Key Takeaways & Executive Findings
- •• • The TCN model achieved the lowest MAE of 0.505, indicating superior fitting accuracy for coal property prediction, which translates to reduced error in real-time quality assessment for power plant combustion optimization. • • The RF model attained the lowest RMSE of 0.618, demonstrating minimal deviation between predicted and actual values, ensuring reliable quality control in coal blending and supply chain management. • • The TCN model exhibited the lowest CV of 0.042, signifying high consistency and generalization across test sets, which is critical for deploying models across diverse coal sources without recalibration. • • SHAP analysis identified peak 52 (1141.35–1157.65 nm) as the most influential across all models, with specific peaks for ash (peak 46: 1173.89–1179.23 nm), moisture (peak 18: 1382.19–1388.10 nm), and fixed carbon (peaks 5 and 28: 1914.79–1920.46 nm and 1267.38–1270.48 nm), enabling targeted spectral feature selection for enhanced model interpretability and potential sensor design.
Abstract
Rapid and accurate determination of coal properties is critical for process optimization, quality control, and supply chain management in the coal industry. This study integrates near-infrared (NIR) spectroscopy with machine learning algorithms to develop a fast and reliable framework for predicting coal components. Five algorithms—support vector machine (SVM), random forest (RF), temporal convolutional network (TCN), one-dimensional convolutional neural network (1D CNN), and gated recurrent unit (GRU)—were employed for regression modeling. Among these, the TCN model achieved the lowest mean absolute error (MAE) of 0.505, while the RF model exhibited the lowest root mean square error (RMSE) of 0.618, indicating robust predictive accuracy on the test set. The TCN model also demonstrated superior generalization with the lowest coefficient of variation (CV) of 0.042. Single-output models revealed differential performance: TCN was optimal for ash content prediction, while RF excelled for fixed carbon and low calorific value. Considering model parameters, computation time, and accuracy, RF, 1D CNN, and TCN were identified as the most efficient. SHAP value analysis identified key spectral peaks influencing predictions, with peak 52 (1141.35–1157.65 nm) showing the highest impact across all models. Distinct peaks were associated with specific components: peak 46 (1173.89–1179.23 nm) for ash, peak 18 (1382.19–1388.10 nm) for moisture, and peaks 5 and 28 (1914.79–1920.46 nm and 1267.38–1270.48 nm) for fixed carbon. This research provides a novel method for rapid coal testing and offers insights applicable to component analysis in other fields.
1. Introduction
Conventional coal property analysis relies on laboratory-based methods such as proximate analysis and calorimetry, which are time-consuming, labor-intensive, and unsuitable for real-time process control. These methods introduce significant delays between sampling and actionable data, hindering the ability to adjust combustion parameters promptly, thereby reducing efficiency and increasing emissions. Moreover, the heterogeneity of coal necessitates frequent sampling, escalating operational costs and logistical complexity. The coal industry urgently requires rapid, non-destructive, and accurate techniques to overcome these bottlenecks.
Near-infrared (NIR) spectroscopy offers a promising alternative due to its speed, non-destructive nature, and capability to capture molecular information. However, the complexity of NIR spectra and the subtle spectral differences among coal components pose challenges for conventional chemometric methods. This study addresses these limitations by integrating NIR spectroscopy with advanced machine learning algorithms, including SVM, RF, TCN, 1D CNN, and GRU. The systematic comparison of these models, coupled with SHAP value analysis, provides a robust framework for accurate coal component prediction, demonstrating significant improvements in accuracy and efficiency over traditional approaches.
Loading authentic research manuscript (Pages 1–5)...
AN Haiquan, LIU Zhen, LI Ye, PENG Baozi (2026). Identification of Coal Characteristics by Near-Infrared Spectroscopy: Machine Learning Predictions and Experimental Validations. Journal of Fuel Chemistry and Technology. https://doi.org/10.1016/S1872-5813(25)60611-6
Research & Educational Purpose Only: The translations, structured abstracts, analytical annotations, and data reports provided by SinoGreenTechare intended exclusively for academic research, internal corporate R&D, and educational benchmarking. They do not constitute formal engineering, chemical safety, legal, or professional advice.
Copyright & Intellectual Property Notice: Original copyright of the underlying source articles and experimental data remains with the respective authors, institutions, and original publishing journals. SinoGreenTech claims intellectual property only over its proprietary translations, analytical syntheses, and AEO structured enhancements in accordance with international fair use and academic citation principles.
Frequently Asked Questions
What are the failure mechanisms of the proposed models under varying coal moisture levels, and how do they affect prediction accuracy?
The study reports that the TCN model achieved the lowest MAE (0.505) and CV (0.042), indicating stable performance across test sets. However, specific failure modes under extreme moisture conditions were not explicitly analyzed. The SHAP analysis identified peak 18 (1382.19–1388.10 nm) as uniquely associated with moisture, suggesting that spectral variations in this region could impact predictions. For robust deployment, further validation across a wider moisture range is recommended.
How do the computational costs and model parameters of RF, 1D CNN, and TCN compare, and what are the implications for real-time deployment?
The study indicates that RF, 1D CNN, and TCN are more efficient when considering model parameters, computation time, and accuracy. While exact parameter counts and inference times are not provided, the efficiency ranking suggests that these models are suitable for real-time applications. RF is a non-parametric ensemble method with lower training time, whereas TCN and 1D CNN require more computational resources but offer higher accuracy. For industrial deployment, trade-offs between accuracy and latency must be evaluated based on specific process requirements.
Can the identified characteristic spectral peaks (e.g., peak 52 at 1141.35–1157.65 nm) be used to design a low-cost, dedicated NIR sensor for coal quality monitoring?
Yes, the SHAP analysis highlights peak 52 as the most influential across all models, indicating that this spectral region carries critical information for predicting ash, moisture, and fixed carbon. A dedicated sensor could be designed to capture this narrow wavelength range, potentially reducing cost and complexity. However, the study also identifies component-specific peaks (e.g., peak 46 for ash, peak 18 for moisture), suggesting that a multi-wavelength sensor might be necessary for comprehensive analysis. Further validation on diverse coal samples is required to confirm the robustness of these peaks.
What is the expected improvement in prediction accuracy compared to traditional linear regression methods, and how does the model handle non-linear spectral relationships?
The study demonstrates that machine learning models, particularly TCN and RF, achieve low MAE and RMSE values (0.505 and 0.618, respectively), indicating high accuracy. Traditional linear methods often struggle with the non-linear and overlapping spectral features of coal components. The use of non-linear models like TCN and RF allows for capturing complex interactions, as evidenced by the superior performance. The SHAP analysis further reveals non-linear contributions of spectral peaks, underscoring the necessity of advanced algorithms for accurate predictions.
How transferable are these models to different coal types or geographical origins, and what retraining or calibration is needed?
The study does not explicitly address cross-coal-type generalization. However, the low CV (0.042) for TCN suggests good generalization within the test set. For new coal types, spectral variations may require model retraining or transfer learning. The identified characteristic peaks provide a basis for feature selection, potentially reducing the need for extensive recalibration. It is recommended to validate the models on a diverse coal library and implement periodic updates to maintain accuracy.
Related Chinese Research & Cross-Citations
Recent Advances in Carbon-Based Materials for CO2 Capture and Utilization
CO2 capture and utilization (CCU) technologies are critical for mitigating global warming and promoting resource circularity. Carbon-based materials, with tunable pore structures, abundant active sites, high specific surface area, and excellent chemical stability, show significant potential for CO2 capture and conversion. This review systematically analyzes the adsorption behaviors and performance variations of activated carbon, porous carbon, graphene, and carbon nanotubes in CO2 capture. For utilization, recent advances in catalytic applications for methanation, reverse water-gas shift (RWGS), dry reforming of methane (DRM), and alcohol synthesis are emphasized. The benefits and drawbacks of carbon materials regarding adsorption capacity, catalytic activity, and stability are evaluated, and their potential in integrated CCU technologies is discussed. Key strategies for enhancing performance through structural modulation and surface modification are elucidated. This review provides theoretical guidance for future development and large-scale implementation of carbon-based materials in CCU.
Fabrication and Microwave Absorption Performance of FexOy/TiO2/C Composites Derived from Red Mud
Red mud, an industrial solid waste from alumina production, poses severe environmental challenges. This study presents a resource-efficient strategy to convert red mud into high-performance microwave absorbing materials. FexOy/TiO2/C composites were synthesized via a sol-gel method using starch as carbon source, followed by carbothermal reduction. The phase composition and microstructure were optimized by adjusting calcination temperature and raw material ratio. The optimal sample, RmCT-5.4-700, exhibited a minimum reflection loss (RLmin) of -30.2 dB at 14.0 GHz with an effective absorption bandwidth (EAB) of 5.3 GHz at a coating thickness of 2.0 mm. The superior absorption performance is attributed to the synergistic effects of dielectric components (TiO2, graphitized carbon) and magnetic components (Fe3O4/Fe). Carbothermal reduction introduces defects that induce dipole polarization, while the conductive network formed by graphitized carbon and Fe3O4/Fe particles enhances conductive loss. Heterogeneous interfaces between Fe3O4, Fe, TiO2, and the red mud matrix promote interfacial polarization. The magnetic loss of Fe3O4/Fe improves impedance matching, facilitating electromagnetic wave penetration and absorption. This work not only provides a novel route for red mud valorization but also contributes to the high-value utilization of solid wastes.
Damage Mechanism of High Chromia Refractory in the Slag Tapping Hole of Commercial Entrained-Flow Gasifiers
The service life of refractory bricks in the slag tapping hole of entrained-flow gasifiers is a critical bottleneck for long-term stable operation. This study investigated the damage mechanism of high chromia refractories in four commercial coal-water slurry gasifiers by analyzing gasification coal samples and corroded refractory bricks. Slag characteristics, including crystallization and viscosity-temperature behavior, were evaluated. Results revealed that low-viscosity slag induces more severe refractory damage. To mitigate slag crystallization risk, a safe slag tapping temperature range is recommended as tICT−t2.5 when tICT exceeds t25. Interior morphology of corroded bricks exhibited cracks, primarily attributed to molten slag penetration and subsequent reactions with refractory material. SEM-EDS analysis of slag-aggregate and slag-matrix interfaces identified reduction in Cr2O3 content as the earliest damage characteristic. XRD detected no zirconium-containing spinel in cracks, indicating that thermal expansion mismatch between newly formed phases and the refractory matrix drives crack propagation. A damage mechanism is proposed: initial Cr2O3 depletion compromises both matrix and aggregate, facilitating slag ingress and new phase formation, ultimately leading to structural failure. Early detection or prevention of Cr2O3 reduction is essential to prolong refractory service life.
Research advances in the pyrolysis recycling of waste wind turbine blades
The global energy landscape is undergoing a profound transformation, with wind energy gaining increasing prominence due to its clean and renewable nature. However, as installed wind power capacity expands, disposal of waste wind turbine blades (WWTB) has emerged as a significant challenge. These blades are predominantly composed of epoxy resin (EP) polymers, carbon fibers (CFs), and glass fibers (GFs). Improper disposal exacerbates environmental concerns and leads to loss of valuable resources, particularly carbon-based materials. Pyrolysis technology, a versatile and environmentally sustainable method for resource recovery, has garnered considerable attention for WWTB disposal. This work presents a comprehensive review of pyrolytic recycling of WWTB, focusing on principles and classifications of pyrolysis technology, key factors influencing the pyrolysis process, as well as pyrolysis methods, equipment, products, and their applications. Through in-depth analysis of current research, this review identifies critical unresolved issues and provides a forward-looking perspective on emerging research trends. The review highlights that pyrolysis can effectively recover glass fibers and carbon fibers with mechanical property retention depending on process conditions, and that catalytic pyrolysis can enhance the quality of recovered products. Economic analysis indicates that collaborative disposal methods can improve cost-effectiveness. Future research should focus on optimizing process parameters for large-scale industrial application and developing more efficient catalysts to improve product selectivity and fiber quality.
Citric Acid-Modified HUSY Zeolite Catalyzes Alkylation of Phenol with Cyclohexanol for High-Density Aviation Fuel Precursors
Lignin-derived oxygenated aromatics, particularly phenols and aromatic ethers, are promising feedstocks for synthesizing high-density, high-heat-sink aviation fuels via alkylation-hydrogenation processes. This study systematically evaluates the catalytic performance of various zeolites (Hβ, HZSM-5, MCM-41, and HUSY) in the alkylation of phenol with cyclohexanol. Characterization demonstrates that HUSY zeolite exhibits superior catalytic activity due to its favorable pore architecture and well-balanced acid site distribution, which synergistically facilitate molecular diffusion and catalytic transformations. To further enhance catalytic properties, HUSY was modified with citric acid at various concentrations and compared with NaOH and oxalic acid treatments. Results reveal that citric acid treatment preserves crystallinity while modulating acidity and pore structure. All modified zeolites enhance phenol alkylation activity. Notably, HUSY-0.5M, exhibiting the highest medium-strong acid to total acid ratio, achieves superior performance: 80.4% phenol conversion and 99.6% selectivity for alkylation products. The catalyst also shows high activity for various lignin-derived compounds (p-cresol, anisole, guaiacol), demonstrating broad applicability. This work provides a new strategy for valorizing lignin-derived phenols into high-value fuel precursors through alkylation.
Hydrogen Production and Structure Evolution Mechanism during Thermochemical Conversion of Microalgae Pellet in Molten Hydroxide Salts
This study investigates the thermochemical conversion behavior of microalgae pellets in a molten hydroxide salt (80% NaOH-20% Na2CO3) system and its influence on hydrogen production. By comparing temperature evolution, gas release characteristics, and structural evolution of pellets with and without molten salt, and integrating char alkalization experiments, the regulatory mechanism of molten salt on reaction pathways and hydrogen production was systematically analyzed. Results indicate that molten salt significantly enhances internal heat transfer efficiency, achieving a central heating rate of 177 °C/s, effectively alleviating thermal hysteresis. Concurrently, molten salt promotes pore development through penetration, erosion, and catalytic effects, resulting in a porosity increase of 53.2%–104.3% after 10 s of reaction. Conversion efficiency is markedly improved, with the dominant reaction pathway shifting to char alkalization after only 70 s. Furthermore, when heating rate is increased above 600 °C, hydrogen yield from char alkalization improves more significantly, primarily attributed to the synergistic promotion of molten salt catalysis and rapid heating on volatiles reforming. This study provides a theoretical foundation for understanding efficient hydrogen production from biomass in molten hydroxide salts.