Optimization of the selection process of the co-substrates for chicken manure fermentation using neural modeling

Intense development of research equipment leads directly to increasing cognitive abilities. However, along with the raising amount of data generated, the development of the techniques allowing the analysis is also essential. Currently, one of the most dynamically developing branch of computer science and mathematics are the Artificial Neural Networks (ANN). Their main advantage is very high ability to solve the regression and approximation issues. This paper presents the possibility of application of artificial intelligence methods to optimize the selection of cosubstrates intended for methane fermentation of chicken manure. 4-layer MLP network has proven to be the optimal structure modeling the obtained empirical data.

The most important step in ANN generating is to produce an adequate training set containing in its structure encoded empirical data.For this purpose, the Authors have defined the numeric input variables and predicted output variable resulted from the structure of the formulated scientific problem.Five input variables have been adopted as a share of respondents substrates: x GRASS -grass [g Fresh Mass (FM)/reactor], x STRAW -straw [g FM/reactor], x CHICKEN -chicken manure [g FM/reactor], x MAIZE -maize silage [g FM/reactor], x INOCULU -inoculum [g FM/reactor].

Table 1. Fragment of the training file for the ANN simulator (BIOMETHoutput variable)
The simulator of artificial neural networksimplementedin a commercialStatisticapackagehas been usedfor the design ofneural models.The following types ofneural networkshave been subjected totesting: x linear networks.
x RBF networks (Radial Basic Function).
x GRNN networks (Generalized Regression Neural Networkregressive networks).The development of neural modelswas a 2-stage procedure.At the beginning,an effectiveoption supporting the process of ANN designinghas been used, i.e.AutomatedWebDesignerimplementedin Statistica computer system.This toolallowed toautomateandsimplify theprocedures of preliminaryseeking of a set with predictiveneural networksmodeling the investigated process.In the second stage of neural models development another useful device has been used i.e.User Networks Designer.It offers the possibility of advanced interference in the parameters and training methods of generated neural networks.This tool has been activated frequently in order to modify both the initial settings of parameters, learning algorithms and the ANN structure itself.

Artificial neural networks
From 100of generatedneural models,a fileof10isolatedneuraltopologies has been extracted as it showsTable 2 (ANN no.10is the bestnetwork).Unidirectional MLP neural networks are commonly used in ANN topology practice.Multilayer Perceptron represents the so-called category of parametric neural models.Where characteristic is that the number of neurons constituting its structure is considerably less than the number of cases of the training set.The basic characteristics of MLP network include the following features: x MLP is a unidirectional network, x MLP is trained by "with-a-teacher" method, x has a multi-layer structure, with the following layers: input, hidden, output x architecture of the connections within the network allows communication only between the neurons located in contiguous layers, x neurons being a part of ANN, MLP type,aggregate the inputdataby defining theinputs weightedsums(using the linearformula ofaggregation), x activation functionof the inputneuronsis linear, hidden neurons-non-linear, whilethe nature ofoutput neuronsis generallynonlinear, x due to thesaturation level present(in the sigmoidactivationfunctions), all the data processed by the networkrequire an appropriaterescaling(preprocessing andpost processing).
The quality of the generated MLP network, as a predictive tool, is identified by the so-called statistics of regression issues, which are shown in Table 3. between the reference value and the value obtained at the output) for the output variable.-Error S.D. -the standard deviation of the errors for the output variable.-Abs E. Mean -mean absolute error (the difference between the reference value and the value obtained at the output) for the output variable.-S.D. Ratio -quotient of the standard deviations both for errors and data.This is the main indicator of the quality of the regression model developed by the network.-Correlation -standard R.Pearson correlation coefficient for the setpoint value and the value obtained at the output.Table 3shows thatthe correlation isat the level of0.99for the following files:training, validationandtest one, while the quotient of standard deviations for errorsand thedatarangesfrom 0.08in case ofvalidation file up to0.09forthe test file.The assessment of sensitivity of developed MLP on individual input variables has been performed in order to determine the level of significance of representative parameters used to build the neural model.The procedure of sensitivity analysis is implemented in the Statistica package as a tool for assessing the impact of the various input variables on the quality performance of generated neural model.The sensitivity analysis provides an insight into the usefulness of particular input variables.Moreover, it indicates high-ranking variables that without any loss of quality of the network can be omitted.Furthermore it also points the key variables (low rank value), which must not be ignored.-Rank -significance level of input variable, organizes the variables by importance: 1 is the dominant variable.-Error -the network quality in the absence of a variable: the lower the rank number of the ANN input variable, the higher error of reduced network (without the input variable).-Ratio -quotient of the error of the reduced network by the ANN error obtained using all variables.If the Ratio is less than one, than removal of the variable improves the quality of the network.
The sensitivity analysis of the MLPneural model5-8-4-1onthe input variables of analyzed processshowed thatthe most importantparametersin the process of neural estimation of the amount ofgeneratedbiomethaneare(in order): x

4.Discussion
The analysis of the empirical resultsobtainedduring the testsor receivedin the course of industrial processes controlsometimesis very complex.Usuallyit is related to a largeamount ofdata obtained.Duringcontinuous processessuch asindustrial production ofbiogasthroughoutmethane fermentationwe have to dealwith a specific case.In the aforementioned situationit is possible toobtain a veryextensivetraining setbased ona number of variablesanalyzedover a long periodof time [12,13].In order to properlyinterpret the multifactorialresultsit is necessary torefer toadvancedstatistical methodsbased onArtificial Intelligence.According to [14; 15] application of thistype of analysis enables defining the trends, dependencies and propercontrol of the processes,not only in thelaboratory scale butmainlyin the industrial one.

Conclusions
The resultsanalyzedusingANNhave shownthat the optimumstructuremodelingobtainedempirical dataisa 4layer MLP network.The receivedvisualizationsand statementsare consistentwith the experience ofa biogas plantstaffandscientists studyingthe efficiencyof biogassubstrates.It proves that chosen method can besuccessfully implemented into the plannedapplicationoptimizingthe selection of the co-substratesfor fermentation of the chicken manurein order to obtainthe highest possibleof biomethane production.
Rank 1: STRAW x Rank 2: MAIZE x Rank 3: GRASS In order to visualize the behavior of the generated neural model, depending on the values of the main descriptors (STRAW, MAIZE, GRASS) Figure 3 shows the three surfaces of ANN responds illustrating the biomethane efficiency in a function of key input variables of developed neural model.

Table 2 .
File of 10 generated ANN

Table 3 .
Regression statistics of generated model for the files: training.validation and test

Table 4 .
The values of quotients of errors and rank for the five input variables