This paper proposes a non-invasive prediction method for the junction temperature of discrete IGBT devices based on the relationship between the temperature at the top of the device casing and the junction temperature. Correspondingly, a heat transfer model has been constructed. This method overcomes the limitations of traditional techniques: optical methods and physical contact methods require destructing the packaging; the temperature-sensitive parameter method interferes with the device's operating state. The study first confirms an approximately linear relationship between the device top temperature and the junction temperature under the condition of a single tube running independently. Next, the junction temperature measured using the temperature-sensitive parameter method serves as a reference value, and the top temperature is utilized to predict the junction temperature, with prediction errors calculated through comparison. Finally, the accuracy of this method is further validated by simulating actual operating conditions.

High-voltage, high-capacity AC/AC converters are widely used in various applications. Power conversion imposes diverse requirements on voltage amplitude, frequency, phase angle, and phase number. This paper proposes a novel multi-phase to multi-phase direct AC/AC converter, evolved from the modular multilevel converter (MMC) used for AC/DC conversion. Compared to the traditional modular multilevel matrix converter (M³C), the proposed topology significantly reduces submodule count, simplifies control strategies, and offers excellent scalability. 

This paper reviews HVDC tapping technologies with a focus on Modular Multilevel Converter (MMC) topologies, including series, parallel, and hybrid configurations. A novel MMC-CCV control strategy is proposed, employing an intermediate-frequency AC link to reduce submodule capacitance, improving suitability for low-speed applications. Integration challenges, control strategies, and AC/DC grid interactions are analysed, highlighting the potential of HVDC taps to enhance efficiency and fault tolerance in modern grids

In this paper, based on the fractional-order characteristics of actual inductance and capacitance, the fractional-order mathematical model of quasi-switched boost converter under continuous conduction mode is constructed on the basis of equivalent small parameter method (ESPM). By combining the ESPM with harmonic balance principle, the general state vector differential equation of fractional-order switched boost converter (FO-SBC) and its steady-state analytical appropriate periodic solution are obtained, which reveals that the dc parts and ripples of voltage of C or current of L are all related to the order of FO components. Ultimately, the accuracy and correctness of the theoretical analysis is validated by simulation on Matlab/Simulink platform.

This study proposes a high voltage gain pulse power supply based on the Sepic-Marx circuit, addressing the demands for compactness, portability, and high step-up ratios in low-voltage DC-powered applications such as outdoor portable devices and plasma propulsion. The topology integrates the Sepic structure with the classic Marx generator, retaining its modular design while utilizing Sepic inductors as isolation components to suppress rapid changes of input current. The working principle is analyzed, dividing each switching cycle into four operational modes: inductor charging preparation, high-voltage pulse output, capacitor charging with voltage pumping, and a waiting state. Simulation results in PSIM software show that a five-stage circuit achieves a pulse output of 1.7kV with a 10kHz repetition frequency and 5μs pulse width under a 50V input, yielding a voltage gain of 34. Compared with classical Marx and Boost-Marx generators, the proposed topology requires fewer stages to achieve the same gain, demonstrating distinct advantages in voltage step-up capability and modular simplicity.

In order to reduce the average power generation cost of offshore wind farm, it is necessary to optimize the location of substation and cable topology in the collection system, which is a typical combinatorial optimization problem. At present, the meta-heuristic algorithms represented by particle swarm optimization (PSO) algorithm have been proved to be effective in solving this problem. However, due to the coupling of the internal design variables and the need to consider a variety of constraints, there are generally a large number of local optimal solutions, which makes the conventional algorithms often fail to obtain satisfactory results. In order to improve the global optimization ability of the conventional algorithms, the simulated annealing (SA) strategy is introduced to the PSO algorithm, and several objective functions including cable purchase cost, cable laying cost and power loss cost are considered. Through the study case analysis, it is proved that the proposed method can effectively solve the collection system optimization problem of offshore wind farm, and is better than some traditional methods.

This paper proposes a data-driven approach to simplify Space Vector Pulse Width Modulation (SVPWM) in T-type three-level inverters by replacing conventional modulation logic with supervised learning models. Traditional SVPWM requires sector identification, vector timing computation, and waveform construction, which introduce complexity and hinder real-time applications on embedded platforms. To overcome this, four regression models—a five-layer multi-layer perceptron (MLP), Random Forest, Light Gradient Boosting Machine (LightGBM), and Extreme Gradient Boosting (XGBoost)—are trained to map three-phase reference voltages directly to horn-shaped modulation signals. Training data are generated from conventional SVPWM across varying modulation indices, and models are evaluated in terms of waveform accuracy. Results show that the five-layer MLP achieves the highest accuracy with mean absolute error (MAE) of 0.00125, while Random Forest also demonstrates strong performance. The trained neural model is integrated into a Simulink-based NPC inverter, where it reproduces modulation waveforms with negligible error and maintains Total Harmonic Distortion (THD) at the same level as traditional SVPWM. These findings confirm that machine learning provides an efficient and accurate alternative to conventional SVPWM, enabling reduced computational complexity while preserving modulation quality in inverter control systems.

In recent years, Graph Neural Networks (GNNs) have been extensively applied to wind farm power prediction. GNNs effectively model wind turbine interactions and spatial dependencies within a farm. In this paper, a novel attention-based GNN architecture is proposed to enhance the generalization capability of the prediction model. The architecture incorporates spatio-temporal dimensions and physical principles, to construct dynamic graph connections, for wind farm topology changes caused by turbine downtime. It can better learn the wake characteristics under various operating conditions of turbines and improve the accuracy of power prediction. To verify the effectiveness of the proposed model, a location map of an actual wind farm consisting of 72 wind turbines was established and the SCADA data of the wind farm was used as the learning sample and validation data. The contrast between the results of this model and the conventional GNN shows that the proposed model achieves a superior fit to the actual operational characteristics of the wind farm.

This paper presents a reinforcement learning-based frequency adjustment approach tailored for LLC resonant converters subject to abrupt load transients. To address the limitations of conventional PI controllers in handling rapid load-induced voltage deviations, a single RL agent is integrated into the control framework. Triggered only when a sharp change in output current is detected, the agent applies a one-time frequency correction to mitigate voltage disturbances without interfering with steady-state regulation. The overall control scheme retains the simplicity of PI regulation while enhancing transient response through event-specific learning. A Lyapunov-informed reward structure guides the agent's behavior to ensure convergence and system stability. Experimental results under step-load conditions validate that the proposed method effectively reduces voltage overshoot and accelerates recovery, providing an efficient and scalable enhancement for dynamic control in LLC converters.

Dust deposition on the surface of photovoltaic (PV) panels can degrade the efficiency of photovoltaic power generation. Conventional methods for identifying PV panel dust accumulation status suffer from low recognition accuracy and slow processing speed. To address this issue, this paper develops a DL-DAM-based dust concentration detection model, which quantitatively estimates PV dust accumulation levels using transmittance. Firstly, a dataset is constructed for training the model's transmittance estimation module, leveraging the principle of atmospheric light attenuation combined with image processing techniques. Secondly, multi-scale convolution and a dual attention mechanism are fused to extract image features. Finally, deep learning and wavelet transform are integrated to achieve both transmittance estimation and visual distribution mapping of dust accumulation severity. Experimental results demonstrate that the training error of the DL-DAM model stabilizes at approximately 0.025. In transmittance estimation, the model achieves a mean square error (MSE) of only 0.0005 and a root-mean square error (RMSE) of 0.0224, significantly outperforming the three comparative models. Furthermore, it exhibits high detection accuracy for high-concentration dust accumulation.