Journal Search Engine
Search Advanced Search Adode Reader(link)
Download PDF Export Citaion korean bibliography PMC previewer
ISSN : 2288-1115(Print)
ISSN : 2288-1123(Online)
Korean Journal of Ecology and Environment Vol.46 No.1 pp.1-9
DOI : https://doi.org/10.11614/KSL.2013.46.1.001

Using Artificial Neural Networks for Forecasting Algae Counts in a Surface Water System

Emery A. Coppola Jr.*, Adorable B. Jacinto, Tom Atherholt1, Mary Poulton2, Linda Pasquarello3, Ferenc Szidarvoszky4, Scott Lohbauer
New Jersey Department of Environmental Protection
(NOAH, L.L.C., 610 Lawrence Road, Lawrenceville, New Jersey, 08648-4208, U.S.A.
1Division of Science, Research, and Technology.
New Jersey Department of Environmental Protection, 401 East State Street,
Trenton, New Jersey, 08625-0409, U.S.A.
2Department of Mining and Geological Engineering, University of Arizona,
Tucson, Arizona, 85721-0012, U.S.A.
3Passaic Water Valley Commission, 1525 Main Avenue, Clifton, NJ, 07011, U.S.A.
4Department of Systems and Industrial Engineering, University of Arizona,
Tucson, Arizona 85721-0020, U.S.A.)
Manuscript received 4 July 2012, Revised 18 November 2012, Revision accepted 12 February 2013

Abstract

Algal blooms in potable water supplies are becoming an increasingly prevalent andserious water quality problem around the world. In addition to precipitating tasteand odor problems, blooms damage the environment, and some classes like cyanobacteria(blue-green algae) release toxins that can threaten human health, evencausing death. There is a recognized need in the water industry for models that canaccurately forecast in real-time algal bloom events for planning and mitigation purposes.In this study, using data for an interconnected system of rivers and reservoirsoperated by a New Jersey water utility, various ANN models, including both discreteprediction and classification models, were developed and tested for forecasting countsof three different algal classes for one-week and two-weeks ahead periods. Predictormodel inputs included physical, meteorological, chemical, and biological variables,and two different temporal schemes for processing inputs relative to the predictionevent were used. Despite relatively limited historical data, the discrete predictionANN models generally performed well during validation, achieving relatively highcorrelation coefficients, and often predicting the formation and dissipation of highalgae count periods. The ANN classification models also performed well, with averageclassification percentages averaging 94 percent accuracy. Despite relatively limiteddata events, this study demonstrates that with adequate data collection, both in termsof the number of historical events and availability of important predictor variables,ANNs can provide accurate real-time forecasts of algal population counts, as well asfoster increased understanding of important cause and effect relationships, whichcan be used to both improve monitoring programs and forecasting efforts.

01-Using(12-34)(JH)컬러.155452.pdf301.4KB

INTRODUCTION

 Although there is general consensus among scientists that the incidence of algal blooms (AB) world-wide is increasing at an alarming rate (Smith et al., 2006), and the detrimental effects of these blooms on the environment and water supplies are well documented, controversy remains over the important factors and mechanisms responsible for their occurrence, and the most effective means for both modeling and forecasting this phenomena. Given the multitude, interplay, and complexity of various weather, water quality, biological, and hydrologic factors, many of which will vary over space and/or time, complicated by random events, some researchers argue that there are no fast and true “rules” for predicting algal biomass and blooms. Researchers recognize that reliance upon mechanistic models for predicting algal biomass is insufficient, given our inadequate “level of understanding of how these complicated ecological systems work” (Pelley, 2005).

Researchers have identified fundamental “nutrient” variables, primarily nitrogen and phosphorous, as “limiting” to the growth of algae. Because nitrogen and phosphorous are often strongly correlated with the quantity of algal biomass in water systems, researchers frequently develop models that predict algal biomass as a function of one or more of these compounds. Limiting the models to such highly reduced input-output relationships, while efficient and sometimes effective, not only ignores the complexity of processes that determine nutrient levels, but also overlooks the myriad of other factors that can influence algal biomass. For example, nitrogen may originate from a number of different sources, and its chemical form and concentration is dictated by different nitrogen processes, such as nitrification, which depends upon the presence of certain bacterial organisms. In some water systems, nutrients that might otherwise be considered limiting on the resident algal populations appear to persist within a range of concentrations that does not significantly affect the organisms. 

 Further complicating the dynamics of algal populations are the variety of physical and biological factors that influence the formation and dissipation of algal blooms. Sunlight is essential for the development of these photosynthetic organisms, and the amount of light that penetrates the water column is controlled by a number of factors, each of which may have multiple effects upon the system. For example, precipitation not only reflects a lower sunlight factor, but also influences the amount of turbidity in the water column via sediment transfer from rainfall run-off. The degree of run-off depends not only upon the quantity of precipitation, but also the size and characteristics of the watershed. Precipitation also increases surface water flow velocities which can stir up and suspend sediments from the bottom and scour sediments from the banks. Other factors that influence algal levels, like dissolved oxygen are similarly affected by other conditions of the system, such as water temperature, wind speed and direction, and the presence of other competing or even predatory organisms.

 In a publication by the American Water Works Association Research Foundation entitled “Early Warning Management of Surface Water Taste and Odor Events” (Taylor et al., 2006), the authors evaluate six existing mechanistic-based computer simulation models used by the water industry “to predict the timing, magnitude, and duration of taste and odor events” associated with algal populations. The authors discount four of the models on the basis that they are not capable of simulating detailed hydrodynamic conditions, such as stratification and mixing, which influence algal populations. The authors assert: “A good program must be able to model many water quality variables so that it can simulate a wide range of reservoirs.” They note that only the two most advanced models can simulate three algae classes. However, they caution: “it is a very involved process that requires extensive data collection and model calibration and validation. The large front-end effort precludes these models from being used within time frames required for managing specific taste and odor events.” The authors carefully draw a distinction between simulation of algae growth, which two of the mechanistic models can perform “reasonably well”, and the associated taste and odor problems caused by gesomin production and release from “blue-green algae concentrations.”

As an alternative paradigm to mechanistic models often used for forecasting algal blooms, artificial neural networks (ANNs) were used in this study. In this study, various ANN algal forecasting models were developed for three surface water sampling stations monitored as part of a drinking water supply system of interconnected rivers and reservoirs located in northeastern New Jersey. Using weather, water quality, hydrologic, and biological input variables, counts for three different algal classes were forecasted one-week and two-weeks ahead. The algal classes considered in this study included the toxic cyanobacteria, better known as blue-green algae, chrysophyta (gold algae), and chlorophyta (green algae) classes.

 There is previous work in the scientific literature where ANNs were developed and tested for predicting algal blooms. Rechnagel and others (1997) applies the technology to four different freshwater systems. The paper first introduces the complexity and non-linearity of algal bloom dynamics and used at least six and up to ten years of what appear to be weekly data consisting of limiting nutrients, water temperature, light conditions, and, in one case, density data of zooplankton groups to train the ANN models to predict phytoplankton organisms. Maier and others (1998) forecasted cyanobacteria in the River Murray using seven years of weekly data, consisting of eight input variables, was used to provide forecast algae counts four weeks into the future. As in the work presented in this paper, the authors concluded that adequate nutrients were available for algal growth, and hence were not limiting factors for the river system studied. Lee and others (2003) developed ANNs to predict algae bloom dynamics of a coastal water system using water quality data that could be used in real-time. Kim and others (2012) used ANNs to forecast algae in a surface water reservoir using hydrologic, weather, and water quality data collected from an automatic data collection system.

 The three primary study objectives in this project were: 1) to assess the feasibility of using artificial neural networks (ANNs) as a real-time tool for accurately forecasting algae counts of three species, in surface water systems; 2) identify critical climate, hydrologic, and water quality factors (i.e. variables) that may influence algae levels; 3) related to points 1 and 2 above; 3) assess the natural time evolution of algae populations, which is particularly relevant to real-time prediction capability.

 This research differs from previous work in that two different ANN paradigms were developed and tested; multiperceptron nets for predicting final algal counts as a single (i.e. discrete) numerical output value, and radial basis function nets for predicting the bin or classification range of values(pre-specified) within which final algal counts would fall. In addition, two different time schemes were used for assigning input variable values with their corresponding final algal counts; for the first case, model input values corresponded with measurements taken at the beginning of the prediction period, which would be most applicable for real-time forecasting, and for the second case, input value measurements generally corresponded with the conclusion of the prediction period. Lastly, two different data sets were used; the first consisting of more input variables, but fewer historical data events, and the second excluded five select water quality variables that were measured less frequently, but consisting of more historical data events.

ARTIFICIAL NEURAL NETWORKS

 ANN architecture (Fig. 1) is based upon Kolmogorov’s theorem (Sprecher, 1965; Hecht-Nielsen, 1987) which asserts that any continuous function (in this case algal counts) can be represented exactly by a three-layer, feed-forward neural network with n elements in the input layer, 2n+1 elements in the hidden layer, and m elements in the output layer, where n and m are arbitrary positive integers. The presence of common arcs in its architecture allows ANN to identify important inter-relationships that may exist between output variables. ANN technology is a compelling alternative to physical-based modeling approaches (Poulton, 2001). ANN “learns” system behavior by processing representative data through its archi-tecture. ANN is different from physical-based models because it does not rely upon the governing physical laws for making its predictions, and consequently, traditional model parameters are often not required for ANN development and operation.

Fig. 1. Architecture for a simple multi-perceptron ANN.

 In this study, 50% of the available data was used for “training,” that is to “learn” cause and effect relationships if present. Another 25% of the data was used to “verify” the model, to guard against over-training or over-fitting the data. Following training, the remaining 25% of the data was used to validate or assess how well the model learned to generalize system behavior. During training, data patterns are processed through the ANN and “connection weights” are adaptively adjusted until a minimum acceptable error between the ANN-predicted output and the actual output is achieved. At this point, the ANN has “learned” to predict the system behavior of interest (in this case algal counts or classes) in response to the various input parameter values.

 There are a variety of ANN model design features and options. To design an appropriate model, a variety of factors must be considered, including the functional form of the transfer functions, the number of hidden layers and nodes in the architecture, the most appropriate set of input variables, and the algorithm(s) used to minimize the objective function (i.e. training error). This process is typically conducted in an iterative manner within the context of professional judgment and modeling experience. For example, selection of an appropriate set of input parameters during initial ANN development requires a basic understanding of the governing system dynamics (e.g., factors known to influence AB). However, a “sensitivity analysis,” in conjunction with trial and error, can help the modeler converge on the most appropriate and feasible set of predictor variables. The sensitivity analysis, which quantifies the relative importance of each input variable for accurately predicting each output variable, can be used in lieu of common statistical methods.

 ANNs require sufficient data that spans the range of expected system conditions to allow robust learning. Based upon the number of input and output parameters, heuristic equations were used to estimate the minimum number of training data sets required for robust model development. Calculated estimates of the number of training events (data sets) necessary for robust training in this study ranged from 200 to 500, depending upon the ANN model used. Because of the number and complexity of environmental factors and their interactions which control AB dynamics, and given the expected “noise” in the data, the number of required training data sets is probably closer to 500. In this study, the number of data sets available in this study was well below 200.

STUDY AREA AND DATA

 Fig. 2 represents the system modeled in this project, where for water security reasons, specifics are omitted. Two rivers and a reservoir supply water to the water treatment plant (WTP). River A flows into River B upstream of the WTP’s intake canal, while River B water is gravity fed to the WTP intake by way of the canal. Rivers A and B have historically exhibited variable and unique water quality characteristics that impart different treatment challenges. River B is considered to be of lower water quality because of more numerous upstream contaminant sources. However, River A has a higher incidence of AB events.

Fig. 2. Raw water configuration and station locations.

 Because of their desire to forecast algae blooms and monitor overall water quality, the utility has an extensive watershed water quality monitoring program in place to assist with decision making for source water selection and prediction of water quality changes. Grab and online sample data are supplemented by United States Geological Survey flow and water quality monitoring stations located throughout the watershed. The existing algal monitoring program consists of analyzing key water quality parameters and correlating changes in concentrations to predictions of algal concentrations.

 Previously collected (1999~2004) water quality data from Stations 100, 101, and 612 were provided by the utility and used for model development and assessment. A total of 302 measurement events consisting of water quality, hydrologic, weather, pumping, and extraction data collected over the period January 1999 to August 2004 were used in the study. Climate data was obtained from National Oceanographic and Atmospheric Administration (NOAA) and included total daily precipitation, average daily temperature, wind speed, and wind direction. There was no available data for solar radiation; hence data for sky cover, heating degree days, and length of day were also used, with values for the two first variables also obtained from NOAA, and the last obtained from sunrise and sunset tables obtained on-line.

 A listing of the model variables input values, as well as their minimum, average, and maximum values by station, are presented in Table 1.

Table 1. Statistical tabulation for all model variables used by Station.

 System conditions for algal populations, water quality, physical, and weather conditions vary by season. Representative time-series figures depicting algal counts for the three classes are plotted versus dissolved oxygen, nitrate, and water temperature in Figs. 3 through 5, respectively, providing an overview of the complexity and a general lack of transparent consistency in system conditions and algal counts.

Fig. 3. Total algae counts versus dissolved oxygen concentration measured at Station 101.

MODELING METHODOLOGY

 Because reliable weather forecasts generally do not extend beyond one- to two-week time periods, ANN models were developed for one-week and two-week ahead forecasting periods. Two different time schemes were used for computing the values of the ANN model inputs variables. The first, referred to as “original”, consisted of input values measured at the beginning of the prediction periods. The second, referred to as “revised”, used input values measured at the end of the prediction period, coinciding with the final or predicted algal count.

 Both the original and revised approaches were assessed using two distinct data sets. The first set consisted of a smaller number of time-coincident events, but which included a higher number of input variables. The second set, by excluding several less-frequently-sampled water quality parameters (total phosphorous/ortho-phosphate, nitrite/nitrate, sulfate, and total organic carbon for all stations and biological oxygen demand for Stations 101 and 612), consisted of a larger number of data events but with fewer input variables.

 In order to help identify potentially important predictor variables, time series were developed depicting different variables versus algae counts. For example, Figs. 3, 4, and 5 depict oxygen concentrations, temperature, and nitrite/nitrate versus algae counts, respectively. As can be seen, there are not obvious relationships between potential predictor and prediction (i.e., algae count) variables. In addition, a sensitivity analysis was performed by computing a sensitivity ratio, mathematically defined as the root mean squared error of ANN predictions without a particular input variable divided by the mean squared error of ANN predictions with the input variable. For example, a sensitivity ratio of 2.0 indicates that removing the particular input variable increases the RMSE by a factor of 2, while a value of 1.06 indicates that inclusion of the predictor variable has minimal effect on improving prediction accuracy.

Fig. 4. Total algae counts versus temperature measured at Station 612.

Fig. 5. Total algae counts versus Nitrite/Nitrate concentration measured at Station 100.

 As an alternative to developing ANN models that explicitly predict final measured algal counts, RBF nets were developed to predict the pre-specified bins or classification ranges within which the final measured algal counts fall. For this modeling exercise, the following four bins or classification ranges were selected: 0 to 10 counts, 11 to 50 counts, 51 to 200 counts, and 201 and above counts. The original model input sets were used and included both the complete and reduced parameter sets. Station 101 was used for predicting chlorophytes bins one-week ahead and chrysophytes bins two-weeks ahead. Station 612 was selected for predicting chrysophytes bins one-week ahead and chlorophytes bins two-weeks ahead.

 The ANN models consisted of multiple input variables to predict a single output variable, consisting of the algae counts for the multi-perceptron ANN, and the bin or classification range for the radial basis nets. Various different combinations of input variables listed in Table 1 were used to forecast the final algae counts at the end of the forecast period (i.e., one week ahead or two weeks ahead), and using sensitivity analyses, input variables were reduced. The number of historical data events available for the different station and algae ranged from just 32 to 270. The professional software Statistica was used to perform the ANN modeling work presented in this study.

RESULTS

 The models developed with both one-week and two-week ahead prediction periods accurately predicted formation and dissipation of AB events, as well as the relative increase and decrease in cell counts. On the basis of validation correlation coefficients, the ANN models that used inputs measured at the beginning of the prediction period slightly outperformed the models that used inputs measured at the conclusion of the prediction periods, but the difference in validation performance was not significant (r=0.72 vs. 0.69). The importance of this result, however, indicates that real-time predictive accuracy can be achieved.

 The models that forecasted discrete algal count values achieved the highest performance in most cases when the less-frequently measured water quality parameters were excluded as input variables (r of 0.77 versus 0.63). In this case, r represents the correlation coefficient. Correlation coefficients range between -1 and +1, with values close to +1 indicating a strong positive correlation between predicted and measured values, and values closer to 0 indicating little correlation.

 Figs. 6 through 8 provide a visual assessment of model performance for three representative cases, where the validation data show the initial algal count corresponding to the prediction event. The validation series labeled “initial” in the figures designates the initial count measured at the beginning of the prediction period, “final” designates the final algal count measured at the conclusion of the prediction period (i.e., that which is being predicted), and “ANN” designates the final count predicted by the ANN model.

Fig. 6. Time-series plots of measured Chlorophyte counts and ANN one-week ahead predicted values for (a) complete and (b) validations data sets at Station 101 (Revised Model excluding five water quality inputs).

Fig. 7. Time-series plots of measured Chlrophyte counts and ANN one-week ahead predicted values for (a) complete and (b) validations data set at Station 101 (Original Model excluding five water quality inputs).

Fig. 8. Time-series plots of measured Cyanobacteria counts and ANN two-week ahead predicted values for (a) complete and (b) validations data set at Station 612 (Original Model excluding five water quality inputs).

 The models that predicted algal concentration ranges (i.e., classification nets or “bins”) rather than actual counts also achieved high forecasting performance. Three of the eight models that included all of the input parameters achieved 100 percent classification accuracy. The worst-performing net correctly classified 83 percent of the events. For this approach, the models that included the less-frequently-sampled inputs (phosphate, nitrate, etc.) slightly outperformed those that did not, with correct classification percentages of 96 and 92 percent, respectively. However, the models that excluded these parameters had approximately three times the number of available data sets and hence had more events which bordered two adjacent classification bins. All incorrect classifications for all models occurred within an adjacent bin (e.g., a measured count of 8 which placed the count in the 0~10 bin, while the predicted bin was 11~50). Given the inherent imprecision of algal counts (Maier et al., 1998) this performance is more than acceptable for informed water utility decisions.

CONCLUSIONS

 The results of this study demonstrate that ANN technology can be used to accurately forecast algal population counts in real-time for periods ranging from one-week to two-weeks ahead using readily available water quality, hydrologic, weather, and water extraction data. The major findings of this research include:

· Despite a very limited number of available data sets, the ANN models performed well in most cases during validation, accurately predicting large changes in algal cell populations. The degree of accuracy was surprising, given the complexity and non-linear behavior of algal populations, inherent data “noise”, and the relatively small number of historical events available for model training.

·  The ANN models that forecasted algal count values (instead of classification ranges) achieved the highest performance when the less-frequently measured water quality variables (phosphate, nitrate, sulfate, TOC and BOD) were excluded as input variables. This may be due to a data quantity issue rather than inherent importance of these parameters to algal cell growth, but it could also be that, at the concentrations at this WTP, these parameters were not “limiting” algal growth.

·  The Radial Basis Function classification net models classified the counts into the correct concentration ranges with very high accuracy, averaging 94 percent.

·  The ANN models developed with inputs measured at the beginning of the one-week and twoweek ahead prediction periods accurately predicted formation and dissipation of algal bloom events, as well as relative increase and decreases, indicating that there are natural time lags between system conditions and algal population responses. That is, algal populations may on average evolve predictably in response to system conditions, and the trajectory of algal counts over one and two-week forecast periods can be accurately forecasted on the basis of real-time measurements. This may also reflect that open water conditions as influenced by external factors like weather do not typically change significantly in the short-term (e.g. weekly or even bi-weekly), and thus evolving algal populations are not prone to abrupt deviations from trajectory paths. The relatively small changes in conditions over prediction periods is supported by the statistical analyses of the data.

· The small number of historical data events limits the accuracy of the sensitivity analyses performed by measuring the relative increase in RMSE by excluding each input variable. However, some basic trends did emerge, with the most important possibly being the relative nonimportance of the select water quality variables excluded from some models. In particular, the two “limiting nutrients”, total phosphorous/ orthophosphate and nitrite/nitrate, generally did not rank high as important predictor variables. This relative non-importance is weakly supported by the better performance of the models that excluded these variables. The timeseries comparison of these parameters versus algal population also does not reveal an obvious relationship between concentrations and counts.

 In conclusion, ANN-based real-time forecasting capability can be valuable for anticipating algal blooms before they occur and implementing proactive measures for mitigating such blooms accordingly, such as adding chemical treatments, switching to alternative water sources, and issuing health warnings. In addition, ANN technology can be used to help better understand the underlying. These forecasting models also provide value added to expensive data collection systems, and may even be used to optimize sampling strategies, potentially reducing costs.

Reference

1.Hecht-Nielsen, R. 1987. Counterpropagation networks. Proc. Int. Conf. on Neural Networks, II, p. 19-31. New York, IEEE Press.
2.Lee, J.H.W., Y. Huang, M. Dickman and A.W. Jayawardena. 2003. Neural network modelling of coastal algal blooms. Ecological Modelling 159(2-3): 179-201.
3.Maier, H., R.G.C. Dandy and M.D. Burch. 1998. Use of artificial neural networks for modeling cyanobacteria Anabaena spp. In the River Murray, South Australia. Ecological Modelling 105(2-3): 257-272.
4.Pelley, J. 2005. Can nutrient loads predict marine water quality? Environmental Science Technology 39(2): 37A-38A.
5.Poulton, M. 2001. Computational Neural Networks for Geophysical Data Processing. Amsterdam: Pergamon Press Ltd.
6.Smith, V.H., S.B. Joye and R.W. Howarth. 2006. Eutrophication of freshwater and marine systems. Limnology and Oceanography 51(1, part 2): 351-355.
7.Sprecher, D. 1965. On the structure of continuous functions of several variables. Transcations of the American Mathematical Society 115: 340-355.
8.Taylor, W.D., R.F. Losee, M. Torobin, G. Izaguirre, D. Sass, K. Khiari and K. Atasi. 2006. Early Warning and Management of Surface Water Taste-And-Odor Events. IWA Publishing.
9.Wang, P., D.M. Tartakovsky and A.M. Tartakovszky. 2013. Stochastic Forecasting of Algae Blooms in Lakes. Modeling and Simulation in Fluid Dynamics in Porous Media. Proceedings in Mathematics & Statistics 28: 99-108.