropical River Basin Using Artificial Neural Networks

The applicability of artificial neural networks (ANN) for modelling of daily river flows in a humid tropical river basin with seasonal rainfall pattern is investigated and the model performance assessed using the commonly adopted efficiency indices. Although the developed model showed satisfactory results for rainy period, the predicted hydrograph for the low flow period deviate from the observed data considerably. The rainfall and discharge data available for modelling is explored using Self Organizing Maps (SOM) and the subset of data having definite relationship between the selected hydrologic variables identified. The alternate approach for modelling of river flows utilising the knowledge from SOM analysis has improved the model results. The results show that ANN models can be adopted for forecasting of river flows in the humid tropical river basins for the monsoon period. Input data exploration using SOM is found helpful for developing logically sound ANN models.     

坡面产流模式的神经网络模拟

坡面产流是土壤本身特性与外界影响因素相互作用的结果，它们之间具有明显的非线性输入输出关系。在分析坡面产流和神经网络模型具有某些相似的基础上，利用径流站观测资料，建立了小流域坡面产流量的三层前向网络模型（BP算法），并显示了具有较好的模拟预测效果。     

Spatiotemporal Analysis of Droughts Over Different Climate Regions Using Hybrid Clustering Method

Assessment of spatiotemporal variations of drought is an efficient method for implementing drought mitigation strategies and reducing its negative impacts. This study aimed to assess the spatiotemporal pattern of short- to long-term droughts for an area with different climates. Therefore, 31 stations located in Iran were selected and the Standardized Precipitation Index (SPI) series with timescales of 3, 6, and 12 months were computed during the 1951-2016 period. A hybrid methodology including Maximal Overlap Discrete Wavelet Transform (MODWT) and K-means methods was used for obtaining the SPIs time-frequency properties and multiscale zoning of the area. The energy amounts of the decomposed subseries via the MODWT were applied as inputs for K-means. Also, the statistics in drought features (i.e. drought duration, severity, and peak) were assessed and the results showed that shorter term droughts (i.e. SPI-3 and -6) were more frequent and severe in the northern parts where the lowest values were obtained for drought duration. It was observed that the regions with more droughts frequency had the highest energy values. For shorter term droughts a direct relationship was obtained between the energy values and the mean SPIs, drought severity, and drought peak, whereas an inverse relationship was obtained for longer term drought. It was found that increasing the degree of SPI led to more similarity between the stations of each cluster. Also, the homogeneity of stations for the SPI-12 was slightly higher than the SPI-3 and -6.

     

Comparison of performance of statistical models in forecasting monthly streamflow of Kizil River,China

This paper presents the application of autoregressive integrated moving average(ARIMA),seasonal ARIMA(SARIMA),and Jordan-Elman artificial neural networks(ANN)models in forecasting the monthly streamflow of the Kizil River in Xinjiang,China.Two different types of monthly streamflow data(original and deseasonalized data)were used to develop time series and Jordan-Elman ANN models using previous flow conditions as predictors.The one-month-ahead forecasting performances of all models for the testing period(1998-2005)were compared using the average monthly flow data from the Kalabeili gaging station on the Kizil River.The Jordan-Elman ANN models,using previous flow conditions as inputs,resulted in no significant improvement over time series models in one-month-ahead forecasting.The results suggest that the simple time series models(ARIMA and SARIMA)can be used in one-month-ahead streamflow forecasting at the study site with a simple and explicit model structure and a model performance similar to the Jordan-Elman ANN models.     

Forecasting Surface Water Level Fluctuations of Lake Van by Artificial Neural Networks

Lake Van in eastern Turkey has been subject to water level rise during the last decade and, consequently, the low-lying areas along the shore are inundated, giving problems to local administrators, governmental officials, irrigation activities and to people's property. Therefore, forecasting water levels of the Lake has started to attract the attention of the researchers in the country. An attempt has been made to use artificial neural networks (ANN) for modeling the temporal change water levels of Lake Van. A back-propagation algorithm is used for training. The study indicated that neural networks can successfully model the complex relationship between the rainfall and consecutive water levels. Three different cases were considered with the network trained for different arrangements of input nodes, such as current and antecedent lake levels, rainfall amounts. All of the three models yields relatively close results to each other. The neural network model is simpler and more reliable than the conventional methods such as autoregressive (AR), moving average (MA), and autoregressive moving average with exogenous input (ARMAX) models. It is shown that the relative errors for these two different models, are below 10% which is acceptable for engineering studies. In this study, dynamic changes of the lake level are evaluated. In contrast to classical methods, ANNs do not require strict assumptions such as linearity, normality, homoscadacity etc.     

Streamflow Forecasting Using Four Wavelet Transformation Combinations Approaches with Data-Driven Models: A Comparative Study

This study investigates the use of wavelet transformation (WT) as preprocessing tool in data-driven models (DDMs) for forecasting streamflow 7 days ahead. WT used are Continuous wavelet transformation (CWT), discrete wavelet transformation (DWT), and a new proposed combination of CWT and DWT, namely discrete continuous wavelet transformation (DCWT). In addition to these three different WTs, the single DDMs were used also to create four different schematic layouts. The DDMs applied were artificial neural network (ANN), adaptive neuro-fuzzy inference system (ANFIS), and support vector machines (SVM). The lagged rainfall, temperature, and streamflow were incorporated as inputs into the WT-DDMs. It was found that CWT improved the forecasting accuracy of models which only included the rainfall and temperature but not the streamflow. Moreover, DWT improved the performance dramatically for the models with streamflow. Notably, DWT layout outperformed CWT layout in general while CWT layouts resulted in higher improvement to the models with rainfall and temperature only. The proposed DCWT in which CWT applied on the rainfall and temperature variables and DWT applied on the streamflow improved the forecasting ability in several models combinations when ANN was applied. Nevertheless, improvement in the forecasting accuracy was deteriorated in those with SVM while no improvement was observed with ANFIS. ANN outperformed both ANFIS and SVM while ANFIS performed better than SVM.     

River Discharges Forecasting In Northern Iraq Using Different ANN Techniques

The Upper and Lower Zab Rivers are two of main and most important tributaries of Tigris River in Northern Iraq region. They supply Tigris River with more than 40 % of its yield. The forecasting of flows for these rivers is very important in operation of the existing Dokan Dam on the Lower Zab River and the proposed Bakhma Dam on the Upper Zab River for flood mitigation and also in drought periods. Three types of Artificial Neural Networks (ANNs) are investigated and evaluated for flow forecasting of both rivers. The ANN techniques are the feedforward neural networks (FFNN), generalized regression neural networks (GRNN), and the radial basis function neural networks (RBF). The networks’ performance varied with different cases involved in the study; however, the FFNN was almost better than other networks. The effect of including a time index within the inputs of the networks is investigated. In addition, the ANNs’ performance is investigated in forecasting the high and low peaks and in forecasting river flows using the data of the other river.     

Short-Term Water Demand Forecast Modelling at IIT Kanpur Using Artificial Neural Networks

The efficient operation and management of an existing water supply system require short-term water demand forecasts as inputs. Conventionally, regression and time series analysis have been employed in modelling short-term water demand forecasts. The relatively new technique of artificial neural networks has been proposed as an efficient tool for modelling and forecasting in recent years. The primary objective of this study is to investigate the relatively new technique of artificial neural networks for use in forecasting short-term water demand at the Indian Institute of Technology, Kanpur. Other techniques investigated in this study include regression and time series analysis for comparison purposes. The secondary objective of this study is to investigate the validity of the following two hypotheses: 1) the short-term water demand process at the Indian Institute of Technology, Kanpur campus is a dynamic process mainly driven by the maximum air temperature and interrupted by rainfall occurrences, and 2) occurrence of rainfall is a more significant variable than the rainfall amount itself in modelling the short-term water demand forecasts. The data employed in this study consist of weekly water demand at the Indian Institute of Technology, Kanpur campus, and total weekly rainfall and weekly average maximum air temperature from the City of Kanpur, India. Six different artificial neural network models, five regression models, and two time series models have been developed and compared. The artificial neural network models consistently outperformed the regression and time series models developed in this study. An average absolute error in forecasting of 2.41% was achieved from the best artificial neural network model, which also showed the best correlation between the modelled and targeted water demands. It has been found that the water demand at the Indian Institute of Technology, Kanpur campus is better correlated with the rainfall occurrence rather than the amount of rainfall itself.     

Estimate of Suspended Sediment Concentration from Monitored Data of Turbidity and Water Level Using Artificial Neural Networks

Artificial neural networks (ANNs) are promising alternatives for the estimation of suspended sediment concentration (SSC), but they are dependent on the availability data. This study investigates the use of an ANN model for forecasting SSC using turbidity and water level. It is used an original method, idealized to investigate the minimum complexity of the ANN that does not present, in relation to more complex networks, loss of efficiency when applied to other samples, and to perform its training avoiding the overfitting even when data availability is insufficient to use the cross-validation technique. The use of a validation procedure by resampling, the control of overfitting through a previously researched condition of training completion, as well as training repetitions to provide robustness are important aspects of the method. Turbidity and water level data, related to 59 SSC values, collected between June 2013 and October 2015, were used. The development of the proposed ANN was preceded by the training of an ANN, without the use of the new resources, which clearly showed the overfitting occurrence when resources were not used to avoid it, with Nash-Sutcliffe efficiency (NS) equals to 0.995 in the training and NS = 0.788 in the verification. The proposed method generated efficient models (NS = 0.953 for verification), with well distributed errors and with great capacity of generalization for future applications. The final obtained model enabled the SSC calculation, from water level and turbidity data, even when few samples were available for the training and verification procedures.