U ovom radu je predstavljena realizacija DC/AC konvertora
pomoću pojačavača u klasi D. DC/AC konvertori se prave u
širokom opsegu snaga, a ovde će biti prikazan DC/AC
konvertor snage 10 W. U ovom radu je predstavljena
elektronika za realizaciju DC/AC konvertora uz sva detaljna
objašnjenja i matematičke formule.

The design of Application-Specific Embedded Processors
(ASEPs) traditionally relies on manual, expert-driven
efforts, while automated exploration often employs
High-Level Synthesis (HLS) to generate data for Design
Space Exploration. However, HLS is constrained by its high
level of abstraction, limited model fidelity, and a
relatively small number of design variants that it can
effectively explore. To tackle the challenge of automated
Application-Specific Embedded Processor (ASEP) generation,
this paper proposes a novel four-tier optimization
framework that automates generation of designs from the
assembly level to RTL, systematically exploring the
application algorithm, instruction sets, hardware
architecture, and code generation. The proposed framework
keeps functional correctness for all generated designs and
demonstrates exceptional scalability. Two simple case
studies are investigated here to illustrate the validity of
the proposed framework, a 2D affine transformation, and
Gaussian filtering of images. Based on the same hardware
constraints, the framework has automatically generated and
optimized 192 and 1,728 extremely small RTL designs for the
2D affine transformation and Gaussian filter algorithms,
respectively, revealing a large theoretical design space.
ASEPs are traditionally designed manually, which is a
costly process requiring highly skilled expertise. To
reduce design cost, larger general-purpose processors are
frequently used instead of ASEPs. The automation framework
proposed here reduces the need for expert design skills and
can quickly generate and optimize application-specific
processors that have superior performance to their
general-purpose counterparts and are more than an order of
magnitude smaller. The small size is key to ultra-low power
implementations. In summary, this work provides a
systematic automated solution for constructing high-quality
ASEP by exploring large design spaces.

Digital predistortion (DPD) is widely recognized as an
effective technique for power amplifier (PA) linearization.
By compensating for nonlinearities in the PA transfer
characteristic, DPD ensures compliance of wireless
infrastructure with telecommunication standard
requirements, including bit error rate (BER), transmit
spectrum mask (TSM), and error vector magnitude (EVM). In
addition to improving linearity, DPD contributes to reduced
operational costs by enhancing PA energy efficiency. We
employed different algorithms to implement DPD. Proposed
DPD methods rely on complex valued memory polynomials and
neural networks. The methods are described in the paper
from mathematical standpoint. Simulation results are also
presented in the paper.

Power output measurements at solar plants are routinely
affected by gaps caused by faults such as communication
failures and data logging issues. When such gaps are
present, the usual practice of discarding the corresponding
samples reduces the amount of training data available for
machine learning models and can introduce bias if gaps are
not randomly distributed. This paper proposes a method for
imputing missing power output values using
k-nearest-neighbor (kNN) search based on Mahalanobis
distance computed over co-located meteorological
measurements. The method was applied to data from a
prosumer-regime solar plant with a 60 kW installed capacity
and validated through a hold-out experiment for a range of
values of k, in which 5% of complete samples were
artificially left-out, then recovered by the proposed
method and the left-out values are compared to
corresponding imputed values. Results show that the mean
absolute error of imputed values on the validation set is
below 5 kW for all tested values of k, with the lowest MAE
of 4.41 kW and RMSE of 7.53 kW achieved for k = 10.

This paper presents the design and implementation of a
proximity-based collision detection system using an HC-SR04
ultrasonic sensor and an Arduino Uno R3 microcontroller.
The system provides audible feedback through a passive
buzzer whose beeping frequency is dynamically adjusted
according to measured distance. Three distinct operational
zones are defined: continuous slow beeping for distances
above 15 cm, progressively increasing beep rate in the 5–15
cm range using the Arduino map() function, and a constant
tone for distances below 5 cm. The ultrasonic ranging
function is implemented manually without any external
library. Experimental results confirm consistent distance
readings within ±1 cm over the rated 2–400 cm range, and
audible zone transitions proved intuitive during user
testing. The design is compact, low-cost, and directly
extensible to multi-sensor or visual-feedback
configurations.

This paper investigates XGBoost-based approaches for
predicting photovoltaic power output at a 156 kW
installation in southeastern Serbia. Building on prior
LSTM- and MARS-based research, we introduce lag
features—derived from the plant's own production
history—and cyclical time encodings as novel feature
engineering steps. Three model configurations are compared:
a global plain model, a K-means clustered model routing
predictions by weather regime, and a calendar-season model.
A persistence baseline is established as a reference. Lag
features yield a 13% reduction in daytime root mean square
error (RMSE) relative to persistence. K-means clustering
(k=3) discovers irradiance and time-of-day regimes rather
than seasonal patterns, and per-cluster models provide no
measurable accuracy improvement over the global model.
Calendar-season splitting, tested to assess whether
dedicated seasonal models capture season-specific effects,
does not outperform the global model due to the reduced
training data available per season. The plain XGBoost model
with lag features is recommended as the optimal
configuration.

This manuscript presents the development of a real-time
simulation environment and the application of
Hardware-In-the-Loop (HIL) technology for testing an
industrial controller. The research includes implementing a
power system model in the Typhoon HIL Control Center
software, using its built-in SCADA interface for management
and monitoring. Special attention is paid to implementing
communication between the simulator and the industrial
controller using the Modbus TCP protocol, and to enabling
two-way system management. Finally, the conducted tests
verify the functionality of the developed system and assess
the applicability of the HIL approach for testing control
solutions in the power industry.

Increasing power density represents one of the key
development trends in modern power electronic converters.
This trend has encouraged the adoption of planar
transformers and integrated magnetic structures. However,
integrated magnetics are still often avoided in practical
designs, while many solutions reported in the literature
rely on custom-designed ferrite cores that are typically
either not commercially available or overly complex to
implement in compact enclosure. This paper presents a
practical method for magnetic integration in an LLC
resonant converter using commercially available cores. The
proposed approach combines simplicity and ease of
implementation with effective magnetic decoupling, achieved
through analytical design method. As a result, flux in the
shared leg is reduced and a nearly uniform flux
distribution within the ferrite core is ensured. The
proposed solution is validated through analytical approach,
simulations, and experimental results obtained from a 300 W
LLC converter implementing an integrated planar
transformer. Experimental results confirm the feasibility
of the proposed approach and show approximately 1%
efficiency improvement and reduced magnetic volume compared
to a conventional discrete magnetic implementation.

This paper presents an analysis of the oscillations of
disk-shaped metal endings (sonotrodes) designed for use in
ultrasonic transducers. The research aims to study and
predict resonant frequencies and vibration modes, with a
particular focus on the coupling of thickness and radial
oscillations. For modeling and displacement distribution
analysis, an analytical 3D approximate matrix method was
used, enabling the efficient prediction of thickness and
radial oscillations, as well as their mutual coupling. In
addition to the analytical model, a numerical model using
the finite element analysis software package was applied,
allowing for the prediction of the system's dynamic
behavior. Experimental measurements were carried out using
an industrial ultrasonic resonant analyzer in the frequency
range from 15 kHz to 50 kHz. The results show high
agreement between the analytical model, numerical
simulations, and experimental measurements. The research is
specifically focused on the identification and
characterization of axisymmetric flexural (bending) modes
that occur within the measurement range for steel and
duralumin sonotrodes, with dimensions most commonly used in
ultrasonic transducers. The proposed approach provides
reliable guidelines for optimizing sonotrodes geometry and
avoiding coupling with unwanted parasitic modes.