Umeå University's logo

Accurate Cloud Behavior Prediction (CBP), also referred to as forecasting in this context, is essential for Solar Power Prediction (SPP), as well as for weather forecasting, climate analysis, and satellite imaging. However, the nonlinear and dynamic nature of clouds, combined with other limitations, presents significant challenges to advancing CBP. Recent developments, particularly the integration of Machine Learning (ML), Numerical Weather Prediction (NWP), and other innovative approaches, show strong potential for improving CBP and, in turn, enhancing SPP and related applications. This review presents a bibliometric analysis of 467 publications from 1970 to 2024, retrieved from the Scopus database using CBP-related keywords. It identifies trends, influential studies, major subject areas, leading authors, contributing countries, and key publishers. The study further categorizes the essential steps in CBP and provides a detailed review of the most relevant literature on cloud cover, cloud motion (including vector-based methods), and cloud image prediction. Additionally, it examines critical factors affecting model performance and introduces a framework for evaluating predictive methods based on input types, methodologies, prediction horizons, results, and evaluation metrics. Several key challenges are highlighted, including the nonlinearity of cloud behavior, limited data availability, image quality issues, and model accuracy. In response, actionable recommendations are offered, such as expanding data sources, applying hybrid imaging and modeling approaches, managing uncertainty, improving postprocessing techniques, and incorporating cloud content estimation. Given the relatively limited research in this field, this study serves as a valuable benchmark for researchers, engineers, and policymakers engaged in real-time SPP and other cloud-dependent domains.

This study explores strategies to maximize salinity gradient power (SGP) generation using reverse electrodialysis (RED), focusing on key operating parameters under Swedish environmental conditions. Herein, using a full-factorial experimental design, seawater salinity, flow velocities, and water temperature is varied across three levels to assess their impact on SGP output. machine learning methods predict power density (PD), including 1) ensemble learning with decision tree (DT), 2) gaussian process regression (GPR), and 3) artificial neural network (ANN). Fivefold cross-validation confirms the ANN's high accuracy (root mean squared error (RMSE): 1.173%, R2: 99.35%), closely followed by GPR (RMSE: 1.95%, R2: 99.17%). A feature and trend pattern analysis among the input factors reveals sea salinity as the primary influence on PD, with temperature as the secondary contributor. Complementing this, a life cycle assessment examines the environmental impact of RED systems, identifying the Seawater River RED and brine-wastewater treatment plant RED systems as having environmental effects, particularly on ozone layer depletion and freshwater toxicity. Carbon fiber-based (CF) electrodes, especially lignin CF, demonstrate a lower impact, yet concerns remain over key sustainability challenges. These findings highlight SGP's potential as a viable renewable source, highlighting areas for future material selection and system efficiency improvements.

Ground fault relays (GFRs) in resonant-grounded medium voltage distribution networks shall not operate during phase-to-ground (Ph-G) fault inception, allowing the Petersen coil to suppress self-extinguishing faults, but the designated GFR must operate during permanent faults. In order to enhance the performance of GFRs, particularly during high-impedance faults, the scope of this paper is to propose a straightforward, machine-learning-based protection framework. The enhanced GFR is modeled as a classification task. Depending on the GFR's position and the Ph-G fault location in the network, fault samples are labeled as “no operation,” “primary,” “backup,” or “backup of backup,” forming two-class, three-class, and four-class GFR setups, respectively. This assures selective operation across three protection zones and improves the reliability of all GFRs. The proposed protection scheme employs backward optimal feature selection to identify the most relevant discrete features obtained from measured zero-sequence current and voltage waveforms. An ensemble of k-nearest neighbor classifiers is utilized for accurate classification, simulating the GFR operating conditions, with measurement errors and sensitivity incorporated in the preprocessing. A 20 kV case study network validates the proposed framework, achieving F1-scores exceeding 96 %. The maximum operation delay of the protection scheme for an enhanced GFR is 225 ms, accommodating the required time window (200 ms), prediction time (5 ms), and change detection time (20 ms), thus assuring safe operation. Compared to other machine-learning-based methods used for Ph-G fault protection in resonant-grounded radial networks, this framework is high-performing, fast, and easy to implement, utilizing a simpler structure than neural networks.

Electromobility contributes to decreasing environmental pollution and fossil fuel dependence, as well as increasing the integration of renewable energy resources. The increasing interest in using electric vehicles (EVs), enhanced by machine learning (ML) algorithms for intelligent automation, has reduced the reliance on. This shift has created an interdependence between power, automatically, and transportation networks, adding complexity to their management and scheduling. Moreover, due to complex charging infrastructures, such as variations in power supply, efficiency, driver behaviors, charging demand, and electricity price, advanced techniques should be applied to predict a wide range of variables in EV performance. As the adoption of EVs continues to accelerate, the integration of ML and especially deep learning (DL) algorithms will play a pivotal role in shaping the future of sustainable transportation. This paper provides a mini review of the ML impacts on mobility electrification. The applications of ML are evaluated in various aspects of e-mobility, including battery management, range prediction, charging infrastructure optimization, autonomous driving, energy management, predictive maintenance, traffic management, vehicle-to-grid (V2G), and fleet management. The main advantages and challenges of models in the years 2013–2024 have been represented for all mentioned applications. Also, all new trends for future work and the strengths and weaknesses of ML models in various aspects of mobility transportation are covered. By discussing and reviewing research papers in this field, it is revealed that leveraging ML models can accelerate the transition to electric mobility, leading to cleaner, safer, and more sustainable transportation systems. This paper states that the dependence on big data for training, the high uncertainty of parameters affecting the performance of electric vehicles, and cybersecurity are the main challenges of ML in the e-mobility sector.

This paper proposes a computationally light adaptivefiltering approach, normalized least mean squares (NLMS), to model the nonlinearity caused by the current transformer (CT) iron core saturation. A simplified CT model was used to generate adataset considering four different nonlinear iron core magnetic characteristics. The preliminary results show satisfactory results in the cases where the CT iron core nonlinearity is within certain limits.

Power-quality standards provide limited guidance on frequency quality for short time scales, such as less than one hour. Capturing frequency variations and events requires high time resolutions, e.g., 0.1 seconds or less, resulting in significant data storage requirements. However, power-quality monitors typically report averaged values at intervals of 10 minutes, 15 minutes, or one hour, depending on the disturbance type. To address these challenges, we propose calculating statistic indices from high-resolution data within a 15-minute or one-hour window length, thus avoiding storing high-resolution data. We apply the basic statistic indices to frequency data measured in Finland for a single day in June 2023, demonstrating their effectiveness in capturing frequency variations and two significant events during that day.

The paper investigates the application of machine learning (ML) for voltage sag source localization (VSSL) in electrical power systems. To overcome feature-selection challenges for traditional ML methods and provide more meaningful sequential features for deep learning methods, the paper proposes three time-sample-based feature forms, and evaluates an existing feature form. The effectiveness of these feature forms is assessed using k-means clustering with k = 2 referred to as downstream and upstream classes, according to the direction of voltage sag origins. Through extensive voltage sag simulations, including noises in a regional electrical power network, k-means identifies a sequence involving the multiplication of positive-sequence current magnitude with the sine of its angle as the most prominent feature form. The study develops further traditional ML methods such as decision trees (DT), support vector machine (SVM), random forest (RF), k-nearest neighbor (KNN), an ensemble learning (EL), and a designed one-dimensional convolutional neural network (1D-CNN). The results found that the combination of 1D-CNN or SVM with the most prominent feature achieved the highest accuracies of 99.37% and 99.13%, respectively, with acceptable/fast prediction times, enhancing VSSL. The exceptional performance of the CNN was also approved by field measurements in a real power network. However, selecting the best ML methods for deployment requires a trade-off between accuracy and real-time implementation requirements. The research findings benefit network operators, large factory owners, and renewable energy park producers. They enable preventive maintenance, reduce equipment downtime/damage in industry and electrical power systems, mitigate financial losses, and facilitate the assignment of power-quality penalties to responsible parties.

This paper addresses the lack of consideration of short time scales, below one hour, such as sub-15-min and sub-1-hr, in grid codes for frequency quality analysis. These time scales are becoming increasingly important due to the flexible market-based operation of power systems as well as the rising penetration of renewable energy sources and battery energy storage systems. For this, firstly, a set of frequency-quality indices is considered, complementing established statistical indices commonly used in power-quality standards. These indices provide valuable insights for quantifying variations, events, fluctuations, and outliers specific to the discussed time scales. Among all the implemented indices, the proposed indices are based on over/under frequency events (6 indices), fast frequency rise/drop events (6 indices), and summation of positive and negative peaks (1 index), of which the 5 with the lowest thresholds are identified as the most dominant. Secondly, k-means and k-medoids clustering methods in a learning scheme are employed to identify typical patterns within the discussed time windows, in which the number of clusters is determined based on prior knowledge linked to reality. In order to clarify the frequency variations and patterns, three frequency case studies are analyzed: case 1 (sub-15-min scale, 10-s values, 6 months), case 2 (sub-1-hr scale, 10-s values, 6 months), and case 3 (sub-1-hr, 3-min values, the year 2021). Results obtained from the indices and learning methods demonstrate a full picture of the information within the windows. The maximum value of the highest frequency value minus the lowest one over the windows is about 0.35 Hz for cases 1 and 2 and 0.25 Hz for case 3. Over-frequency values (with a typical 0.1% threshold) slightly dominates under-frequency values in cases 1 and 2, while the opposite is observed in case 3. Medium fluctuations occur in 35% of windows for cases 1 and 2 and 41% for case 3. Outlier values are detected using the quartile method in 70% of windows for case 2, surpassing the other two cases. About six or seven typical patterns are also extracted using the presented learning scheme, revealing the frequency trends within the short time windows. The proposed approaches offer a simpler alternative than tracking frequency single values and also capture more comprehensive information than existing approaches that analyze the aggregated frequency values at the end of the specific time windows without considering the frequency trends. In this way, the network operators have the possibility to monitor the frequency quality and trends within short time scales using the most dominant indices and typical patterns.

Understanding the influence of teaching culture (tradition) within academic departments is crucial for new teachers navigating the complex landscape of higher education. This paper investigates the possible impact of the department’s teaching culture on the pedagogical approaches of new teachers, forming their teaching style, concentrating on insights gathered from interviews with experienced colleagues in a Swedish university department with a focus on engineering education. By exploring the department’s teaching traditions and identifying potential challenges faced by new teachers, this study offers valuable insights into enhancing teaching styles and fostering student engagement. Drawing upon both experiential knowledge and insights from pedagogic literature and courses, the authors provide practical strategies to overcome obstacles and promote operative teaching practices. Ultimately, the outcomes of this study aim to empower new teachers to create enriching learning environments that promote student motivation, engagement, and overall academic success, aligning with the findings of existing literature on pedagogy and student learning outcomes.