Researchers introduce GraphToolbox, an open-source Python framework that unifies graph construction, model training, aggregation, and interpretation for GNN-based electricity forecasting in one configuration-driven pipeline built on PyTorch Geometric. It supports 51 of 65 PyTorch Geometric convolutions and includes case studies on French regional load and net-load forecasting, showing where graph models beat classical baselines and where they fall short.
An arXiv paper evaluates peak-aware short-term load forecasting at three distribution grid aggregation levels (area codes, secondary substations, LV feeders) using UK and Swiss open datasets. Chronos-2 delivers the best high-demand performance at every level, cutting mean HD-NMAE by 20-51% versus LightGBM and XGBoost, with fast enough inference for practical deployment.
An arXiv paper evaluates peak-aware short-term load forecasting at three distribution grid aggregation levels (area codes, secondary substations, LV feeders) using UK and Swiss open datasets. Chronos-2 delivers the best high-demand performance at every level, cutting mean HD-NMAE by 20-51% versus LightGBM and XGBoost, with fast enough inference for practical deployment.
A new arXiv paper benchmarks nine modern deep learning models for time-series forecasting on real-world smart meter data, addressing a gap in comparative performance studies. The work evaluates linear, MLP-based, and other architectures for short- and long-term energy forecasting, with implications for operational costs and grid management.
A new arXiv paper benchmarks nine modern deep learning models for time-series forecasting on real-world smart meter data, addressing a gap in comparative performance studies. The work evaluates linear, MLP-based, and other architectures for short- and long-term energy forecasting, with implications for operational costs and grid management.
An arXiv paper introduces physics-aware random walk fingerprints for classifying cascading failures on power-grid graphs, targeting benchmarks like PowerGraph. The approach aims to offer an alternative to GNNs, which need end-to-end training and model-specific tuning and yield latent representations that are hard to tie to physical propagation patterns.
An arXiv paper introduces physics-aware random walk fingerprints for classifying cascading failures on power-grid graphs, targeting benchmarks like PowerGraph. The approach aims to offer an alternative to GNNs, which need end-to-end training and model-specific tuning and yield latent representations that are hard to tie to physical propagation patterns.
A new arXiv paper audits common spatiotemporal forecasting benchmarks, finding that GNN methods are often evaluated on a narrow set of datasets and that simple baselines frequently match or beat them. The authors call for more rigorous evaluation protocols.
A new arXiv paper audits common spatiotemporal forecasting benchmarks, finding that GNN methods are often evaluated on a narrow set of datasets and that simple baselines frequently match or beat them. The authors call for more rigorous evaluation protocols.
A new arXiv study systematically evaluates TabPFN-TS, a zero-shot time-series foundation model, for probabilistic heat load forecasting in district heating networks. The research compares TabPFN-TS against other time-series foundation models and trained machine learning baselines, finding it a promising alternative that adapts at inference time without retraining.
A new arXiv study systematically evaluates TabPFN-TS, a zero-shot time-series foundation model, for probabilistic heat load forecasting in district heating networks. The research compares TabPFN-TS against other time-series foundation models and trained machine learning baselines, finding it a promising alternative that adapts at inference time without retraining.