Energy Digest

Analytical Power-Aware Provisioning for Prefill-Decode Disaggregated AI Inference develops an analytical framework that jointly considers serving capacity and power consumption in large-scale inference serving, providing limited analytical insight into how provisioning decisions shape the tradeoff between serving capacity and power consumption. The framework models the serving capacity and average power consumption of a provisioned deployment based on hardware constraints, workload characteristics, and KV-cache reservations. It determines the serving capacity--power Pareto front among candidate provisioned deployments to enable service providers to choose an optimal provisioned deployment based on changing workloads or available power.

Offline reinforcement learning is used for line-selective tripping in distribution grids, achieving high precision and recall rates of 0.9993 and 0.9496, respectively, with a combined input and CQL weight of α=0.9. The model selects the correct line-trip action correctly in 98.13% of fault episodes but also trips incorrectly in 72.73% of non-fault episodes. The results demonstrate strong faulted-line selection on simulated fault episodes using conservative Q-learning (CQL).

Grid-forming frequency shaping control can shape post-contingency aggregate system dynamics into ideal first-order systems with prescribed rate of change of frequency and steady-state frequency deviation, reducing transient control effort by allocating resources to achieve coherent dynamics. The squared $\mathcal{H}_2$ norm is used as a quantification of transient control cost under step power imbalances through system transformation. A constrained optimization problem can be solved using successive convexification method to optimize resource allocation.

CE-ESMs used for optimizing Nordic energy systems fail to capture grid stability requirements, resulting in 1.1% to 17.4% contingencies with voltage violations depending on grid code and automation requirements. Grid regions with low generation are particularly vulnerable to voltage insecurity. Implementing power control measures such as Power Park Module voltage control can improve voltage security in the modeled future system.

The article proposes an equivalent modeling procedure for load-side systems with distributed renewable generation that considers fault responses and nodal voltage coherence, leading to improved power responses across various operating conditions and fault types, particularly under severe unbalanced faults. The method generates a reusable physical electromagnetic transient equivalent from representative operating conditions and uses it to analyze transmission-system dynamic behavior. This approach outperforms direct capacity aggregation in predicting terminal responses under different scenarios.

Spatial coherence in renewable forecast scenarios has negligible operational value for single-period economic dispatch, with a maximum gain of 0.64% of dispatch cost. Decision-focused training is more effective, delivering gains of 2.82-5.19%. A parametric Gaussian-copula approximation may be useful at realistic error magnitudes but performs poorly under extreme stress conditions.

GraphToolbox is an open-source Python framework for graph neural network (GNN) forecasting, unifying graph construction, model selection, training, and interpretation in a single configuration-driven pipeline. The framework offers data-driven graph construction, automatic model instantiation, online expert aggregation, and significance testing on cached forecasts, improving accuracy over classical baselines. Evaluations using French regional load and net-load case studies show improved performance with GraphToolbox, especially when aggregating multiple graphs.

Artificial representative trees (ARTs) combined with leaf-wise Mondrian conformal predictive systems (CPS) provide compact, structurally stable trees with substantially more reproducible split-variable selection. This combination balances predictive performance with interpretability and stability, offering a single model for continuous predictions and calibrated probabilities. ARTs with CPS outperform decision trees in terms of reproducibility and stability.

Remote state estimation under unreliable communication introduces an information asymmetry between a smart sensor and a remote estimator due to imperfect state reconstruction by the sensor. The remote estimator's internal state evolves under packet reception, leading to uncertainty captured through a belief conditioned on its information set. A recursive update process is derived to analyze this asymmetry, resulting in a tractable finite Gaussian mixture representation with linearly growing components over time.

Reinforcement learning has emerged as a complementary approach to traditional Operational Research (OR) methodologies, offering strong learning and computational capabilities for sequential decision-making in dynamic and uncertain environments. The integration of RL with OR aims to leverage its learning capabilities to strengthen traditional algorithms, improving solution quality, computational efficiency, and robustness. A systematic review highlights the potential applications and challenges of integrating RL with OR for solving sequential decision-making problems, combinatorial optimization, and extended reality analysis.