The exhaustion of non-renewable fossil fuels has heightened awareness about environmental issues. As a result, biomass energy has come into the spotlight as a promising renewable alternative, particularly in the context of bio-oil production through pyrolysis from waste biomasses. Unfortunately, physics-aware models pose dificulties when modelling bio-oil production, prompting researchers to lean towards data-centric approaches. To cope with this problem, this paper showcases a comprehensive dataset of nearly one thousand records sourced from prior literature about bio-oil production. Besides collecting, cleaning and organising the gathered data, we also used machine learning techniques to evaluate the resulting dataset, with the most promising result yielding a mean absolute error of 2.6 and an adjusted R-squared of 0.9 in prediction of bio-oil yield. To the best of our knowledge, this paper delivers introduces the most comprehensive dataset ever collated in this domain. The assembly of such an exhaustive dataset is pivotal for sustainable process engineering because it fosters precise modelling, thus better-addressing uncertainties inherent in the process.
CEUR ceur-ws.org
The accelerated depletion of non-renewable fossil fuel reserves has led to critical issues related to
environmental degradation. Consequently, biomass energy, which is abundant and sustainable,
has attracted considerable attention [
The uncertainty in the yield of bio-oil production and its qualities and the optimization of
process variables can be managed with mathematical models. Simulating pyrolysis for bio-oil
production using physics-based models is challenging. As artificial intelligence, and in particular,
the subfield of machine learning evolves, many data-centric approaches have been proposed
for estimating bio-oil yields from waste biomass properties and plant operational conditions,
acknowledging the non-linear correlations [
Section 2, ”Materials and Methods”, introduces the collected dataset related to waste biomasses and bio-oils needed to enforce data-centric research and the procedures used for handling missing data; Section 3, ”Results”, presents the comparison between diferent machine learning algorithms for predicting the yield of bio-oil; Section 4, ”Discussion with Conclusions” presents the conclusions elucidating the novelty of this work about the existing literature.
The following section introduces the dataset, which has been gathered from the literature to contribute to AI and renewable energy research in the context of waste biomasses. The missing data in the dataset has been addressed before establishing a machine-learning benchmark for predicting bio-oil yield, as shown in section 3. Section 2.1 presents the details of the dataset, while Section 2.2 describes the frameworks used to handle the missing data.
This study’s collected dataset, consisting of 1057 entries, includes proximate analysis of biomass like ash fixed-carbon and volatiles, ultimate analysis of biomass like carbon hydrogen oxygen and nitrogen, lignocellulosic content, and plant operative conditions as temperature, heating rate, particle size, and nitrogen flow rate. The target variable in this research work is the bio-oil yield.
The extent of missing data varies across diferent variables. For the lignocellulosic content, missing data account for between 40 and 50 percent of total entries, whereas for ultimate and proximate analyses, the missing data are close to 5 percent and around 10 percent, respectively. This prevalence of missing data necessitates the implementation of data imputation techniques before applying models to estimate bio-oil yield.
In this work, the issue of missing data was addressed by leveraging four distinct imputation
techniques, utilizing an iterative method in a round-robin fashion as explained in the
scikitlearn iterative-imputer doc1, a framework inspired on [
Lastly, a fourth approach, the Variational Autoencoder (VAE), was used for data imputation
purposes, as seen in [
Four versions of the complete dataset were created without missing values from the original dataset using the four imputation algorithms above. Machine learning algorithms were then employed to give a benchmark on bio-oil yields on this newly collected dataset. In Tables 1, 2, 3, and 4, the comparison of machine learning algorithms to predict the bio-oil yield is shown on the imputed versions of the datasets using KNN, random forest, support vector regressor, and variational autoencoder, respectively.
Models PMeeracennAtabgseolEurtreor Tree 0.1 Linear Regression 0.2 GBTree 0.1 RF 0.1 KNN7 0.2 MLP 0.2 LibSVM 0.2 RBFRegressor 0.2 KNN3 0.1 KNN5 0.1 AdditiveRegression 0.2
In Tables 1, 2, 3, and 4, it is shown that the handling of missing values with the iterative imputation using Random Forest combined with random forest for regression on bio-oil yield produces the best results, resulting in the mean absolute error of 2.6 and adjusted R-squared of 0.9. The models were trained using a hold-out split of 70-15-15 for training, validation, and testing.
The produced bio-oil yield from biomass waste was successfully predicted with random forest regression combined with an iterative imputation method based on Random Forest to tackle the problem of missing data, achieving an R-squared value of approximately 0.9 and a mean absolute error of about 2.6.
To the best of our knowledge, this is the benchmark on bio-oil yield on the more extensive
dataset collected in literature in the context of pyrolysis of biomass, which hence contains a
broader range of biomass properties and plant operative variables, concerning other studies
on diferent and smaller datasets like [