Energy Digest

X-OPM introduces a feature engineering framework using tree-based models to capture feature interactions, complemented by linear models for prediction, and incorporates human-in-the-loop workflow to balance model accuracy against modeling effort. It achieves $R^2 > 0.93$ across all workloads with low sampling window size, outperforming state-of-the-art methods. X-OPM incurs an area overhead below $0.1\%$, comparable to lightweight tree-based and linear models.

Probabilistic machine-learning forecasts using published German transmission records outperformed baseline models in terms of accuracy and reliability. Raw LightGBM achieved the highest normalized weighted interval score (nWIS) with 25% improvement over the seasonal baseline, while rolling calibration improved its performance further. Boosted-tree models, such as LightGBM and Transformer, showed accurate probabilistic forecasts under target delays, but nominal aggregate validity did not ensure reliability during extreme congestion events.

Gaussian process-based estimators are used to extend complex regression techniques to closed-loop multiple-input single-output dynamic systems. These estimators can identify individual transfer functions in such systems by analyzing analytical conditions and converging to true values in the zero noise limit. The approach offers a nonparametric alternative to existing methods, while providing probabilistic characterization of estimation uncertainty.

The proposed framework learns common demand behavior through a shared model while adapting only a compact subset of parameters to improve local accuracy. It uses separate adaptation components for each load to preserve shared knowledge across customers while providing flexibility for heterogeneous and mixed load compositions. The approach shows consistent improvements in deterministic accuracy and probabilistic quality over various baselines with low storage and inference costs.

Battery energy storage systems must respect strict state-of-charge limits due to measurement noise and nonlinear dynamics, but traditional data-enabled predictive control methods can fail when faced with abrupt changes or uncertainties. Scenario-based data-enabled predictive control (Scenario-DeePC) improves constraint robustness by predicting future behavior directly from observed prediction errors rather than assuming a disturbance distribution, resulting in comparable tracking performance with fewer constraint violations. Scenario-DeePC achieves this with an adaptive scenario buffer that tightens constraint handling automatically to improve robustness against unpredictable events.

The proposed mixed-integer linear programming model predictive control framework optimizes heat pump demand flexibility with grid-scale battery energy storage systems and voltage-regulation devices, reducing peak demand by up to 23.1% and overheating issues in distribution networks. The framework also dispatches electric-resistance heaters for preheating, adding an additional 2.7 percentage points of peak reduction. By co-optimizing these elements, the proposed method can accommodate high penetration rates of heat pumps with minimal impact on grid operation.