Photovoltaic digital twin for AI-assisted solar fleet operations
- Customer
- AXS Energia
- Project manager on the customer side
-
- IT Provider
- InterSystems
- Year of project completion
- 2026
- Project timeline
- october, 2025 — september, 2026
- Project scope
- 30000 subscribers
- Goals
- AXS Energia is a Brazilian energytech founded in 2021 with no legacy systems and one premise: treat data as core business infrastructure from day one. The company sells solar energy by subscription and operates its own distributed generation fleet, delivering one to two new plants per month. The goal was to grow at this pace without losing control: unify plant telemetry, IoT sensors, SCADA, weather stations, ERP, CRM, finance, maintenance and connections to more than ten utility distributors in a single operational layer; move decisions from individual experience to history, patterns and integrated data; and solve a sector-specific problem, the cold start: machine-learning forecasting models need about three years of history, and a fast-growing fleet always has new plants without it.
- Project Results
- Published 2025 indicators: OPEX per installed MWp down 38% (reflecting Orion and scale gains), mean time to diagnosis down 32%, MTTF up 27%, and above 92% accuracy in generation forecasting; forecasting results (92% and 81.68%) were measured jointly with InterSystems. Customer and environmental impact in 2025: about R$ 144 million in direct savings for customers and about 83,000 tonnes of CO2 avoided, serving more than 33,000 customers, growing toward 50,000, with waiting lists for new plants. Scale: 72 plants in operation across five Brazilian states, heading to 130 plants and 335 MWp, with one to two new plants delivered per month. Net revenue grew from R$ 53 million to R$ 100 million in one year (88% growth, independently audited). The case was presented at InterSystems READY 2026 (Washington, DC), at Intersolar South America 2026 and at a data & AI roundtable at the Gartner CIO & IT Executive Conference 2026.
The uniqueness of the project
Orion is not an AI layer bolted on later: AXS was born without legacy systems, and Orion is the direct consequence of treating data as infrastructure since day one. The intelligence layer runs on InterSystems IRIS with a data fabric architecture and four analysis pillars (descriptive, diagnostic, relational and predictive). Vector search relates current events to operational history, automatically identifying similar anomalies and standardizing O&M responses. The signature innovation is a physics-based photovoltaic digital twin (pvlib, embedded Python inside IRIS) that writes, with the laws of physics, the history a new plant does not yet have: in a proof of concept validated against 208 days of real generation, it reached 81.68% baseline forecasting accuracy on the plant's first day, rising above 92% as real data accumulates. On top of this foundation, Ask ORION brings generative AI that answers from Orion's own data, not from model assumptions.- Used software
- InterSystems IRIS as the data backbone of Orion: data fabric, real-time integration and vector search. Machine-learning models for generation forecasting (Random Forest and regression). Physics-based photovoltaic digital twin (pvlib) running as embedded Python inside IRIS, generating synthetic generation time series. Generative AI grounded in Orion's data (Ask ORION). Integration with SCADA, IoT sensors, weather stations, ERP, CRM and the systems of more than ten utility distributors; comparison of real generation against PVsyst simulations.
- Difficulty of implementation
-
Five main challenges:
- real-time integration of heterogeneous sources: plant telemetry, IoT, SCADA, weather stations, ERP, CRM, finance, maintenance and more than ten utility distributors, each with its own formats;
- scalability: onboarding new plants, distributors and systems without restructuring the platform, in a portfolio heading to 130 plants and 335 MWp;
- the cold start of predictive models: ML needs about three years of history that new plants do not have, solved with the physics-based digital twin;
- moving decisions from individual experience to history, patterns and integrated data, with standardized O&M responses;
- Southern Hemisphere physics in the digital twin: plant orientation, direct versus diffuse light separation, and the mismatch between temperature-efficiency rules and real module behavior.
- Project Description
- Initial situation. Founded in 2021 with no legacy systems, AXS decided to treat data as business infrastructure from day one; I joined as the company's first employee and built the technology organization from zero. The challenge was turning that premise into a real operation in a business that never stops growing: every new plant brings telemetry, IoT, SCADA and weather stations; every new region brings another utility distributor (AXS connects to more than ten); and ERP, CRM, finance and maintenance must talk to plant operations. Without a single foundation, each expansion would multiply silos, rework and cost. The journey. 2021, architecture decision: after evaluating market alternatives, AXS chose InterSystems IRIS for scalability, flexibility, cost efficiency and interoperability. 2022, production: the first users were O&M and the engineering team that qualifies new plants by comparing real generation against PVsyst simulations; each new plant and distributor is onboarded without restructuring the platform. 2025, intelligence layer: Orion organized into four pillars, with vector search relating current events to history and ML models forecasting generation. 2026, digital twin proof of concept: to solve the cold start of new plants, AXS and BPlus built a physics-based digital twin that generates the history a plant does not yet have; validated against 208 days of real generation, it reached 81.68% baseline accuracy on day one. Today, Ask ORION brings generative AI to operational data, and Orion is evolving into the data fabric layer of the whole company. Implementation specifics. The hardest parts were the real-time integration of heterogeneous sources at fleet scale; an architecture that absorbs new plants, distributors and systems without restructuring; and Southern Hemisphere physics in the digital twin: plant orientation, separating direct and diffuse light, and the mismatch between temperature-efficiency rules and real module behavior. Changes after implementation. Decisions that depended on individual experience now rely on history, patterns and integrated data; similar anomalies are identified automatically and O&M responses are standardized. The portfolio expands without complexity and cost growing in the same proportion.
- Project geography
- Brazil. 72 solar plants in operation across five states (MG, PR, SP, MT and GO), all operated through a single platform; more than 33,000 customers nationwide; the technology organization is based in Florianopolis, Santa Catarina.
Events