Testing Trustworthy AI for Green Hydrogen: Insights from CARTIF’s Hybrid Energy System

When an Algorithm Meets an Electrolyser  

In a research facility in Valladolid, Spain, a 500 NL/h electrolyser is doing something unusual. Rather than operating on fixed schedules or rule-based controllers, it's being guided by Artificial Intelligence (AI) — systems trained to anticipate faults, predict maintenance needs, and optimise performance in real time, against real renewable energy inputs. 

This work is part of a broader European effort under the Digital Europe Programme, where AI Testing and Experimentation Facilities (TEFs) are helping bridge the gap between research and real-world deployment by providing environments to rigorously test AI in operational conditions. 

Deploying AI in real-world energy systems, especially in critical infrastructures such as hydrogen systems, requires more than promising algorithms. It demands robust validation, transparency and trust. 

This is not a simulation. And that distinction matters more than it might seem. 

CARTIF's hybrid hydrogen facility is one of the experimentation environments at the heart of EnergyGuard, a pan-European project building a network to test and validate AI solutions for the energy sector.  

The goal is straightforward but demanding: move AI from promising concept to deployable, trustworthy technology — in environments that reflect the complexity of actual energy systems. 

 

The Problem with Electrolysers in the Real World  

Green hydrogen depends on electrolysers, and electrolysers depend on renewable energy. That's where things get complicated. 

Renewable generation is inherently intermittent. Wind and solar inputs fluctuate, sometimes rapidly. Anion Exchange Membrane (AEM) electrolysers — well-suited for decentralised hydrogen production — have electrochemical constraints that introduce inertia and limit how quickly they can respond.  

Frequent start-stop cycles accelerate degradation. And the conventional fix — adding batteries or power electronics to buffer the variability — increases both system complexity and cost.  

Instead of relying solely on additional hardware, AI offers a smarter alternative: adapting the operation of the system to changing conditions, working with its constraints rather than around them to reduce complexity and improve performance. 

CARTIF’s Hybrid Hydrogen System: A Platform Built for Real-World Complexity  

CARTIF's facility was designed with this in mind. It's a hybrid energy platform where hydrogen doesn't operate in isolation — it interacts with electrical and thermal systems in a dynamic, integrated way.  

The installation includes: 

  • An AEM electrolyser with a production capacity of 500 NL/h  
  • A PEM fuel cell for electricity and heat generation  
  • A metal hydride hydrogen storage system (1.6 kg capacity)  
  • Electrical storage (batteries) and thermal storage (300 L tank + heat pump)  
  • A simulation environment for renewable generation and demand profiles  

Additionally, the control architecture is PLC-based and designed for seamless AI integration, while real-time data acquisition enables researchers to capture how the system truly behaves — from inefficiencies and response delays to real operational constraints — going beyond theoretical models to understand its performance in practice. 
 

3 Areas of AI Experimentation  

Within the EnergyGuard framework, CARTIF is using this facility to validate AI-driven approaches for electrolysis systems, focusing on 3 main areas.  

1. AI-Based Diagnostics and Fault Detection 

One of the primary goals is to detect operational issues early by analysing system data in real time. AI models are trained to detect abnormal operating conditions, communication failures, and deviations from expected performance in real time — catching issues early in systems where components interact in complex ways and faults aren't always immediately visible.  


2. Prognostics and Predictive Maintenance 

Beyond detecting faults, AI is used to anticipate future failures and degradation. By analysing historical and real-time data, the system can identify optimal time windows (e.g. during low renewable production). This approach enables a shift from reactive to predictive maintenance, reducing downtime and operational costs. 

 
3. AI-Driven Control and System Optimisation 

AI-based controllers coordinate the electrolyser, storage systems, and renewable inputs to maximise renewable energy use, minimise efficiency losses, and avoid harmful operating patterns like rapid ramping or frequent cycling. By integrating forecasts of both generation and demand, the system can operate in a genuinely adaptive way.   

Why This Matters Beyond Valladolid  

The immediate benefits are measurable: extended system lifetime, lower maintenance costs, improved energy efficiency, and reduced hardware requirements.  

These improvements contribute directly to reducing the Levelised Cost of Hydrogen (LCoH), a key barrier to large-scale adoption.  

But the broader value lies in replicability. The insights from CARTIF's facility — the sizing criteria, the validated control strategies, the lessons from real-world operation — are designed to be extrapolated.  

The point isn't just to prove AI works in one lab in Spain; it's to generate the evidence base that enables other installations across Europe to follow with confidence. 

That confidence will extend to trust. Deploying AI in critical energy infrastructure raises legitimate questions about transparency, accountability, and alignment with EU regulatory frameworks, including the AI Act. 

The work at CARTIF addresses this directly by validating AI on real operational data, ensuring decisions are traceable, and developing control strategies that are interpretable — not just effective. 

 

The Bigger Picture: Ensuring Trustworthy AI in Energy Applications 

As Europe's energy system becomes more decentralised and more dependent on variable renewables, hydrogen's role as a flexibility enabler will only grow.  

What EnergyGuard — and facilities like CARTIF's — are building is the foundation of trust that large-scale deployment will require. 

AI can make electrolysis systems smarter. Real-world experimentation is what makes that claim credible. And credibility, ultimately, is what accelerates adoption. 

To take a deep dive into how EnergyGuard is testing and validating AI solutions for the energy sector, please visit https://energy-guard.eu 

This article was produced by José L. Hernández, Senior Researcher at CARTIF. 

Tag
Energy TEF AI Innovation