Journal of Electrical and Power System Engineering

The Race for Power in the AI Era: A Systematic Review of Time-to-Power Constraints, Data-Centre Location, and Sustainable AI Infrastructure

Praveen — 2026-10-14

Artificial intelligence (AI), generative AI, large language models, and accelerated computing are redefining the physical demands of digital architecture. The feasibility of deploying large-scale AI infrastructure is not just determined by computational efficiency, but even more so by the availability and local timing of electrical capacity to energize the systems that power them: whether nearby land has access to abundant deep-well water, which is increasingly attracting interest; what sources of low-carbon electricity are available, when they are available, and whether they can be delivered in time; what cooling resources are available across regions; and where high-bandwidth, low-latency network connectivity exists are also being examined. This review brings together literature on AI compute growth, electricity demand from data centers, data-centre siting and span of control in grids, renewable-energy procurement and scheduling, carbon-aware computing and cooling, and water use. Evidence published primarily from 2020 to 2026 is synthesized using a structured systematic-review approach, consistent with PRISMA 2020 principles, while earlier work laying the foundations of this evidence base remains referenced. The evidence base consists of empirical studies, engineering analyses and modelling work, industry reports and systems research that cannot be compared statistically through meta-analysis; hence, thematic synthesis is used. The central concept developed in this review is Time-to-Power (TTP), defined as the elapsed period between a sufficiently advanced infrastructure commitment (site control, interconnection request or equivalent planning milestone) and the availability of the required continuous electrical capacity at the point of connection. TTP is treated as a conceptual deployment constraint rather than an established industry-standard metric. The review further demonstrates that AI data-centre location is becoming a multi-resource decision involving grid capacity, interconnection timing, land, fiber, renewable resources, tariffs, climate, cooling options, and water availability. Sustainability requires temporal and spatial consideration of carbon intensity, renewable matching, workload flexibility, and both direct and indirect water use. These findings are organized into the proposed AI Infrastructure Readiness and Sustainability Framework (AIRSF), comprising five layers: compute demand, power readiness, location, sustainability, and deployment decision criteria. AIRSF is not presented as an empirically validated optimizer; it provides a research and decision architecture that can be operationalized with real project, grid, workload, cost, and environmental datasets. Future priorities include standardized TTP definitions, spatial-temporal workload orchestration, grid-interactive AI data centers, integrated carbon–water–power optimization and empirical AIRSF validation.

Active Filtering-Based Decoupling of Fluctuating Power in Single-Phase Converters

Shaikh Zubair A. R. — 2026-08-19

Single-phase AC/DC and DC/AC converters are widely used in renewable energy systems, electric vehicle chargers, uninterruptible power supplies, and distributed generation applications. Despite their widespread adoption, these converters inherently experience low-frequency power pulsations caused by the double-line-frequency ripple power arising from the instantaneous mismatch between the AC-side and DC-side power. Conventionally, this ripple is mitigated by incorporating bulky electrolytic capacitors in the DC-link. Although passive filtering effectively suppresses voltage fluctuations, it increases the converter size, reduces power density, and limits system reliability due to the relatively short lifetime and ageing characteristics of electrolytic capacitors. To address these limitations, active power decoupling (APD) techniques have gained significant attention as an effective alternative. APD diverts the low-frequency ripple power to auxiliary energy storage elements, thereby reducing the dependence on large DC-link capacitors and improving the overall performance, efficiency, and lifetime of the converter. However, most existing APD topologies require additional passive components, such as inductors or large film capacitors, together with auxiliary switching devices. These extra components increase circuit complexity, component count, installation space, and overall system cost, limiting their practical implementation in compact power electronic systems. The proposed approach employs a symmetrical half-bridge active power decoupling circuit that provides effective ripple power compensation with a reduced number of components. The topology utilizes only two low-value capacitors and two active switches to absorb and regulate the fluctuating power, eliminating the need for bulky energy storage elements. As a result, the proposed design offers a compact, cost-effective, and reliable solution with improved power density, enhanced efficiency, and extended converter lifetime, making it a promising choice for next-generation single-phase power conversion applications.

A Hybrid Ensemble Learning Framework for Accurate Battery Remaining Useful Life Prediction in Electric Vehicles

Ravisankar Kampara — 2026-10-15

Reliable calculation of the remaining usable life (RUL) of lithium-ion batteries is vital for boosting the reliability, operating safety, and maintenance preparation of electric vehicles (EVs). Since battery degradation is highly nonlinear and is affected by a number of operating parameters, estimating the RUL presents a difficult problem. The paper puts forward an explainable hybrid machine learning approach that combines feature selection, Bayesian hyperparameter optimization, and Extreme Gradient Boosting (XGBoost) to achieve accurate prediction of the battery’s RUL. First, the battery data are preprocessed, and the most relevant health indicators are found through correlation-based screening, mutual information analysis, and Recursive Feature Elimination (RFE) to reduce feature redundancy and enhance model performance. Bayesian optimization is then used to find the best XGBoost hyperparameters, thus improving prediction accuracy while at the same time reducing overfitting. The framework is assessed against Random Forest, Decision Tree, K-Nearest Neighbours, and Naïve Bayes by means of standard performance measures like MAE, MSE, RMSE, and R². The results show that the optimized XGBoost model offers better prediction accuracy and interpretability, and as such can be considered a practical solution for intelligent Battery Management Systems in next-generation EVs.

STATCOM-Based Reactive Power Compensation in Distribution Networks: A Review of Sizing, Control, and Renewable Integration Literature

Tejaskumar Ashokrao Rewatkar — 2026-09-19

Distribution feeders were not built to handle power flowing both ways. Rooftop solar has changed that. When solar panels send power back up the line, voltage can rise too high. Old capacitor banks cannot fix this fast enough because they only switch on and off in steps. This paper looks at a device called STATCOM, and its smaller version, D-STATCOM, which can push or pull reactive power almost instantly. It works much faster than old methods. This paper reviews research on four things. First, where to place STATCOM devices and how big they should be. Second, how their controllers are built. Third, how they work together with solar and wind power. Fourth, how they compare to an older device called SVC. Sizing methods have changed over time. Older studies used simple trial-and-error search. Newer studies use mathematical methods that guarantee a better answer. Controller design has also changed, moving from simple PI controllers to smarter fuzzy and AI-based controllers. But these three research areas barely talk to each other. Also, almost every study on sizing tests its method on a made-up IEEE test network, not a real power line. This paper also looks at real loss numbers from Indian power companies, both nationally and in Maharashtra state. This shows how big this research gap really is. At the end, the paper suggests ways to connect sizing, control, and real utility data better than before.