Thermokinetics, reaction modelling, and machine learning DNN and LSTM analysis of heat-induced pulverised Citrus clementina peel

For the first time, this study investigates the use of Citrus clementina (CP) peel as a pyrolysis feedstock, analysing its thermal decomposition profiles through established kinetic methods and deep learning models, specifically, deep neural networks (DNN) and long short-term memory (LSTM) networks. The pulverized CP biomass material exhibits lignocellulosic properties, confirmed by physicochemical characterization that revealed a high carbon content of 42.7 %, oxygen content of 50.6 %, and a particle size distribution ranging from 150 to 300 µm. Thermogravimetric (TGA) analysis conducted at heating rates of 5, 10, 15, and 20 °C.min⁻¹ , over a temperature range of 30–950 °C, indicates that the thermal degradation of the material is a complex multi-step process, revealing distinct deconvoluted derivative thermogravimetric (DTG) peaks and overlapping temperature ranges signifying moisture removal (30 – 185 °C), as well as the decomposition of hemicellulose (140 – 305 °C), cellulose (224 – 460 °C), and lignin (274 – 745 °C) contents of the biomass. Thermokinetic analysis of the experimental measurements identified that the Zhuravlev, Lesokin, Tempelman (D5) diffusion model exhibited the best fit, achieving an R-squared value of 0.98, closely followed by the second-order (F2) reaction model and the Avrami-Erofe’ve (n = 1.5) model. The estimated average activation energies were determined to be 103.2 ± 6.6 kJ·mol⁻¹ across the entire thermal spectrum (2.5–95 % conversion), and 128.2 ± 7.8 kJ·mol⁻¹ for the primary volatile matter release (10–90 % conversion). These values were calculated using multiple iterative and model-free methods, specifically Flynn-Wall-Ozawa (FWO-Ir), Kissinger-Akahira-Sunose (KAS-Ir), Ortega (OM), modified Ortega (MOM), Vyazovkin (VyM), and Friedman (FR). The estimated thermodynamic parameters further indicate that the formation of activated complexes is endothermic, non-spontaneous, and exhibits less disorder. Furthermore, modelling and simulation utilising the selected machine learning techniques demonstrated a high degree of confidence for both the DNN (R-squared value of 0.995) and LSTM (R-squared value of 0.978) architectures, following a series of modelling adjustments and extended training periods. It is anticipated that the methodologies employed in this study, along with the estimated data, will provide valuable insights to energy experts, environmental analysts, and policymakers regarding the potential utilization of CP biomass material as a viable pyrolysis feedstock for bioenergy and biochar generation.