PSI - Issue 77

Michal Holubčík et al. / Procedia Structural Integrity 77 (2026) 413– 423 “Mi chal Holubčík” / Structural Integrity Procedia 00 (2026) 000 – 000

418

6

Table 3 : Comparison of advantages and disadvantages of types of neural networks in biomass boiler regulation [8], [9]

Neural Network Type

Abb.

Data Type

Advantages

Limitations

Feedforward Neural Network Recurrent Neural Network Long Short Term Memory Self Organizing Map Transformer / Seq2Seq Models

Simple structure, effective for direct relationships (fuel–air– emissions), fast training Models sequential combustion data; suitable for predicting emission trends Captures long-term dependencies; effective for predicting future states (emissions, output) Identifies hidden operating states; useful for classifying stable vs unstable combustion Superior for multi-step predictions; strong potential for real-time adaptive regulation

Single-instance mapping

Cannot capture time dependencies; less accurate under dynamic load changes

FNN

Training instability (vanishing gradients), requires large datasets

RNN

Time-series

Computationally demanding; slower to train

LSTM

Time-series

Pattern clustering (unsupervised)

No direct prediction; requires interpretation for control actions

SOM

Time-series forecasting

Very data-hungry; requires high computational power

–

3. Implementation of smart control in experimental small-scale biomass boiler setup

The experimental work was carried out on a small - scale biomass boiler LOKCA Úspor, a 15 kW unit designed for domestic heating. The boiler operates on wood pellets with a typical moisture content of 17% (class A1). In its original configuration, the system allowed control only of the primary combustion air, with start–stop fuel feeding and a ten-level fan adjustment. While suitable for basic operation, this limited flexibility highlighted the need for modernization and intelligent regulation. Figure 4 illustrates multiple systems that were implemented and that are necessary in order to convert this boiler into a smart-enabled machine.