HHVannPRO: ANN trained to predict the biomass HHV based on the results of the proximate analysis

Published: 20-01-2021| Version 2 | DOI: 10.17632/dx8kbs5yxc.2
Contributor:
Olga Jakšić

Description

For the comparative analysis of different algorithms for training the ANNs the same network structure is used but with diferent training functions. In all cases this network was fed with the same sets of input data. Data sets were formed on the basis of literature data on measured HHV of biomass characterized by the proximate analysis. In order to improve the ANNs response, the training data sets were augmented with the additional literature data on measured HHV and biomass composition in terms of the proximate analysis, so that the final set of data consisted of 318 training input records and corresponding 318 outputs. The custom designed MATLAB functions were then applied to a new set of input data completely unknown to ANNs and the results were recorded. Folder ANNmatlabDefault comprises files related to training functions Levenberg-Marquardt, Bayesian Regularization ans Scaled Conjugate Gradient. Other folders are related to other training functions: Gradient Descent, Gradient Descent with Momentum, Variable Learning Rate Gradient Descent, One Step Secant, Polak-Ribiére Conjugate Gradient, Fletcher-Powell Conjugate Gradient, Conjugate Gradient with Powell/Beale Restarts, Resilient Backpropagation and BFGS Quasi-Newton.

Files

Steps to reproduce

Input for Matlab/octave functions is a 3xN matrix with 3 raws referring to percentages of fixed carbon, volatile materials and ash (results of the proximate analysis of biomass), and N is the number of different examples. Folder ANNmatlabDefault comprises files related to training functions Levenberg-Marquardt, Bayesian Regularization ans Scaled Conjugate Gradient. Other folders are related to other training functions: Gradient Descent, Gradient Descent with Momentum, Variable Learning Rate Gradient Descent, One Step Secant, Polak-Ribiére Conjugate Gradient, Fletcher-Powell Conjugate Gradient, Conjugate Gradient with Powell/Beale Restarts, Resilient Backpropagation and BFGS Quasi-Newton. First step is to import data from Microsoft Excel (or elsewhere) to MATLAB Workspace so that each result of proximate analysis is a separate variable: FC, VM, ASH, and if exixts a column with related measured HHV named HHVJkg. functions named ann???.m where ??? stays for the ANN algorithm acronym calculate HHV of the biomass based on the input consisted of FC, VM and ASH. Scripts named train???name-of-the-training-function.m use sets of example data to train new ANNs. They are editable and one can change the network structure and parameters. Data are given in three ways: in Excel file with references, in Excel file without references and as a saved MATLAB Workspace. The output is teh higher heating value of the biomass, given in MJ/kg.