Energy Digest

Tempered Christoffel-weighted polynomial chaos expansion is a method for uncertainty quantification in power systems, which balances numerical stability and tail fidelity. It reduces 95th percentile deviation by 16% and improves regression stability index by over 130%, with controlling weighting intensity directly influencing both stability and tail prediction accuracy. The proposed method outperforms traditional sparse polynomial chaos expansion methods in terms of distributional accuracy.

The proposed hybrid knowledge-data-driven approach leverages collaboration between a large language model (LLM) agent and a reinforcement learning (RL) agent to achieve two-stage voltage control in active distribution networks, incorporating day-ahead forecasts and semantic-based grid codes. The LLM agent generates scheduling strategies for OLTC and SCs at the region level, while the RL agent refines terminal voltages with reactive power generation strategies for PV inverters based on accurate node-level measurements. This approach enhances training efficiency and improves voltage control performance by effectively utilizing the inherent knowledge and reasoning capabilities of both agents.

A novel three-stage diagnosis-driven co-planning (DDCP) framework is proposed to optimize network reinforcement and battery energy storage system installations for distribution grids with high penetration of electric vehicles. The framework diagnoses critical bottleneck lines, upgrades cables exclusively at these lines, and then optimizes BESS deployment using a network-enhanced model. The DDCP framework achieves techno-economic superiority in addressing high-EV-penetration challenges.

Selecting the right deep learning model for power grid forecasting is challenging and depends heavily on the available data. The paper benchmarks five models, including state space models, Transformers, and a traditional LSTM, across six US power grids with varying forecast windows, revealing that there is no single best model for all situations. Models perform differently depending on the task, such as patchTST excelling on solar generation forecasting and state space models performing better on wind and wholesale prices forecasting.

This paper proposes a federated cross-client interdependency learning methodology for decentralized root cause analysis in nonlinear dynamical systems, allowing for the incorporation of diverse feature spaces and proprietary client models without requiring access to raw sensor streams or model modification. The approach uses machine learning models augmented with cross-client interdependencies to learn representation consistency while preserving privacy through calibrated differential privacy noise. The method is validated on extensive simulations and a real-world industrial cybersecurity dataset.