PSI - Issue 77

Michal Holubčík et al. / Procedia Structural Integrity 77 (2026) 413– 423 "Michal Holubčík" / Structural Integrity Procedia 00 (2026) 000 – 000

417

5

In practice, an excess air coefficient λ of 1.5 –2.0 is required to ensure stable combustion. However, deviations in fuel moisture content (10–30%) significantly shift the optimum ratio, leading to either elevated CO concentrations (>500 mg/m³) or increased flue gas losses if not properly adjusted. Model predictive control (MPC) further extends this principle by optimizing multiple parameters simultaneously. Experimental studies have shown that MPC can reduce CO emissions by up to 40% compared to fixed-ratio methods while maintaining boiler efficiency above 85%. Dynamic tests under varying load conditions demonstrated quicker stabilization times —often under 2–3 minutes—compared with over 5 minutes for conventional PID controllers [5].

Figure 3 : Model based control of a small-scale boiler (flowchart) [6]

2.3. Intelligent based control – Neural networks Neural networks provide a powerful framework for intelligent combustion control, enabling real -time prediction of boiler performance and emissions. By learning complex, nonlinear relationships between fuel, air, and output, they surpass conventional method s. Their adaptability makes them essential for achieving stable operation, improved efficiency, and significant emission reductions. Belany et al. (2024) confirmed that artificial neural networks consistently outperform polynomial regression in predicting energy consumption, highlighting the advantage of advanced machine learning tools for smart energy applications [7].They consist of interconnected nodes, or neurons, that process information and learn from data. Key components include: 1. Input layer – receives data and passes it to hidden layers. 2. Hidden layers – extract patterns; their number and size affect performance. 3. Output layer – provides final predictions or classifications. Training occurs via backpropagation , which optimizes weights and biases by minimizing error. This enables networks to achieve high accuracy in prediction and optimization tasks. Applications range from image and audio analysis to process control, making neural networks highly flexible in modern research. Table 3 presents a summary of various neural networks, their data type, and their main advantages and disadvantages in the domain of combustion control [8], [9]. 4. Weights and bias – define the importance of signals and adjust neuron activation. 5. Activation functions (e.g., sigmoid, ReLU, tanh) – capture non-linear patterns.