Application of Artificial Neural Networks (ANNs) for the prediction of ...

11th International Conference on Urban Drainage, Edinburgh, Scotland, UK, 2008______

Application of Artificial Neural Networks (ANNs) for the prediction of CSO discharges Andreas Kurth1*, Prof. Adrian Saul1, Dr. Steve Mounce1, Dr. Will Shepherd1, David Hanson2 1Pennine Water Group, Department of Civil and Structural Engineering, University of Sheffield, Sir Frederick Mappin Building, Mappin Street, Sheffield, S1 3JD 2Yorkshire Water Systems Services, Bradford, BD3 7YD * Corresponding author, e-mail [email protected]

ABSTRACT Combined sewer overflows (CSOs) represent a common feature in combined urban drainage systems and are used to discharge excess water to the environment during heavy storms. Due to problems arising within the system, such as blockages, the release structures can unfortunately operate even in dry weather conditions spilling the polluted foul water directly to the watercourse. In order to prevent this from occurring, research regarding the prediction of the performance of CSO assets is being carried out in three UK catchments in collaboration with Yorkshire Water Systems Services. The overall strategy is to develop a monitoring, modelling and predictive operational strategy that utilizes rainfall input from weather radar to ultimately predict the hydraulic performance. Investigating an alternative approach to hydraulic models, an Artificial Neural Network (ANN) was utilised to predict the CSO performance. A three hiddenlayer feed-forward multilayer perceptron (MLP) was trained, validated and tested with data recorded from the system. Preliminary results from the first catchment suggest that the underlying relationship between local rainfall and water depth related to the weir crest within the CSO structure can be captured successfully in order to predict the normal CSO performance three time steps ahead.

KEYWORDS Combined sewer overflow (CSO); Artificial Neural Network Perceptron (MLP); Monitored data; Catchment; Rainfall Radar Data

(ANN);

Multilayer

INTRODUCTION Combined sewer overflow (CSO) structures are common assets within the UK’s combined urban drainage systems with the purpose to protect downstream sewers and Waste Water Treatment Plants (WWTP) from hydraulic overloads and flooding during extreme rainfall events by discharging the excess water directly to a receiving watercourse. They are designed to operate during heavier storm events only as the inherent pollutants of the foul water are then diluted and the impact on the environment is comparatively less. However, problem CSOs do exist and they may produce unnecessary and un-consented flooding/ spill incidents. If a CSO is operating during dry weather flow condition the undiluted, untreated

Andreas Kurth et al.

1

11th International Conference on Urban Drainage, Edinburgh, Scotland, UK, 2008______ foul water is diverted to a receiving river with the potential of causing severe harm to the stream and on the environment. The general public, regulators such as EA (Environment Agency), and water companies such as Yorkshire Water Systems Services have a great interest in preventing this from happening. Yorkshire Water Systems Services have therefore in recent years invested in methods of constantly monitoring their systems to provide reliable information in order to develop methodologies to improve the management of the sewer network. Ultimately it is envisaged to develop a practical tool that can be used in water management which utilizes phased alarm levels to create an automated alarming system to enable the operating water company to organise remedial action pro-actively in order to significantly reduce the amount of un-consented discharge events. In general, a sewer network is a very complex and sophisticated system and to understand its performance it is necessary to take into account the hydrological and hydraulic processes that occur within the system. It is necessary to model them mathematically such that the predictions of performance may be made for different inputs, especially for predicting spill incidents of CSOs. Although it is generally possible to predict CSO discharges with hydraulic modelling software, there is no known reliable system in the UK which is able to predict un-consented discharges efficiently in real time. Therefore an alternative approach is being investigated in the course of this research project by applying Artificial Neural Network (ANN) techniques. In recent years several studies have been carried out in this area. Fernando et al. (2006) applied a feed-forward, back-propagation ANN model to forecast the occurrences of wastewater overflows in a combined sewerage system in Australia. The data used included the traditional modelpredicted overflow rates for one overflow structure and the synthetically generated rainfall for the rain-gauge in the closest proximity. Choosing two different data sets, the first contained antecedent overflow rates and antecedent rainfall, whereas the second took antecedent rainfall data into account. It was shown that the availability of antecedent flow rate data at the CSO was essential to predict the overflow rates accurately, while the second model failed to generalise the problem. Sumer et al. (2007) has researched the feasibility of real-time detecting sanitary sewer overflows (SSOs) using time series analysis and ANN techniques in two case studies in Arizona, USA. In order to identify whether a SSO is occurring, control limit theory was used to detect important deviations between measured and expected depth and flow data.

THEORY General background on CSOs Background and design information about CSO structures can be found e.g. in the industry standard guideline FR048 (Balmforth et al., 1994), in the specific design guide for CSOs incorporating screens (WaPUG, 2001) and in terms of primary operational functions in Saul et al. (1997).

2

Application of ANNs for the prediction of CSO discharges

11th International Conference on Urban Drainage, Edinburgh, Scotland, UK, 2008______ General background on ANNs Generally, an Artificial Neural Network (ANN) is an information processing system that is inspired by the way biological nervous systems, such as the brain, process information (Figure 1). The key element of this system is the novel structure of the information processing system. It is composed of a large number of highly interconnected processing elements (neurons) working together to solve specific problems. ANNs, like human beings, learn from examples. An ANN is configured for a specific application, such as pattern recognition or data classification, through a learning process. Learning in biological systems involves adjustments to the synaptic connections that exist between the neurons which is true for ANNs as well (typically called weights).

Figure 1: Principles of a biological neuron (Mehrotra, K., 2000)

Figure 2: Principles of a basic artificial neuron

Neural networks rely on training data to initialize and update the system. Thus, an ANN requires an appropriate training set that allows the system to learn and generalize on future input data. A combination of inputs very similar to previously seen training data are recognized and result in a similar output, while new data (or incomplete and/or noisy data) can be matched as closely as possible to patterns previously learned by the system (Medsker et al., 1999). Regarding the history and development of ANNs the scientific contributions of McCulloch and Pitts (1943), Hebb (1949), Rosenblatt (1958, 1962) and Minsky and Papert (1969) are very significant. However, it was not until 1986 when Rumelhart, McClelland and Williams published their paper detailing a method for training a multi-layer perceptron (MLP) that the problems posed by Minsky and Papert (1969) were solved. MLPs trained with backpropagation are, in theory and given sufficient training data, universal computing machines capable of arbitrary function approximation. Generally error backpropagation is a form of supervised learning for multi-layer networks. Error data at the output layer is “backpropagated” to earlier layers, allowing incoming weights to these layers to be updated. It is most often used as training algorithm in current neural network applications. Subsequently the field of ANNs flourished and increasing research, development and application led to the founding of journals, conferences and companies. Scientists and engineers such as Hopfield (1985), Carpenter et al. (1987) and Kohonen (1984, 1988) developed other ANN architectures and applications that demonstrated the potential of ANN technology. Regarding the topic of this paper, a major property of ANNs is the fact that they do not need to have explicit formulation of the hydrological-hydraulic interrelation of the system in advance as


3

11th International Conference on Urban Drainage, Edinburgh, Scotland, UK, 2008______ this data driven approach will learn the underlying characteristic pattern simply from historical data. There are many examples of successful application of the ANN technique in the water industry, mostly applied to river flow/level and flood forecasting and to rainfall-runoff modelling (e. g. Aqil et al., 2006; Dawson et al., 2002; Dawson et al., 2006; Imrie et al., 2000; Thirumalaiah et al., 2000; Wu et al., 2005 and Zealand et al. 1999), but just a few ones are related to CSO performance (e. g. Fernando et al., 2006; Sumer et al., 2007). Project hypothesis The ANN developed in this project is designed to learn and generalise the underlying relationship between local rainfall and water level in a CSO chamber which will occur with a certain lag time. With a rainfall event occurring in the contributing catchment, the CSO reacts with a rising water level in its chamber, whereas under normal conditions during dry weather the release structure presents a relatively stable diurnal water level pattern. It is envisaged that the outcome of this project will be specifically applicable to those CSOs which are at a high risk of un-consented discharge.

DATA Case study catchments Yorkshire Water Systems Services as the responsible water company in Yorkshire are currently monitoring many of their CSO assets with telemetered Hawkeye data monitors that record water levels relative to the weir height. For a first case study, Masborough Street CSO in the Masborough & Kimberworth catchment was chosen. It is one of eleven drainage areas in Rotherham covering a total area of 7.5 km² and serving a population of 17,000. As part of the strategic partnership between Yorkshire Water Systems Services and the University of Sheffield, two strategic catchments were selected, namely Ilkley and Thornton. Within these catchments sewerage assets are being monitored for rainfall (raingauge and radar data), flow depth, flow magnitude and quality over a proposed 10 year period since 2007 and 2008 respectively. The data gathered from these catchments will be used to meet the research needs of Yorkshire Water Systems Services and the University of Sheffield. When using ANNs it is crucial to apply representative time series data covering the most important pattern of the system to be investigated. From Masborough Street CSO, sufficient historic data was available so that it could be confidently used as first case study. In the future it is proposed to utilize the data from the two data rich strategic catchments, Ilkley and Thornton, as the second and third case study catchment to apply and prove the developed methodology. Working very closely with Yorkshire Water Systems Services, all the measured data used for this study are available from their databases.

4


11th International Conference on Urban Drainage, Edinburgh, Scotland, UK, 2008______ Data sources for Masborough Street CSO Worsbrough Dale STW Rain gauge Mexborough STW Rain gauge

Masborough&Kimberworth Catchment Blackburn Meadows WwTW – Rain gauge

Masborough Street CSO ~3 km

Figure 3:

First case study catchment Masborough and Kimberworth with three rain gauges (Source: Yorkshire Water Systems Services, GIS system “Odyssey”)

The CSO used in this study is constructed as a single high-side weir chamber being known to spill a few times per year. Apart from water level data within Masborough Street CSO (with 100% signifying the spill level), rain intensity data (mm/hr) from three permanent tipping bucket rain gauges located in the further proximity of the CSO were utilised (see Figure 3). The time series data used ranges from 01.07.2005-31.03.2007 with a time resolution of 5 minutes (i.e. one reading every 5 minutes).

0

180

5

160

10

140

15

120

20

111.15

100

25

80

30

60

35

40

40

20

45

0

50

Rain intensity (mm/hr)

200

15:00 15:05 15:10 15:15 15:20 15:25 15:30 15:35 15:40 15:45 15:50 15:55 16:00 16:05 16:10 16:15 16:20 16:25 16:30 16:35 16:40 16:45 16:50 16:55 17:00 17:05 17:10 17:15 17:20 17:25 17:30 17:35 17:40 17:45 17:50 17:55 18:00 18:05 18:10 18:15 18:20 18:25 18:30 18:35 18:40 18:45 18:50 18:55 19:00 19:05 19:10 19:15 19:20 19:25 19:30 19:35 19:40 19:45 19:50 19:55 20:00 20:05 20:10 20:15 20:20 20:25 20:30 20:35 20:40 20:45 20:50 20:55 21:00

Water level in CSO (% of weir height)

Relationship Masborough CSO - Blackburn, Mexborough and Worsborough Raingauges (28/09/05)

Time

Blackburn Meadows RG

Figure 4:

Mexborough STW RG

Worsborough STW RG

Masborough Street CSO

Weir crest/Spill level

Relationship between CSO performance (water level) and local rainfall in 09/05

The runoff of a storm contributing to a catchment eventually enters the sewer system to be carried downstream. Hence, there is a delay until the runoff has got an impact on the flow rate in the sewer system and the depth within the CSO respectively. Figures 4 and 5 provide representative examples of the relationship between local rainfall and the resultant change in flow depth. As expected, there is a lag time between the peak of rainfall and peak of flow. It can be seen that the two raingauges Blackburn Meadows and Mexborough STW (Sewage Treatment Works) give similar results and better reflect the CSO performance. Data from the Worsborough STW raingauge was subsequently not used in the study.


5


Relationship Masborough CSO - Blackburn, Mexborough and Worsborough Raingauges (06/10/06)

10

160 (% of weir height)

Water level in CSO

180

140

20

120 109.4

100

30

80 40

60 40

Rainfall intensity in mm/hr

0

200

50

20 60 06:00 06:05 06:10 06:15 06:20 06:25 06:30 06:35 06:40 06:45 06:50 06:55 07:00 07:05 07:10 07:15 07:20 07:25 07:30 07:35 07:40 07:45 07:50 07:55 08:00 08:05 08:10 08:15 08:20 08:25 08:30 08:35 08:40 08:45 08:50 08:55 09:00 09:05 09:10 09:15 09:20 09:25 09:30 09:35 09:40 09:45 09:50 09:55 10:00 10:05 10:10 10:15 10:20 10:25 10:30 10:35 10:40 10:45 10:50 10:55 11:00 11:05 11:10 11:15 11:20 11:25 11:30 11:35 11:40 11:45 11:50 11:55 12:00

0

Time

Blackburn Meadows RG

Figure 5:

Mexborough STW RG

Worsborough STW RG

Masborough Street CSO

Weir crest/Spill level

Relationship between CSO performance (water level) and local rainfall in 10/06

Although an obvious relationship can be observed, the interrelation between CSO performance and storm occurrence is non-linear which makes it suitable for a neural network application. Hence, local rainfall and water level data within the CSO chamber are used as the input data to the ANN in order to find the underlying relationship. CSOs in Ilkley and Thornton catchments Additional to the data available in Masborough & Kimberworth, the two strategic catchments are equipped with several flow and depth monitors installed in the pipes close to CSOs under investigation. Furthermore, several rain gauges have been installed to capture the storms contributing to the catchments. Both catchments are also covered by Rainfall Radar data with a resolution of 1 km2.

METHODS The software chosen to build and run the ANN model is NeuroSolutions (Version 5.06; NeuroDimension, Inc.); data pre-processing was carried out using Matlab and Excel. The raw data was pruned of outliers and single missing values in the depth data were filled by linear interpolation. Normalisation was applied to the input data using a range from -1 to 1. Generally the ability to forecast beyond the range of data the ANN was trained and validated with is limited with this approach. However, the data being used for training and cross-validation was selected in such a way as to ensure that all important patterns were covered to overcome this limitation. A total of 60 % of the data was used for training, a further 20 % each for crossvalidation (CV) and testing. Determination of input data and architecture As suggested by Fernando et al. (2006), cross-correlation was applied to establish the most suitable amount of time steps into the past (t-i) to be taken into account regarding rainfall data. As can be seen in Figure 6, Blackburn and Mexborough raingauges present the best correlation with the CSO water level from what was to be expected from Figures 4 and 5. The information delivered by Worsborough raingauge is less relevant for this particular CSO and will not be taken into account for the modelling approach.

6


11th International Conference on Urban Drainage, Edinburgh, Scotland, UK, 2008______ Starting with t=0 the increase of the correlation coefficient is proportional to the lag time and starts decreasing at a lag time of t= -5. Correlation between CSO and different raingauges 0.5

Correlation coefficient

0.45 0.4 0.35

Masborough Street CSOBlackburn RG

0.3

Masborough Street CSOMexborough RG

0.25 0.2

Masborough Street CSOWorsborough

0.15 0.1 0.05 t=0 t=-1 t=-2 t=-3 t=-4 t=-5 t=-6 t=-7 t=-8 t=-9 t=-10 t=-11 t=-12 t=-13 t=-14 t=-15 t=-16 t=-17 t=-18

0

Time lag

Figure 6:

Cross correlation between water depth in CSO and rainfall

In order to predict water levels 3 time steps into the future (DM(t=+3)), with DM being water depth data within Masborough Street CSO chamber, the following input variables have been chosen. The rainfall considered to be relevant to capture the inherent problem are RB(t=0)-RB(t=-12) with RB being raingauge data of Blackburn Meadows RG (rain gauge) and RM(t=0)-RM(t=-12) with RM being raingauge data of Mexborough STW RG. On the basis of conducted serial correlation (Fernanda et al., 2006), water depth input values taken into account are DM(t=0)-DM(t=-5). The right choice of input pattern is as important as the determination of an architecture capable of learning and generalizing the inherent input-output function of a particular problem. Numerous trials were carried out and for the data available for Masborough Street CSO a 3 hidden-layer feed-forward MLP trained with back-propagation gave a good representation of measured system performance. It comprises of 32 PEs (Processing Elements) or layer nodes in the input layer and 1 on the output layer, whereas the 3 hidden layer (HL) comprise 50 PEs (HL1), 30 PEs (HL2) and 15 PEs (HL3). The ANN was trained and cross-validated several times with the best weights stored to be used for testing/forecasting. The stopping criteria (MSE termination) was based on the cross validation data set as it tends to be an appropriate indication of generalization the ANN has achieved. This approach stops the network when the MSE of the CV data set begins to increase indicating that the network starts overtraining. Overtraining is when the network purely memorizes the training data, hence, becomes specialised to it. Because it would fail to generalize the underlying problem, overtraining has to be prevented.


7

11th International Conference on Urban Drainage, Edinburgh, Scotland, UK, 2008______ Once trained and cross-validated with 80 % of the data set, the frozen best weights are used to test the network on the remaining 20 % unseen data.

RESULTS AND DISCUSSION The result shows that the chosen 3 hidden-layer Multilayer Perceptron (MLP) trained with backpropagation is capable of learning the underlying relationship between local rainfall occurrence and CSO response. Fig. 7 reveals the comparison of the 3 time steps ahead predicted water level values (model output, pink line) with the desired actual data of the unseen test set (blue line). Comparison of actual with model output (3 time steps ahead) 120

Water level in CSO (% of weir height)

100 80 60 40 20 0 1

1125 2249 3373 4497 5621 6745 7869 8993 10117 11241 12365 13489 14613 15737 16861 17985 19109 20233 21357 22481 23605 24729 25853 26977 28101 29225 30349 31473

Data points (5 min intervals)

Masborough t=3

Figure 7:

Masborough t=3 Model Output

Results of modeled versus measured CSO water level using a 3-HL MLP

Initially it was assumed that the majority of data collected during dry weather condition had an impact on the network’s ability to pick up the peak flows during storm events. However, with the methodology chosen, the network was able to generalize the inherent input-output function. The model performance criteria of the test run with a MSE (Mean Squared Error) of 7.18 using the best weights achieved during training proves the success. As can be seen in Fig. 7 the pattern is nicely matched, too. Further important performance measures are NMSE (Normalised Mean Squared Error) and MAE (Mean Absolute Error) which are listed in Figure 8. Model performance criteria

Masborough Street CSO (t=+3)

MSE NMSE MAE

7.18 0.12 1.11

Figure 8:

Model performance

Although the network is able to match most of the storms correctly and in this case the one spill incident has been forecasted, too, generally major rainfall events have proven to be more difficult to be modelled precisely regarding peak values (Figure 7). With a MAE of 1.11 it is indicated that a good estimate 15 minutes (3 time steps) into the future on unseen data has been accomplished. However, this is mainly influenced by the very good MAE achieved for dry weather flow periods. The MAE for the storm periods only would be much higher. Further

8


11th International Conference on Urban Drainage, Edinburgh, Scotland, UK, 2008______ research will be carried out using additional data available in the two strategic catchments mentioned earlier.

CONCLUSIONS AND FUTURE WORK The work carried out has shown the potential of ANNs being able to capture the underlying relationship between contributing local rainfall and consequences on the water level within a CSO structure downstream. The identification of normal CSO behaviour depending on rainfall is envisaged to contribute to a straightforward methodology. It seeks to become able to distinguish between a rise in water levels caused by rainfall events and rising patterns not significantly related to storm events and therefore indicating anomalies. Further research will be conducted on this matter. Being able to differentiate between the two types of triggers of rising patterns, it should be feasible as a next step, to create an automated alarm system, that could be implemented in companies dealing with waste water management. It is proposed that Yorkshire Water Systems Services complete a comparison of ANN predicted pattern with real time measured water level. The outcome will be written into a decision support tool based on phased alarm levels depending on the degree of deviation. Further research will be carried out in two further catchments using additional monitored data, namely flow, velocity and depth in the upstream sewer and rainfall radar data, in order to model the processes more precisely and to take the research to the next level of developing a decision support tool regarding whether action needs to be taken to prevent un-consented discharges. Due to the fact that a prediction of 15 minutes (3 time steps) into the future as presented in this paper does not signify enough time to act pro-actively, it is planned to create a link with Rainfall Radar Data. If the development and movement of a storm is known even hours before it hits a catchment, this information will be used together with the CSO performance prediction model, to capture the entire process upon which a decision support tool can be based. Making maintenance routines in sewer network related waste water management more efficient, it will be possible to make a contribution to a safe, non-hazardous and clean environment.

ACKNOWLEDGEMENT The support from Yorkshire Water Systems Services regarding the provision of data is gratefully acknowledged.

REFERENCES Aqil, M.; Kita, I.; Yano, A.; Soichi, N. (2006): ‘Decision Support System for Flood Crisis Management using Artificial Neural Networks, International Journal of Intelligent Technology, Vol. 1, No. 1, ISSN 13056417, pp. 70-76 Balmforth, D.J..; Saul, A. J.; Clifforde, I. T. (1994): ‘Guide to the Design of Combined Sewer Overflow Structures’, FWR Report No FR0488 Carpenter, G.; Grossberg, S. (1987): ‘A Massively Parallel Architecture for a Self-organizing Neural Pattern Recognition Machine’, Computer Vision, Graphics and Image Processing, Vol. 37, pp. 54-115


9


Dawson, C. W.; Abrahart, R. J.; Shamseldin, A. Y.; Wilby, R. L. (2006): ‘Flood estimation at ungauged sites using artificial neural networks’, Journal of Hydrology, Vol. 319, pp. 391-409 Dawson, C. W.; Harpham, C.; Wilby, R. L.; Chen, Y. (2002): ‘Evaluation of artificial neural network techniques for flow forecasting in the River Yangtze, China’, Hydrology and Earth System Sciences, 6(4), pp. 619626 Hebb, D. O. (1949): ‘The Organization of Behaviour’, New York: Wiley Hopfield, J. (1985): ‘Neural Networks and Physical Systems with Emergent Collective Computational Abilities’, Proceedings of the National Academy of Sciences of the United States of America, Vol. 79 Imrie, C. E.; Durucan, S.; Korre, A. (2000): ‘River flow prediction using artificial neural networks: generalisation beyond the calibration range’, Journal of Hydrology, Vol. 233, pp. 138-153 Kohonen, T. (1984): ‘Self Organization and Associative Memory’, Berlin: Springer-Verlag Kohonen, T. (1988): ‘An Introduction to Neural Computing’, Neural Networks, Vol. 1, No. 1 McCulloch and Pitts (1943): ‘A logical calculus of the ideas immanent in nervous activity’, Bulletin of Mathematical Biology 5, pp.115-133 Medsker, L. and Liebowith, J. (1999): ‘Design and development of expert systems and neural networks’, ISBN 0-02-380131-X Mehrotra, K.; Kohan, C. K.; Ranka, S. (2000): ‘Elements of Artificial Neural Networks’, MIT Press, Cambridge, Massachusetts, 2nd printing, ISBN 0-262-13328-8 Minsky, M. and Papert, S. A. (1969):’ Perceptron’, MIT Press, Cambridge, MA Rosenblat, R. (1958): ‘The perceptron, a probabilistic model for information storage and organization in the brain’, Psychol. Rev., Vol. 65, pp. 386-408 Rosenblatt, R. (1962): ‘Principles of Neurodynamics: Perceptrons and the Theory of Brain Mechanisms’, Spartan, New York Rumelhard, D. E.; Hinton, G. E. and Williams, R. J. (1986): ‘Learning internal representations by error propagation’, in Parallel Distributed Processing 1 (8), edited by D. E. Rumelhard and J. L. Mc Clelland, MIT Press, Cambridge, Massachusetts Saul, A. J., et al. (1997): ‘Sewers – Rehabilitation and new construction, repair and renovation’, edited by Geoffrey F. Read with I. Vickridge, published by Arnold, London, ISBN 0340544724 Sumer, D., Gonzalez, J., Lansey, K. (2007): ‘Real-Time Detection of Sanitary Sewer Overflows Using Neural Networks and Time Series Analysis’, Journal of Environmental Engineering, Vol. 133, No. 4, pp. 353-363 Swingler, Kevin (1996): ‘Applying Neural Networks – A Practical Guide’, ISBN 0-12-679170-8 Taylor, J.G. (1993): ‘The Promise of Neural Networks’, ISBN 3-540-1977-7 Thirumalaiah, K.; Deo, M. C. (1998): ‘River Stage Forecasting Using Artificial Neural Networks’, Journal of Hydrologic Engineering, Vol. 3, No. 1, ISSN 1084-0699, pp. 26-32 WaPUG Guide (2001): ‘The Design of CSO Chambers to Incorporate Screens’, 18 pages, June 2001, http://www.wapug.org.uk/ Wu, J. S.; Han, J.; Annambhotla, S.; Bryant, S. (2005): ‘Artificial Neural Networks for Forecasting Watershed Runoff and Stream Flow’, Journal of Hydrologic Engineering, Vol. 10, No. 3, pp. 216-222 Zealand, C. M.; Burn, D. H.; Simonovic, S. P. (1999): ‘Short term streamflow forecasting using artificial neural networks’, Journal of Hydrology, Vol. 214, pp. 32-48

10
