The increasing penetration of renewable energy sources has introduced significant variability into modern power systems, making accurate short-term forecasting essential for reliable planning and operation. Among renewable resources, wind energy is particularly challenging due to its strong intermittency and nonlinear temporal behavior. This study presents a comparative evaluation of classical statistical and deep learning models for short-term wind speed forecasting, including Moving Average (MA), Simple Exponential Smoothing (SES), Double Exponential Smoothing (DES), Triple Exponential Smoothing (TES), Autoregressive Integrated Moving Average (ARIMA), and Long Short-Term Memory (LSTM). The models are assessed using a real-world wind speed dataset from California covering the years 2021--2022 at a 5-minute resolution. To ensure a fair comparison, model hyperparameters are tuned using a rolling-window validation framework, after which final performance is reported on a 30\% hold-out test set. In addition, robustness is investigated under different additive white Gaussian noise (AWGN) levels. The results show that LSTM achieves the best overall forecasting accuracy, while ARIMA remains highly competitive and robust under noisy conditions, which shows the importance of modeling temporal dependence and noise resilience for reliable short-term wind speed forecasting.

Data centres need low carbon electricity, but they cannot compromise backup power. This paper presents a sizing method for photovoltaic generation, battery storage, grid supply, and an existing diesel generator. The diesel generator remains in the system and supplies only the outage energy that the photovoltaic and battery system cannot cover. The method tests 49 designs with the same Monte Carlo outage events. For a five megawatt data centre in Abu Dhabi, the lowest cost design that meets the illustrative outage autonomy target uses twenty megawatts of photovoltaic capacity and forty megawatt hours of battery capacity. It covers 45.8 percent of sampled outages without diesel and reduces diesel served outage energy by 68.2 percent. Sensitivity checks show that the results depend on weather, outage mix, battery power, tariffs, and battery life.

Battery management systems must select charging protocols in real time while keeping operation within safe limits, yet current systems enforce fixed, non-adaptive constraints that sacrifice performance over the battery lifetime. We propose an Adaptive Safety Certificate (ASC) framework that defines state-dependent safety constraints derived from public aging data and a gradient-boosting risk classifier, tightening automatically as battery health degrades. Integrated into a safe online learning policy, the framework progressively eliminates empirically unsafe protocols as state of health decreases, concentrating selections on protocols that balance charging speed with long-term sustainability. In simulation over 2,000 charging batches, the ASC achieves 41% longer expected battery life than a random baseline, corresponding to $3,528 in deferred replacement cost per vehicle and 2,578 tonnes of avoided CO2 per 1,000-vehicle fleet. Ablation studies con- firm that state-of-health-dependent adaptation is the single most important property of the framework, with direct implications for sustainable fleet management policy.

The reliability of power grid transformers is a cornerstone of socio-economic stability in the Cauca Department, Colombia, where diverse geographical conditions and intense ceraunic activity pose significant risks to critical infrastructure. Traditional predictive maintenance models, while accurate, often suffer from high computational overhead and latency, limiting their deployment in edge-computing environments. This paper introduces a Hyperdimensional Regression (HDR) framework as a computationally agile, energy-efficient alternative to conventional ensemble learners for power systems regression. Leveraging a comprehensive dataset of 15,873 transformers, we employ 2019 operational data for model training and 2020 data for rigorous temporal validation, ensuring practical applicability in predicting future infrastructure states. Experimental results demonstrate that while ensemble methods such as XGBoost yield high peak accuracy, the proposed HDR framework achieves robust predictive performance with $R^2$ scores reaching $65.5\%$ at a fraction of the computational cost. Notably, the HDR models maintain sub-second, deterministic inference latencies, proving to be approximately $17\times$ faster than high-performing baselines like XGBoost. This study establishes HDR as a superior high-fidelity, low-latency alternative for real-time fault detection in power distribution networks, providing utility providers with a scalable tool for optimizing maintenance schedules in both rural and urban grids.