Skip to content
World clockEU--:--UK--:--USA--:--CN--:--PLDEFRIT中文EN

portal about AI and technologyevents · analysis · interviews · technical background

Search
LIVE
›

Grid software turns to space weather forecasts and energy-aware GPU sharing

Microsoft Research says a machine learning pipeline can flag space-weather risk at 66,935 US substations 30 to 60 minutes before impact. A SOSP '26 paper reports 1.4-2.3x energy savings from smarter GPU sharing.

TechnologyNewsRachel NwosuPublished: 30 September 20264 min readSources 10
Grid software turns to space weather forecasts and energy-aware GPU sharing
Microsoft Research says a machine learning pipeline can flag space-weather risk at 66,935 US substations 30 to

Two pieces of grid-adjacent software research surfaced this week. Both target the same problem: the infrastructure behind power and compute must do more with less warning and less energy.

The newest dated item comes from Microsoft Research. It published a blog post on 30 September describing a forecasting pipeline for space weather risk on power grids.

Space weather, substation by substation

According to Microsoft Research, the system uses forecast-time solar-wind information. It produces location-specific risk estimates for 66,935 substations in the continental United States. It combines forecasts of the Auroral Electrojet (AE) and Disturbance Storm Time (Dst) indices with local latitude, geology and ground conductivity. The blog says the pipeline detected nearly 80% of major space-weather events during the evaluation period. It can warn grid operators 30 to 60 minutes before a specific risk appears.

The post frames the problem in operational terms. Knowing a storm is coming is not the same as knowing when and where its effects will be worst. Geomagnetically induced currents (GICs) depend on local conditions, according to the same post. Regions with resistive bedrock can see stronger currents than regions with more conductive geology. The May 2024 geomagnetic storm, which pushed auroras far beyond their usual range, is cited as the motivating example.

The work was done during a summer internship at Microsoft Research, the blog says. The author describes it as an end-to-end machine learning pipeline rather than a physics simulation. That distinction matters for grid operators. They need a forecast they can act on, not a research model that needs interpretation.

Energy, not just utilization, in GPU clusters

A second paper, published at SOSP '26 on 28 September, takes aim at the other side of the energy problem. The authors are Prasoon Sinha, Dimitrios Liakopoulos, Nathan Lemma and Neeraja J. Yadwadkar. In "Beyond Utilization: Energy-Conscious GPU Sharing for Inference Serving", they argue that optimizing only for GPU utilization can counterintuitively increase energy consumption. Their system, EnerTune, uses analytical models to estimate per-model performance and power. It also estimates the power draw of colocated models on shared GPUs. It feeds them into an energy-aware bin-packing algorithm.

The paper reports that EnerTune meets performance SLOs while reducing energy consumption by 1.4-2.3x. It cuts power draw by 1.3-2.6x over state-of-the-art baselines. The authors note that profiling every possible configuration at scale is prohibitively expensive. That is why the analytical models matter. Each model can be deployed under hundreds of configurations, and profiling itself costs energy.

Both items land in the same week as a flurry of grid software activity. Energy Digital Magazine published a list of top 10 smart grid providers on 30 September. Sustainable Views argued the same day that Europe must build and operate its grid as one system. Utilities Middle East ran a Siemens piece on freeing up grid capacity, also on 30 September.

On the storage side, Energy Vault secured New South Wales state development approval for its 125 MW / 1 GWh Stoney Creek BESS, according to Business Wire on 29 September. That clears the path for construction. Energy Source & Distribution reported the same approval on 30 September. The same week, Energy-Storage.News published a piece on why availability matters as a BESS performance metric and how to calculate it. It also reported that SMA's battery inverter became the first to win grid-forming certification for the German inertia market.

Forecasting and flexibility

Forecasting is not only a space-weather problem. pv magazine Global reported on 29 September on a digital map that forecasts Spain's grid constraints and curtailment hotspots for 2027. The same day, Business Wire reported that Critical Loop delivered flexible interconnection for California industry. It cited the first power controller to achieve CSIP 2030.5 certification and UL 3141 listing.

On 29 September, Spectra by MHI published an explainer on how virtual power plants can future-proof grids. Business Wire reported that Willow is putting autonomous agents to work across Fortune 100 building portfolios. MarketScale reported on an IFS survey finding that only 33% of utility executives call their asset management advanced. That number sits awkwardly next to the automation claims.

The common thread is that grid software is moving from dashboards to decisions. Microsoft's pipeline does not just show a storm. It estimates risk per substation with a lead time. EnerTune does not just pack models onto GPUs. It trades off energy against latency and throughput. Whether operators trust either system enough to act on its output is a separate question. Neither item addresses it directly.

Comments 0

Sources

10
  1. 01Forecasting space weather risks on power gridsEN
  2. 02Beyond Utilization: Energy-Conscious GPU Sharing for Inference ServingEN
  3. 03CHOMPI portable sampler instrument is now open-sourceEN
  4. 04The open-source AI platforms vying to become China's Hugging FaceEN
  5. 05Report: Kakao is shutting down Tapas webtoon platformEN
  6. 06Justice Department and FBI Seize Platforms Operated and Used by China State-Sponsored HackersEN
  7. 07Japan's fusion energy startup to begin power-on tests at pilot reactor in 2027EN
  8. 08Dear Software MakersEN
  9. 09OpenClaw Enterprise – The Open Agent PlatformEN
  10. 10Estimating software projects without fooling yourselfEN

All figures and quotations in this text come from the sources listed below.

Content prepared by the editorial team with AI assistance.

Rachel Nwosu

Rachel Nwosu

AI, models and technology

Rachel Nwosu covers AI, models and technology for FLASH24, working from public model documentation, benchmark releases and repository histories rather than press summaries, and she skips announcements that arrive without reproducible numbers. She checks training-data claims against dataset cards and reruns reported metrics where code is available. She spends much of her week interviewing researchers and engineers, tracking model launch calendars, and comparing vendor benchmarks with independent evaluations. Outside the desk she runs 3D printers, restores old computers, and tests how models learn from internet junk. She does not publish benchmark figures she cannot trace to a source.

Newsroom →

Comments

0
  1. No comments yet — be the first.

Write a comment

Comments are public. We do not publish abuse, spam or advertising.