To tackle inaccurate and expensive forecast services, Microsoft and TempoQuest collaborate with AceCAST to accelerate wind energy forecasts. The new platform is more precise and timely than earlier CPU-based models, delivering renewable power generation predictions in megawatts on an hourly basis.

For proper forecasts of renewable energy production companies require accurate weather modeling. This further helps in preparing a better plan of action to deal with natural disasters. In a recent GTC session, Microsoft and TempoQuest mentioned and detailed their work with NVIDIA.

The project is to address energy and climate issues. It is the need of the hour because non-forecasted and ineffective weather forecasts already cost about $714 billion in 2022.

To tackle this, companies need cost-effective but fast and accurate weather forecast models. In the trio, TempoQuest enables hyper-local low-latency environmental and weather forecasts.

It is a member of the NVIDIA Inception Program. The company said, “Our team is multidisciplinary, covering atmospheric science, meteorology, HPC, AI, ML, engineering, and more. We have been a leading adopter of bringing GPUs to the environmental sector.”

Other things quoted about the company include the following:

  • It is the first to develop GPU software in a way that can be used as a service weather forecast system.
  • First company to port Weather Research and Forecasting (WRF) to GPUs
  • Also, the first one is to create fast and high-resolution forecasts that are cheaper than CPU-based forecasts.

The company benefitted from NVIDIA on Microsoft Azure to move traditional Weather Research and Forecasting software to GPUs. Also, deliver resolutions less than kilometer and time resolutions of 1-minute to 1-hour. With this it also enables faster predictions of power generated by solar and wind resources.

Challenges faced by Utilities to Integrate Renewables

Environmental factors like wind speed and cloud coverage influence wind and solar power. That makes it challenging to maintain the grids primarily with wind and solar power.

On days of insufficient power generation from renewable resources, utility companies use spinning reserves, that is, carbon-based power produced by generators. With better, faster, and accurate weather forecasting it will be easier to predict renewable power generation.

To incorporate more renewable to grid the utility companies have to build higher voltage transmission lines and towers. But more importantly, new generation sites and all this will increase operational and capital costs.

Accelerating Weather Research and Forecasting with GPUs

AceCAST or Accelerated Forecast is a result of running WRF which is used by 50,000 users in 160 countries. WRF is ported to run on x86 systems with NVIDIA GPUs.

It is using proprietary OpenACC and CUDA, and it is scaled on multi-node and mullet-GPU systems. All major WRF name list options, physics scheme, and dynamics are supported in AceCAST.

Plus, it is a drop-in replacement for all existing WRF configurations. Faster solutions, great accuracy, higher resolution, reduced computation costs, and great localized awareness of weather phenomena are several benefits of AceCAST.

Cost Analysis of AceCAST Validation and Performance

CPU WRF to GPU WRF differences is checked whether they are within an acceptable tolerance range. Model performance was tested across spatial forecast and several temporal ranges.

Lastly, thousands of test cases were validated to ensure the results produced by AceCAST are the same as CPU WRF. Large differences in performance and cost were revealed after running performance tests on Microsoft Azure. Take a look to the

1. CPU-based WRF – Standard HB120rs_v3 VMs (HBv3):

  • 120 AMD EPYCâ„¢ 7V73X-series (Milan-X) CPU cores
  • 450 GB RAM (350 GB/sec memory bandwidth)
  • 200 Gb/sec HDR InfiniBand
  • 2 x 1 TB NVME SSD disks
  • NCAR WRF 4.2.2
  • Uses Parallel net-CDF
  • Compiled with Intel Compilers and MPI

2. GPU-accelerated WRF – Standard_ND96amsr_A100_v4 (NDmv4):

  • 8 NVIDIA A100 Tensor Core GPUs (80GB)
  • NVLink 3.0 (200 Gb/s HDR InfiniBand)
  • 96 AMD EPYCâ„¢ 7V12-series (Rome) CPU cores
  • 8 x 1 TB NVME SSD disks
  • AceCAST 2.1
  • Proprietary implementation using OpenACC and CUDA
  • Scales on multi-node and multi-GPU using MPI

3. Azure Managed Lustre File System

  • 40TiB Storage Azure Managed Capacity
  • 10000 MB/s max throughput

As per the acquired results, AceCAST achieved ~9x acceleration than CPU-based WRF. Results achieved from 18 CPU nodes are similar to the ones received from 1 GPU node.

These results show that utilities can accurately predict renewable power generation. With this excessive power outages could be avoided, and reliable power can be delivered.

For another test on AceCAST 3.0.1, a nested domain with outer domain was used. The outer domain was 5 million grid points (430x331x38v) with 15-kilometer grid spacing, and inner domain was 80 million grid points (1551x1361x38v) with 3-kilometer grid spacing.

From the results obtained it was concluded that AceCAST runs 16.8x faster than WRF inner domain. Also, it runs about 7% faster and at 75% lower cost in comparison to CPU-based WRF. It means utilities will accurately do power prediction in megawatts at specific sites on an hourly basis every day.

Microsoft and TempoQuest collaborate with AceCAST to accelerate wind energy forecasts. With this, there will be a major societal and global change.

Source: NVIDIA


Elliot is a passionate environmentalist and blogger who has dedicated his life to spreading awareness about conservation, green energy, and renewable energy. With a background in environmental science, he has a deep understanding of the issues facing our planet and is committed to educating others on how they can make a difference.

Leave A Reply