Metenergy Solutions - high-resolution modeling of meteorological variables optimized with AI, applied to the energy market

• Grant number: 24/04530-9
• Support Opportunities: Research Grants - Innovative Research in Small Business - PIPE
• Start date: November 01, 2024
• End date: October 31, 2025
• Field of knowledge: Physical Sciences and Mathematics - Geosciences - Meteorology
• Agreement: SEBRAE-SP
• Principal Investigator: Felipe Tartaro Pereira
• Grantee: Felipe Tartaro Pereira
• Company: METENERGY SOFTWARES LTDA
• CNAE: Desenvolvimento e licenciamento de programas de computador customizáveis ; Pesquisa e desenvolvimento experimental em ciências físicas e naturais
• Associated researchers: Lucas Tartaro Pereira ; Vinicius Akira Imaizumi

Abstract

We are a team originating from the Mesoscale Meteorology Laboratory (LMM) at the University of São Paulo (USP) - IAG and we participated in the NIDUS innovation residency at the USP Innovation Center (INOVA USP). Currently, we have the participation of Prof. Dr. Ricardo Hallak from LMM-USP and a qualified team, composed of the imminent graduate in Meteorology Felipe Tartaro, master's student in Meteorology Lucas Tartaro and Electrical Engineer Vinícius Akira, with skills in numerical modeling, artificial intelligence, software development and cloud architecture. Our tutor professor has already had experience with the job market working in numerical modeling at the companies SOMAR and CPTEC, totaling more than 4 years of good relationship with the strategic sectors of meteorology. Our technology-based startup seeks solutions to the problems of energy waste and high risks in investment and planning for agents in the current energy market, caused by errors associated with estimates in energy pricing, arising from erroneous predictions of precipitation and temperature of future scenarios.Our solution is based on the generation of prediction and analysis products from meteorological data simulated with the WRF numerical model, customized in nested grids of high spatial and temporal resolution and refined using AI in a scheme called Physics-Informed Neural Networks (PINN , from the English Physics Informed Neural Network). Thus, we obtain optimization of parameterizations from several atmospheric and oceanic models recognized by NOAA, such as ERA5, ORA5 and ECMWF. From this refined data, we are particularly interested in creating forecast models to operate in regions with wind, hydroelectric and solar plants, as well as forecasts of the load required by the National Interconnected System (SIN), mainly dependent on temperature and rainfall.We aim to optimize the preparation of energy contracts drawn up by generating and trading agents, improve the predictability of the load generated by the plants, as well as monitoring and forecasting the Difference Settlement Price (PLD) and the Marginal Operating Cost (CMO ), making the free energy market in Brazil less risky for agents and more sustainable. Initially, we will focus on the main river basins where the largest generation of the Brazilian energy matrix is located and, later, the project will extend to wind and solar farms. We developed a product that has already proven efficient for other study groups. We followed the lecture by POLI-USP professor, Anna Reali Costa, who leads a research group where PINN coupling was able to minimize the prediction errors of the model used (SOFS) by around 10% in very high resolution in-channel modeling. of Santos, in São Paulo.Our business model has a pioneering perspective in the Brazilian energy sector, mainly because it uses the Amazon Web Services (AWS) cloud architecture. The current market invests in its own architecture, which demands costs with maintenance, updating and acquisition of computers. AWS already has specific tools for running simulations via WRF in the cloud, as is the case with services such as AWS ParallelCluster, Elastic Fabric Adapter and NICE DCV.We will use two value generation models. The first, to meet scalability demand and enter the market with a more generalized product, we will deliver a standardized, high-value product in subscription form. Therefore, we can guarantee 24 hours of product availability for standard configuration for each of the main regions available. The second, in the "pay for use" format, will be aimed at customers who opt for a solution more applied to their area of activity, with pre-defined options that increase the processing cost. (AU)