Research progress in modeling and intelligent control of anaerobic fermentation processes for agricultural wastes

Anaerobic fermentation converts crop residues, livestock manure, and other agricultural wastes into energy-rich products, particularly methane and hydrogen, simultaneously enabling waste treatment and resource recovery. However, stable operation and industrial scale-up remain difficult because feedstock properties, microbial metabolism, fluid behavior, and mass transfer vary continuously throughout fermentation. Their interactions generate nonlinear, time-varying, and multiscale dynamics that are difficult to observe, predict, and regulate using conventional methods. This review synthesizes progress in modeling and intelligent control for the anaerobic fermentation of agricultural wastes, with particular attention to biochemical hydrogen production. The reviewed advances are organized around a hierarchical data-mechanism-control framework that connects process characterization, hybrid modeling, online estimation, optimization, and system-level decision support. First, the review examines heterogeneous data generated during feedstock characterization, pretreatment, reactor operation, microbial conversion, and product formation. Laboratory and pilot-scale datasets are often sparse, noisy, imbalanced, and dependent on specific feedstocks or reactor configurations. Second, the review compares major modeling strategies and clarifies their respective roles. Random forests and support vector machines are effective for modeling feedstock characteristics and pretreatment responses, whereas artificial neural networks and long short-term memory networks are capable of capturing nonlinear relationships and temporal dependencies. Gaussian process regression supports uncertainty analysis under limited data, while transfer learning can reuse information across related feedstocks or operating scenarios. Despite these capabilities, purely data-driven models may generate unreliable predictions outside their training domains and may violate conservation laws, thermodynamic constraints, or microbial kinetics. Mechanistic models, including anaerobic digestion model No. 1 (ADM1), provide interpretable process descriptions but require extensive calibration and may inadequately represent complex slurry transport and microbial succession. Physics-informed neural networks (PINNs) provide a gray-box modeling approach by embedding conservation relationships and microbial kinetic equations into the learning objective. Explainable artificial intelligence (XAI), particularly SHapley Additive exPlanations (SHAP), complements this approach by identifying influential variables, interactions, inhibitory conditions, and potential process thresholds. These tools can translate model outputs into operationally meaningful evidence for process diagnosis and control, although their reliability under practical operating conditions remains to be validated in reactors treating agricultural wastes. Third, the review links hybrid modeling to closed-loop process control. Internet of things (IoT) sensor networks and edge computing can support local filtering, drift correction, feature extraction, communication management, and rapid fault detection. Soft sensors can estimate difficult-to-measure state variables, including volatile fatty acids, chemical oxygen demand, alkalinity, and microbial activity, from readily available online measurements. Multi-objective optimization and model predictive control (MPC) can then balance gas productivity, process stability, and energy consumption while satisfying equipment and safety constraints. At the system level, digital twins integrate biochemical kinetics, computational fluid dynamics, sensor data, equipment models, and scenario simulation. Nevertheless, applications in agricultural waste fermentation remain largely conceptual or limited to individual components because biological and hydrodynamic processes are difficult to couple efficiently. Future research should prioritize standardized multimodal databases, shared biochemical and hydrodynamic model libraries, adaptive hybrid models, robust sensing, uncertainty-aware control, and integrated device-edge-cloud architectures. Progress in these areas can advance interpretable prediction, resilient process operation, and reliable bioenergy recovery, thereby supporting the digital and intelligent transformation of agricultural waste management systems.