• No se han encontrado resultados

DIMENSIONAMIENTO DEL SISTEMA DE AGITACIÓN

7 F : presión interna de diseño (MPa)

3.2. DIMENSIONAMIENTO DEL SISTEMA DE AGITACIÓN

ANNs can be successfully trained to describe the influence of energy sources, agricultural operations, and indirect factors on energy consumption in wheat production. The sample size used in this study was 40 farms. Initially, a sample of 30 farms (75%) was randomly selected for training, and the remaining 10 farms (10%) were used for validation. The inputs to the model were the reduced set of inputs found by PCA and an ANN was developed to relate energy consumption (output) to the selected input variables. The selection of the number of inputs and outputs is the first step in developing an ANN.

In this study, several network structures were examined to find the best model using the commercial software package, Peltarion Synapse1 (Appendix A). Specifically, the influence of (1) the number of hidden layers and neurons, (2) learning algorithm, (3) the type of transfer function in each neuron in the hidden layers and output layer in the model, and (4) type of network structure were studied to approximate the actual energy consumption. Genetic optimizer in Peltarion Synapse software was used to optimize weights, learning rates, and number of neurons. As stated before, Genetic algorithms are based on the biological theory of evolution and involve selection, crossover, and mutation of potential solutions to search for the optimum solutions. Optimization is generally a slow process as large search spaces are explored in the optimisation of each parameter.

1

Peltarion Synapse software package is an appropriate tool to design ANN models. Examining different structures and changing most elements in the Peltarion Synapse software are faster than most other modeling software such as Matlab. Furthermore, the software can optimize models using a genetic algorithm and it has a great ability to reduce errors, find the best number of neurons, and optimize weights and learning rates. However, the difficulty to accessing some outputs and the lake of tools to present the graphs and outputs that were necessary for comparing the results of different models, were the most important limitations of this software. The software contained four different operating modes that made up the development of the model including processing, design, training, and post processing (Appendix A).

In the first stage, a simple model with one hidden layer was selected and different learning methods, transfer functions, and other training elements were examined and the model was developed step by step. After initializing the network weights, the training was performed in batch form. Different learning methods, such as Back Propagation, Delta-Bar-Delta, Steepest Descent, Quick Prop, and Gauss–Newton learning methods were tested and the best algorithm was selected to adjust the weights to minimize the mean square error between the actual and predicted outputs.

Peltarion Synapse software provided useful facilities to change various elements in the models. Each model variant was trained for 100 iterations; then, the results, including MSE for training and validation data, were investigated and the best models were saved. Then the models were trained for the next 100 iteration and results were compared for each combination of different model parameters. Several combinations of transfer functions, iteration, learning methods, and numbers of layers created a large number of possibilities. Then, the modelling process extended to multiple hidden layers and several other more complex network structures, such as modular neural networks with a hebbian layer, and the optimisation was repeated.

In summary, the survey was developed carefully, in a step by step manner, to collect as much as information as possible in the easiest way. Total energy consumption was calculated using the most relevant energy conversion coefficient for energy inputs taken from different references. The relationships between the different direct and indirect factors were then examined using the Pearson product-moment correlation coefficient. After data reduction, a group of direct and indirect factors were selected and, based on these variables, and after examining different learning methods, transfer functions and hidden layers, the final ANN model was developed to predict energy consumption under different conditions. In the last step the results from the ANN Model were compared with the Multiple Linear Regression Model.

Chapter 5

Results

_____________________________________________________________________

The main objectives of this study were the determination and modelling of energy use in wheat production; as a secondary objective, it aimed to examine the relationship between energy consumption and direct and indirect parameters in wheat production as such investigation can provide important insights. A clear initial picture was gained through the collected data and their descriptive statistics: such as maximum, minimum, mean, and standard deviation (SD). For making inferences about population data, the 95% confidence interval was estimated for the most important data.

In the first section of this chapter, the most important factors of wheat production are explained; this information would be useful for other related studies. Specifically, correlations between different factors (direct and indirect) were investigated one-by-one through correlation analysis and graphical illustration. Due to the limited of data, it was difficult to present a lengthy discussion on each parameter and correlation; for this reason, they are explained only briefly. It was noticeable that these parameters would influence energy consumption directly or indirectly and investigating the correlations between different variables and energy consumption was necessary for modelling.

In the second section, energy use in wheat production is explored in detail taking into careful account all relevant direct and indirect inputs and operations. In the final section the artificial neural network (ANN) model development and result are explained. For assessing the performance of the final ANN model, it was compared with a multiple linear regression model, the common modelling method used in agricultural studies.