The rapid progress of text-to-speech (TTS) and voice conversion (VC) technologies has enabled the creation of synthetic voices that closely mimic real speakers. While these systems have beneficial applications, they also raise security concerns, being susceptible to misuse by fraud, impersonation, and disinformation. In this paper, we propose a synthetic speech detection system based on conventional convolutional neural network (CNN) models trained from scratch on cepstral-based and constant-Q cepstral coefficients (MFCCs, LFCCs, CQCCs) as inputs. To improve robustness and reduce overfitting, we employ targeted augmentation methods (Mixup, SpecAugment) alongside regularization techniques including batch normalization, dropout, label smoothing, and adaptive learning rate scheduling. Our experiments are carried out on the Fake-or-Real (FoR) corpus, a large-scale benchmark that includes both authentic human recordings and outputs from a variety of TTS and VC systems. Our best system configuration achieves 99.4% validation accuracy and 95.9% test accuracy, outperforming our previous work and demonstrating that carefully designed CNN architectures combined with focused augmentation and regularization can deliver high detection performance.
This paper addresses the complexity of power grid operation behavior by developing a multi-scenario preview and dynamic verification method leveraging virtual reality (VR) technology. The method integrates real-time positioning data, operation ticket information, and environmental variables to construct an immersive 3D virtual scene that closely replicates actual grid equipment and conditions. Within this virtual environment, the behavior and responses of grid operators under various fault scenarios are simulated, enabling early risk identification and assessment. Dynamic verification algorithms are applied to validate operational behavior in real time, providing data-driven decision support for dispatchers. Experimental results demonstrate that the VR-based approach achieves a success rate of at least 85% in scenarios including line faults, equipment faults, load overload, and frequency fluctuations, with an average response time of 16.5 seconds and a behavior simulation recognition accuracy of about 90%. Additionally, the dynamic verification algorithm recorded a MAE (Mean Absolute Error) of 0.03 and an MSE (Mean Squared Error) of 0.002 within 60 seconds. These findings highlight the innovative integration of immersive VR and dynamic verification techniques, offering a robust tool for enhancing the safety and reliability of power grid operations.
Power electronic devices have gradually become the core components of modern energy systems. However, due to their nonlinear dynamic characteristics and operational safety of the system. Therefore, a stability optimization method for power Optimization (IIW-PSO) is proposed. This method utilizes the IIW-PSO to optimize controller parameters, thereby reducing dependence on accurate mathematical models, and combines active damping compensation to adjust damping strength in real time. The voltage tracking error was the lowest at 1.25%, and the Total Harmonic Distortion (THD) was the lowest at 3.3%. In the practical application of stability testing for power electronic equipment, the voltage surge of the improved power electronic equipment system was only 2.9%, the highest steady-state error was only 2.1, and the maximum dynamic disturbance of the output voltage was 20V. The stability method based on the IIW-PSO has excellent anti-interference performance, can meet the high-performance and high reliability requirements, and further promote the development of power electronic equipment towards intelligence and high adaptability.
This paper provides a detailed analysis of the wave energy potential at three locations in the Black Sea, evaluating the performance of various WEC technologies. The paper highlights the difficulties of implementing WEC as well as a series of financial data on the electricity produced. In this regard, the paper may be useful for operators in coastal areas to assess the possibilities of practical use of wave energy.
In recent days, there has been a significant interest in using Machine Learning (ML) techniques in different domains that serves people's lives. Interpolation and extrapolation are fundamental techniques in data prediction and ML. In contrast, interpolation estimates values within a known data range, whereas extrapolation attempts to predict beyond the observed data points. This study compares the accuracy and limitations of these approaches using synthetic datasets modeled on sinusoidal functions with noise. We employ cubic interpolation and polynomial regression to assess their predictive reliability. Results show that interpolation provides highly accurate estimates within the given range, whereas extrapolation introduces significant deviations as it moves further from known data. Our findings highlight the risks of relying on extrapolation in ML applications and emphasize the importance of selecting appropriate methods based on data distribution and predictive constraints. To further support the analysis, we implemented linear regression and a Moving Average model as a non-parametric alternative to Random Forest. Performance was evaluated using Root Mean Square Error (RMSE) and Mean Absolute Error (MAE), both numerically and analytically. Polynomial regression of degree 5 proved most effective for interpolation, while cubic splines showed instability in extrapolation. These insights reinforce the need for tailored model selection in predictive modelling tasks.
