KMITL

Permanent URI for this communityhttps://dspace.kmitl.ac.th/handle/123456789/1

Browse

Search Results

Now showing 1 - 5 of 5

Some of the metrics are blocked by your
consent settings
Advanced Short-Term Wind Power Forecasting Based on CNN-BiLSTM - Lightweight Self-Attention (LWSA)
(2025-01-01)
Huang, Zhibin
;
Jiriwibhakorn, Somchat
Accurate short-term wind power forecasting is critical for maintaining grid stability and enhancing energy dispatch. However, the nonlinear, volatile, and uncertain nature of wind power poses significant challenges to traditional and deep learning models. To address this, a hybrid model named CNN-BiLSTM-LWSA is proposed, which integrates Convolutional Neural Networks (CNN) for local pattern extraction, Bidirectional Long Short-Term Memory (BiLSTM) networks for temporal modeling, and a Lightweight Self-Attention (LWSA) mechanism based on Lin-former. The LWSA module applies low-rank projections to reduce attention complexity from O(n²) to O(n), enabling efficient long-sequence learning while preserving global dependencies. Experi-ments were conducted using a full-year dataset (35,040 records at 15-minute intervals) from the Mahuangshan First Wind Farm in Ningxia, China. The model was tested under various input win-dow lengths (1h, 3h, 12h, 24h, and 32h). Results show that CNN-BiLSTM-LWSA consistently out-performs CNN-BiLSTM and CNN-BiLSTM-Attention in both accuracy and efficiency. Under a 24-hour input, it achieves an RMSE of 53.4 kW, MAE of 23.2 kW, and R<sup>2</sup> of 0.955 while reducing training and testing time by 54.8% and 47.1%, respectively, compared to the attention-based base-line. Even with a 32-hour input, the model maintains low prediction errors and stable R<sup>2</sup>, validating its scalability. The experimental results fully confirm that CNN-BiLSTM-LWSA effectively balances forecasting accuracy and computational cost across different temporal settings, offering a robust, efficient, and practical solution for short-term wind power forecasting applications.
Some of the metrics are blocked by your
consent settings
PSD-EEGRepNet: A CNN Architecture with Multi-Branch RepBlocks for Power Spectral Density-Based Motor Imagery EEG Classification in BCI
(2025-01-01)
Thangthong, Kantham
;
Asadi, Fawad
;
Tungjitkusolmun, Supan
Motor Imagery (MI) based Brain-Computer Interfaces (BCIs) utilizing electroencephalography (EEG) offer significant potential, yet progress can be hindered by the computational demands of deep learning classifiers. This study introduces and evaluates a novel lightweight, multi-branch Convolutional Neural Network (CNN), inspired by efficient design principles, specifically for classifying MI tasks from Power Spectral Density (PSD) EEG features. Our objective was to achieve strong classification performance coupled with favorable development phase computational characteristics. Evaluated on 10 subjects from a public PhysioNet dataset using 5-fold cross-validation and two data overlap conditions (80%, 90%), the proposed model (~6.32M parameters, ~20 MFLOPs) demonstrated high mean classification accuracies (80.38% for 80% overlap, 85.69% for 90% overlap) and efficient training times (avg. 8.8s and 20.4s per fold, respectively). While performance scaled positively with data augmentation, inter-subject variability was noted. We conclude that the proposed architecture effectively balances high accuracy with significant offline computational efficiency offering a valuable tool for BCI research and a promising foundation for developing practical MI-BCI systems.
Some of the metrics are blocked by your
consent settings
Deep Learning to Classify Bacterial Species in the same Genus
(2024-01-01)
Sheela, Sherin
;
Piang, May Phu
;
Sakorntanant, Sakda
;
Poomrittigul, Suvit
;
Treebupachatsakul, Treesukon
Bacterial strains in the same genus share highly similar morphology, gram-staining characteristics, colony sizes, and spatial arrangements. Therefore, identifying them by deep learning can be quite challenging. This study aimed to assess the classification of 7 species of bacteria from 2 genera of Bacillus and Vibrio by using 8 Convolutional Neural Network (CNN) models. We implemented Python programming along with Keras API within the Jupyter Notebook. The models were constructed and evaluated under unbalanced and balanced datasets by augmentation (rotation, flip, etc.). Transfer learning with fine-tuning, and pre-processing of mixup and label smoothing were also applied to reduce overfitting and enhance generalization. Based on the experimental results on private dataset, the results of InceptionResNetV2 emerged as the top-performing model with a notable accuracy of 82.8%, 88.6% precision, 78.4% recall, and 78.0% F1-score when label smoothing was applied at 0.5 on balanced dataset.
Some of the metrics are blocked by your
consent settings
Classification of Depression Audio Data by Deep Learning
(2022-01-01)
Homsiang, Phanomkorn
;
Treebupachatsakul, Treesukon
;
Kiatrungrit, Komsan
;
Poomrittigul, Suvit
Due to many factors such as anxiety from contracting the disease and concern about the socioeconomic impacts, Thai people have accumulated stress and are at risk of depression. The diagnosis of depression can be primarily assessed by testing the assessments such as PHQ8, PHQ-9, and CES-D. The applied deep learning technology in medicine has received research interest and has been developing. In this research, we tried the classification of depression and non-depression audio datasets with the implementation of 4 model architectures: 1D CNN, 2D CNN, LSTM, and GRU. By converting wave audio format (WAV) of Daic-woz database to the Melfrequency cepstrum (MFC). We have done the training and evaluated the 4 model architectures and compared the results between non-augmented and augmented datasets. The highest accuracy was obtained from 1D CNN with a non-data augmentation of 95%, and a 2D CNN with a data augmentation of 75%. These results confirm that human voices can differentiate between depression and non-depression.
Some of the metrics are blocked by your
consent settings
Vision-based attitude estimation for spacecraft docking operation through deep learning algorithm
(2020-02-01)
Phisannupawong, Thaweerath
;
Kamsing, Patcharin
;
Tortceka, Peerapong
;
Yooyen, Soemsak
On-orbital services, especially in docking operation and other space object interaction. The missions need accurate, reliable, and robust detection to be an accurate in joining any interaction concerned. Two spacecrafts with an unknown mathematical model to predict the position and orientation, a computer vision-based attitude estimation system to detect the poses of spacecraft via camera is the key option of the mission. In astronautics control, the position coordinates are normally represented as the cartesian coordinate system and used a quaternions coordinate system for orientation representation because quaternions can represent the orientation of spacecraft better than physical angle and can overcome the problem of singularity. This paper aims to construct a model for both position and orientation estimation with public data. The input images are the dataset of Soyuz in the resolution of 1280x960, which is simulated by Unreal Engine 4. The implementation of this paper use GoogLeNet for a convolutional neural network model with the mathematical model of loss subject to direct regression. The result shows that a position estimation is significantly accurate with having distance error smaller than 1 meter and trand to reduce when setting a proper scaling factor for loss function. The result demonstrates a high error for orientation estimation. However, the experiment expresses that both position and orientation estimation can be improved in case of selecting a suitable scaling factor of loss function.

KMITL

Browse

Filters

Settings

Sort By

Results per page

Search Results