Development of neural network-based systems to rank feedstocks for biodiesel production

The transition from fossil fuels to renewable energy sources is a pressing concern in the current century. Second-generation biodiesel derived from agroindustrial residues are playing a key role in this transition, offering a sustainable alternative that does not compete with food resources. In addition, biodiesel provides a solution for the utilisation of low-cost commercial raw materials, thus reducing waste. Substantial investments in improved mixing and storage facilities contribute to a robust and competitive biodiesel industry. 

Although numerous plant species have been identified as suitable for biodiesel produc- tion, the challenge lies in identifying the most suitable feedstock for biodiesel production for specific situations, taking into account various factors that characterise and qualify biodiesel, including fuel properties, engine performance, and emission characteristics. 

To bridge this gap, an innovative approach was developed for ranking feedstocks using secondary data sources from the published literature, employing machine learning tech- niques such as Random Forests (RF), Artificial Neural Networks (ANN), and Genetic Programming (GP). The uniqueness and novelty of this research lie in the development of these innovative models, which utilise data sources extracted from published litera- ture to rank feedstocks based on their composition, physico-chemical properties, and relevant economic, environmental and engine parameters. By training these models on known feedstock properties, one may be able to predict target properties for more than a hundred feedstocks from incomplete information. ANN models provide the highest accuracy, while GP offers greater transparency. 

This study consists of three main phases : data preprocessing and feature selection, predictive modelling and optimisation of key biodiesel characteristics, and feedstock ranking. Using advanced methodologies such as the multiple imputation by chained equations (MICE) algorithm for addressing missing data, RF, ANN and GP for predictive modelling, and multi-objective optimisation (MOO) and weighted single-objective optimisation (WSOO) approaches for feedstock ranking, this research yields significant insights into the selection of feedstocks in the production of biodiesel. The integration of these approaches enables informed decision making and promotes the advancement of sustainable and effective biodiesel production methods. 

To optimise objectives or target variables, two different approaches are employed. The first approach in which MOO maps projected properties and compositions to the nearest feedstocks in the database, allowing manufacturers to choose from a range of non- dominated solutions that offer possible trade-offs between different feedstock targets. i ii The other approach where WSOO involves assigning weights to different objectives, effectively converting them into a single objective, thus resulting in a linear ranking of feedstocks based on fixed priorities or relative weights. 

Through the application of proportional weights to prioritise the different target vari- ables, namely Production Cost, Yield, Load, Conversion rate, and CO, CO2, HC and NOx emissions across three different scenarios, a comprehensive and unbiased assessment framework for feedstock suitability has been established. This holistic approach ensures that economic viability, physicochemical efficiency, environmental impact, and long- term sustainability are thoroughly evaluated, thus propelling the progress of effective and sustainable biodiesel production practices. 

Furthermore, the properties of the biodiesels evaluated were found to meet the stan- dards of ASTM D6751-2, EN 14214, as well as the Australian biodiesel standard. Among the evaluated feedstocks, Millettia pinnata (Karanja), Madhuca indica (Mahua), Persea americana (Avocado), Terminalia catappa (Almond), and Eruca sativa (Taramira) demon- strated exceptional performance in different parametric settings. These feedstocks were identified as the top five choices based on the desired performance characteristics. On the other hand, Coconut (Cocos nucifera) and Croton (Croton megalocarpus) secures the tenth and eleventh ranks, while Groundnut (Arachis hypogeae) and Beauty Leaf Tree (Calophyllum inophyllum) comes in the sixteenth and seventeenth ranks, respectively. 

Using the potential of state-of-the-art modelling paradigms based on artificial intelli- gence and machine learning, the present research offers a robust and efficient method to identify the most suitable feedstocks for biodiesel production, optimise objectives, and achieve sustainable first- and second-generation biodiesels. This transformative approach seeks to empower investors and manufacturers to make informed decisions while selecting the most appropriate feedstocks, thus enhancing the overall sustainability of biodiesel production.