Between hype and reality
Artificial intelligence holds great potential to revolutionize battery monitoring. However, many solutions still fail due to a fundamental issue: they rely on generic models that don’t reflect the complexity of real-world energy storage systems. This kind of “standard AI” fails to account for the fact that real batteries differ significantly depending on cell chemistry, configuration, sensor setup, and usage scenarios.
One common rookie mistake is developing a cell model in the lab and assuming it will behave the same in the field. Another is working with insufficient data – for example, only evaluating the extreme values of a few cells or using raw, unverified measurement data.
At NOVUM, we don’t rely on standard AI. We intentionally adapt and extend it to reflect real-world conditions and the specific requirements of each battery system. That’s the only way to generate truly reliable results.
Typical mistakes when using AI in battery analytics
The following examples show which mistakes are particularly common in practice. New providers in particular often underestimate the complexity of real battery systems — with consequences for the validity of their analyses.
Here are typical mistakes that we at NOVUM repeatedly observe and consciously avoid:
1. Taking lab results as the gold standard
A common misconception is to assume that lab-based insights translate directly into real-world operation. But real battery systems never behave exactly like lab cells – even with the same chemistry. Usage scenarios, wiring, sensor placement, and environmental factors all significantly influence cell behavior. That’s why we treat lab results as just one of many inputs. Our AI models are always based on the specific system in its actual operating environment.
2. Capturing too little data
Many manufacturers only provide data from the best and worst-performing cell in a module – typically to reduce the load on battery controllers. But that’s far too little for reliable monitoring: which cell is the “worst” or “best” can change constantly during operation. That’s why we always request individual cell-level data. It’s more effort, but it enables us to develop solutions that are both practical and economically viable.
3. Blindly trusting the data provided
AI is only as good as the data it’s built on. Yet many monitoring providers take current, voltage, or temperature readings at face value. We don’t. At NOVUM, battery scientists and electrical engineers jointly review the data. They check sensor positioning, synchronization, and signal quality – and correct issues directly in the software if necessary. This ensures clean data and trustworthy results.
4. Treating all batteries the same
Cells with the same chemistry can behave very differently depending on how they’re configured, cooled, and used. Despite this, such systems are often lumped together into one training dataset. Our solution: we create a dedicated model for every installation – from the individual cell up to the entire system. This allows us to analyze each level precisely and carry out detailed plausibility checks.
5. Failing to question AI results
AI is often treated as a black box – its results taken at face value. But that can be risky: it may lead to functioning modules being discarded or critical conditions going undetected. That’s why we use an additional “guardian AI” that flags unusual patterns to our human experts. They investigate anomalies, interpret the results, and – if needed – physically inspect the modules to verify their hypotheses. This turns the black box into a white box: transparent and verifiable.
What NOVUM has learned in over 10 years of practice
Our experience shapes not only our analytics but our entire workflow – from the first data set to the final recommendation. We know that AI only delivers reliable results when it’s built on a solid foundation of data and is continuously questioned. That’s why we build an individual model for each battery system, tailored to all its specific characteristics – from chemistry and sensor setup to its real-world use case.
Thorough preparation: We review technical documentation, datasheets, layouts, and available data points to understand every relevant parameter. This is the foundation for building a model that truly fits the system.
Ensuring data quality: We establish a live connection to the storage system and verify data quality. Are timestamps aligned? What’s the resolution? How reliable are the sensors? Only clean data leads to meaningful AI insights.
Custom modeling: Our models adapt to the chemistry in use. LFP cells, for instance, require highly precise state-of-charge estimation, while NMC cells demand fast-reacting safety monitoring. Our models are fine-tuned accordingly.
Control mechanisms: Our AI can recognize when it’s operating in uncertain territory – for example, when encountering behavior it hasn’t seen enough times before. In such cases, it triggers an alert to our battery experts, who intervene as needed.
Validation through practical experience: We regularly validate our results with hands-on inspections. If our AI flags a module as critical, we open it up and investigate the cause – be it self-discharge, lithium plating, or a manufacturing defect. This knowledge flows continuously back into our models.
A clear difference
Many monitoring providers follow the principle: “more data equals more intelligence.” While that may sound reasonable, it’s only half the truth. They recognize correlations but miss causal relationships. Proper interpretation, critical questioning, and contextualization are essential. For example: if you group multiple battery systems with the same chemistry into one model, you might miss how drastically their behavior differs due to cooling or wiring. At NOVUM, we create custom analytics models for every individual installation – right down to the module and cell level.
Our AI is not a black box. We make it a white box: transparent, controlled, and reliable. We achieve this by combining cutting-edge algorithms with deep battery expertise – and the courage to challenge conventional thinking.
