HyunYong Lee

Accurate prediction of Remaining Useful Life (RUL) is critical for effective prognostics and health management (PHM) in industrial systems. However, existing data-driven methods often require large amounts of high-quality data and lack interpretability, limiting their applicability in real-world scenarios. In practice, widely used synthetic benchmark datasets such as CMAPSS and N-CMAPSS are themselves noisy due to sensor drift and mode switching, and limited in the number of available run-to-failure trajectories. In this study, we propose ST-KDFormer (Spatio-Temporal Koopman Dual-branch Transformer), a novel architecture that integrates Koopman operator theory with deep learning to address these challenges. ST-KDFormer employs separate Koopman Autoencoders (KAEs) to learn linearized latent representations of temporal and spatial system dynamics. These embeddings are jointly processed through a dual-branch Transformer encoder to capture contextual dependencies within and across temporal windows and sensor dimensions. To enable efficient end-to-end training, we introduce a unified loss function with adaptive weighting to balance embedding and RUL prediction objectives. We evaluate our approach on the CMAPSS and N-CMAPSS benchmark datasets through extensive experiments, including ablation studies and comparisons with state-of-the-art methods. The results demonstrate that ST-KDFormer achieves superior performance in terms of prediction accuracy and robustness across varying operational conditions, validating its effectiveness for practical RUL estimation tasks.

With the growing penetration of renewable energy sources, ensuring the stability and reliability of Medium-Voltage Direct Current (MVDC) systems has become more critical than ever. A single fault in MVDC systems can cause significant disturbances, necessitating rapid and precise diagnostics to prevent equipment damage and maintain continuous power supply. In this work, we present a Bidirectional Gated Recurrent Unit (Bi-GRU) model that both classifies and locates MVDC faults. By capturing the temporal behavior of voltage signals, the Bi-GRU framework surpasses traditional algorithms such as Convolutional Neural Networks (CNNs) and Bidirectional Long Short-Term Memory (Bi-LSTM) networks. Furthermore, the proposed approach addresses multiple fault scenarios including PTP (Pole-to-Pole), PPTG (Positive Pole-to-Ground), and NPTG (Negative Pole-to-Ground) while preserving real-time diagnostic capabilities. In extensive tests, our model achieves an overall accuracy of 95.54% and an average fault detection time below 1.3 ms, meeting real-world operational requirements. To assess robustness, sensor noise was artificially introduced to emulate realistic conditions. Despite these challenging inputs, our method consistently maintained high diagnostic accuracy, confirming its practicality and reliability. Consequently, the proposed scheme demonstrates a significant contribution toward improving the safety and dependability of MVDC systems, even under noisy conditions.

In this paper, we propose methods for the diagnosis of down conductor cables installed inside blades for the reliable operation of wind turbines. Utilizing the two synchronized drones equipped with X-ray devices to take X-ray images inside the blades, we propose deep learning methods for diagnosis based on X-ray images. The first method is the unsupervised deblurring method for blurry X-ray images that are generated by the drones. The key idea of the proposed method is to apply an image sharpening process to the latent space of the trained model using only blurry X-ray images. Through experiments using the X-ray images of blades, we show that the proposed method makes the outline of the down conductor cable clearer. As the second method, we propose an object detection-based method to provide fast and accurate object detection as the diagnosis. Given an X-ray image, we also propose a method to determine drone flight direction to enable drones to follow the down conductor cable. Through experiments, we show that our method has good performance in object detection (i.e., mAP of 98.21%) and classification (i.e., AUC of 0.9), and drone flight direction determination (i.e., AUC of 0.99).

A reliable performance loss rate of photovoltaic systems requires accurate and reliable performance metrics. This study proposes a systematic method for assessing the performance metrics, particularly predicted power models in terms of both accuracy and uncertainty. The gist of the proposed method is to examine how accurately a predicted power model predicts the manipulated degradation in a controlled environment. For this, the proposed method divides a given dataset evenly into base data (to generate reference performance) and test data (to generate test performance via manipulation) so that the two data have similar features. The proposed method also utilizes the bootstrap iteration to derive a reliable assessment. The novelty of this study is that the proposed method estimates both the accuracy and uncertainty of arbitrary predicted power models, which is missing in existing works. Extensive experiments using the proposed method with real‐world datasets reveal the followings. One interesting observation is that a well‐known machine learning prediction model, not considered in existing works, exhibits the best performance in terms of both accuracy and uncertainty. Existing predicted power models require different experiment settings to produce reliable performance. The number of test data is closely related to uncertainty, but not much related to accuracy.

Demand response (DR) is a voluntary program that encourages related stakeholders, in this case electricity consumers, to cut down on usage during periods of high electricity load. One key to fully exploiting DR is to encourage residential customers to join the DR program. Unlike in the DR programs for commercial and industrial customers, for the successful operation of the residential DR program, several issues have to be addressed, one of which is to provide a group-level incentive to participating customers. In particular, the issue comes up when the incentive calculated for a group is not equal to the aggregated incentives for each customer (i.e., non-equal incentive problem). The non-equal incentive problem deteriorates the successful operation of residential DR by decreasing the motivation of DR operators and customers. We first prove the non-equal incentive problem through mathematical and experimental methods. We then propose the novel single group-based indirect incentive calculation method. The basic idea of our approach is to indirectly calculate the incentive for each customer not using the customer’s data but using the data of other customers of the same DR group. Through experiments involving the electricity usage data of 42,193 households and the real DR events in Korea, we show that our method solves the non-equal incentive problem in most cases. Furthermore, our method improves the accuracy of the baseline estimation (used for calculating the contribution).

A reliable forecasting model is required for photovoltaic (PV) power output because solar energy is highly volatile. Another driver for the need of a reliable forecasting model is concept drift, which means that the statistical properties of the data change over time. In this paper, an online forecasting method to handle concept drift is proposed. First, the problem of forecasting in batch learning is transformed into a forecasting in online learning setting. Then, an online learning algorithm is applied, which is good for handling concept drift. Through experiments using the real‐world data, it is shown that the method noticeably improves performance compared to the case where a trained model is used. Under various concept drift scenarios, the method improves performance by up to 87.3%. It is also shown that the re‐training method (a representative existing method) has several limitations. This method requires several issues to be solved, such as selection of a proper window size, and this is evident through results showing different performance under different settings. In contrast, the method shows a reliable and desirable performance under various concept drift scenarios and thus outperforms the re‐training method. The method improves performance by up to 79%.

In realizing unsupervised pixel-precise anomaly localization by utilizing a generative model, a reference image must be generated (for comparison with an input image) by transforming abnormal patterns of an input image, if any, into normal patterns. In this study, a patch-level operation with adaptive patch control is proposed to improve anomaly localization by generating a better reference image. As a way to exploit a generative model, we divide an image into non-overlapped patches of the same size, generate patch-level reference images, and stitch the patch-level reference images into a single reference image. We then conduct anomaly localization by comparing an input image with the stitched, reconstructed image. To effectively apply the patch-level operation, we propose adaptive patch control to determine the number of non-overlapped patches to be applied. For this, we synthesize defective images using normal images and examine how well the candidate methods with different numbers of patches remove the synthesized defects. In the same way, we utilize adaptive patch control to select a promising model among the candidate generative models. Based on experiments conducted using the MVTec Anomaly Detection dataset, we demonstrate that our method outperforms previous existing methods. Under a real-world scenario, our method shows ROC AUC of 0.926, in contrast to the best value of 0.893 from existing studies. Furthermore, we prove the feasibility of the adaptive patch control by showing that the removal of the synthesized defects and the anomaly localization for real defective images are highly correlated.

In this study, the authors are interested in estimating how much a PV system underperforms than expected by exploiting forecast uncertainty. For this, they first study a forecast accuracy‐related forecast uncertainty metric using the ensemble method based on the dropout technique, which is widely used in deep learning forecasting models. Given the forecast accuracy‐related uncertainty metric, the rationale of the authors' approach is that forecast accuracy is likely to decrease compared to the normal case of similar uncertainty metric values if any performance degradation happens. It is because similar uncertainty metric values are likely to show similar forecast accuracy. Therefore, they generate a standard table by simulating possible performance degradation cases and conduct the performance degradation diagnosis by looking up the standard table based on the uncertainty metric. From the experiments, in the case of persistent degradation, they show that their approach estimates the performance degradation with the estimation error of around 1% while an uncertainty‐unaware approach shows the estimation error of up to 5%. In the case of temporal degradation, their approach shows the estimation error of around 3%, while the uncertainty‐unaware approach does not show meaningful result.

A forecast interval is effective for handling the forecast uncertainty in solar photovoltaic systems. In estimating the forecast interval, most available approaches apply an identical policy to all the point forecasts. This results in an inefficient interval (e.g. an unnecessarily wide interval for an accurate forecast). They also adopt a complex model and even require modification of the available deterministic forecasting model, which may adversely affect their application. To overcome these limitations, the authors introduce a forecast uncertainty‐aware forecast interval. They calculate a forecast accuracy‐related uncertainty metric from an ensemble method based on the dropout technique. The dropout technique is widely used in deep learning models. This implies that the proposed approach can be applied to available deep learning forecasting models without modifying them. Using the uncertainty metric and relevant data of previous forecast results, they estimate the uncertainty‐aware forecast interval. Through experiments using real‐world data, they first demonstrate the close relation of their uncertainty metric to the forecasting accuracy. Then, they demonstrate that the uncertainty‐aware forecast interval reduces the mean interval length by up to 25.7% and decreases the prediction interval coverage probability by 4.07%, compared to available approaches. This illustrates that their approach results in an effective interval.

For better deep learning forecasting systems for photovoltaic systems, confidence information about a point forecast is necessary in practical cases where uncertainties are unavoidable. In this study, using Bayesian deep learning, the authors introduce a confidence‐aware deep learning forecasting system that provides confidence information as well as a point forecast. Through the experiments using the real‐world data, they first solve three main issues caused by when Bayesian deep learning is applied to the forecasting of daily solar irradiance using weather forecast: selection of neural network model, selection of validation data to be used for estimating the confidence information, and ways for estimating the confidence information. Then, they examine the feasibility of the confidence‐aware deep learning forecasting system in estimating the confidence information. From the experiments, classifying the forecast outputs into confident outputs and non‐confident outputs using the confidence information, they show that maximum absolute percentage error of confident forecast outputs and non‐confident forecast outputs are 5 and 22.8% at a specific classification threshold, respectively. This result shows that their confidence‐aware deep learning forecasting system is good to estimate meaningful confidence information that is closely related to the forecast accuracy.