The 92% Barrier: How Power Engineers Are Learning to Speak AI—On Their Own Terms

The Barrier That Nobody Talks About

In a survey of 52 power-systems researchers and engineers, 92% reported hitting at least one wall before they could even run a simple AI model on their data. Not during training. Not during deployment. Before they ran it. The obstacles ranged from not knowing where to start (48% of respondents) to GPU and memory limits (40%) to the mundane frustration of Python environment conflicts (29%). These aren't beginners either—among those who had already completed at least one AI project in the past two years, 95% still reported a barrier.

This is the gap that Junjie Yin, Buxin She, Xinyu Feng, and Fangxing Li set out to close with a new educational framework that pairs core AI concepts directly with power-system problems. Their approach—released as free, cloud-executable Jupyter notebooks and validated through an IEEE webinar series that drew more than 590 live attendees—asks a deceptively simple question: What if the way we teach AI to power engineers actually resembled the problems those engineers solve?

The Science

The conventional path into AI education runs through MNIST. The handwritten-digit database has launched millions of learning journeys: a neural network learns to distinguish a 3 from an 8, and somewhere in that process, a learner absorbs the fundamentals of gradient descent, convolution, and loss functions. It's a sensible scaffold. But 54% of the power-systems professionals surveyed by Yin and colleagues rated image-based examples like MNIST as only "slightly" or "not at all" relevant to their work. Only 11.5% called them "very relevant."

This isn't a judgment on MNIST's pedagogical value. It's a statement about transfer. A power engineer staring at digits on a screen sees no connection to the voltage angles, power flows, and load curves that define their domain. The methods look abstract, the context feels foreign, and the incentive to push through the learning curve evaporates before it begins.

The research team—based at the University of Tennessee's CURENT research center and Kansas State University—started with this insight and built outward. First, they surveyed the community to understand not just what barriers existed, but who experienced them, in what combinations, and with what intensity. Then they designed a progressive framework that maps AI concepts onto recognizable power-system tasks, creating what they call "engineering-grounded AI" (EGAI): workflows where AI follows established engineering and domain rules rather than operating as a task-agnostic black box.

The survey instrument comprised nine questions covering respondent background, current AI adoption, specific barriers encountered, perceived relevance of generic AI tutorials, and interest in domain-specific hands-on materials. Of 52 valid responses, 38.5% were graduate students, 26.9% faculty or research scientists, 19.2% industry engineers, and 15.4% undergraduates. Experience levels were evenly distributed, with roughly 29% each in the under-three-years and five-to-ten-years brackets, and about 27% with more than a decade in the field. This breadth matters: it means the barriers they identified aren't artifacts of a particular career stage or institution type.

The resulting framework organizes six hands-on modules across three difficulty tiers, each released as a Jupyter notebook that runs locally or in a browser via Google Colab—no installation required, no GPU at home necessary. The modules were delivered through an IEEE online course and a two-part webinar series that, in its first installment, attracted enough viewers to rank among the ten most-attended IEEE Power & Energy Society webinars ever.

What They Found

The survey data surfaced three interconnected findings that shaped every design decision.

Interest is broad and cross-domain, but adoption remains uneven. Respondents' work spanned essentially all IEEE PES technical committees, with the heaviest concentration in power system operation, planning, and economics (58%); analytic methods (48%); transmission and distribution (37%); and dynamic performance (31%). This wasn't a niche community. When asked what AI tasks they'd already applied to their work, respondents pointed to forecasting (23 mentions), grid operation and stability (23), protection and control (9), market analysis (8), asset monitoring (7), and cybersecurity (5). Yet 25% had never applied AI beyond conversational tools. There was appetite, but also a gap between appetite and action.

Barriers cluster and persist, even among the experienced. The three most common obstacles were insufficient background knowledge (48%), hardware limitations (40%), and environment-configuration issues (29%). What stands out is how these barriers varied by role: hardware limits were most acute for faculty and research scientists (57%), while environment setup troubled graduate students and industry engineers more than their academic counterparts. Insufficient background knowledge led for every group. Crucially, these weren't problems that disappeared with experience. Among the 39 respondents who had completed one or more AI projects, 95% still reported at least one barrier. AI adoption wasn't curing AI anxiety.

AI Application Domains in Power Systems

Label	Value
Forecasting	23
Grid operation & stability	23
Never used AI	13
Protection & control	9
Market analysis	8
Asset monitoring	7
Customer-side	5
Cybersecurity	5

Source: Community survey, N=52, multiple selections allowed

Generic tutorials feel disconnected from power engineering. When asked how relevant image-based examples like MNIST handwritten-digit recognition were to power-system applications, only 11.5% said "very relevant." The modal response was "neutral" at 35%, with another 25% split between "slightly" and "not at all." Open-ended responses elaborated: publicly available tutorials focused on computer-vision or text tasks while power-system problems centered on time series, waveforms, power-flow quantities, and dynamic responses. The community didn't just want more content; they wanted content that felt like their work.

This demand translated directly into enthusiasm for an alternative. When asked whether they would use a one-click, hands-on tutorial designed specifically for power and energy systems, 65.4% said yes, and another 28.8% said maybe. Combined, that's 94% affirmatively inclined. The survey didn't have to prod for interest; it had to capture the shape of a need that was already there.

AI Learning Barriers in Power Systems

Label	Value
Insufficient background	48 %
Hardware limits	40 %
Environment setup	29 %
OS differences	17 %
Debugging	13 %

Source: Community survey, N=52, multiple selections allowed. Among those who had completed AI projects, 95% still reported barriers.

The researchers responded by building a library of six modules, each deliberately paired with a power-system task. The foundational tier teaches nonlinear regression and function approximation using a feed-forward deep neural network (DNN)—the kind of architecture that, in other contexts, learns to recognize cats in photographs. Here, it learns to predict load curves from historical data. The domain-coupled tier connects a convolutional neural network (CNN) directly to a physics-based power-flow simulator for a modified PJM 5-bus system, predicting bus voltages and line power flows without solving the nonlinear equations online. The frontier tier reaches toward research frontiers: DNN-assisted optimization for security-constrained dispatch, deep reinforcement learning (DRL) for battery storage control, and physics-informed neural networks (PINNs) for electromechanical dynamics.

Every module follows the same notebook skeleton: a Settings block where hyperparameters live, a data generation or loading step, model construction, training, evaluation, and visualization. The learner who understands one module can navigate any other by editing a single configuration block. Defaults are sized to run on a laptop in seconds, with explicit guidance on how to scale up to larger datasets, deeper networks, or more realistic test systems.

The foundational tier centers on a single DNN template that maps a scalar input to a scalar prediction. To demonstrate generality, the Settings block exposes several selectable function modes: sine, cosine, exponential, logarithmic, cubic polynomial, and a user-defined custom option. The same architecture and training pipeline—implemented in TensorFlow/Keras with mean-squared-error loss and the Adam optimizer—reproduce all of them. In evaluations on held-out test data, the network tracked both trigonometric targets and polynomial functions with comparable fidelity, confirming that the template isn't tailored to any single mapping.

Keeping the identical pipeline, the second module (DNN_ load_curve) swaps the analytic generator for measured load data, partitioning the dataset with a 70/30 train-test split. The network learns the input-output relationship from actual measurements rather than a closed-form expression, and performance on the held-out 30% demonstrates what matters: predictive quality must be judged on data the model hasn't seen.

The domain-coupled tier moves from synthetic functions to physics. A CNN—architecturally suited to structured, spatially organized data—is trained as a power-flow surrogate for the modified PJM 5-bus system. The network maps known operating set-points (bus loads and generator dispatch) to solved system states (bus voltage magnitudes and line power flows), bypassing the need to solve nonlinear power-flow equations on the fly. Dataset generation uses Monte Carlo sampling across the load and generation parameter space, with careful attention to train-test leakage prevention: the system topology stays fixed, but operating conditions are randomly varied, and the test set samples from a separate distribution.

Results for voltage prediction show the expected tradeoffs between training epochs, filter counts, and generalization. A model trained for 50 epochs with 24 filters per layer achieved reasonable accuracy but showed room for improvement. Doubling filters to 240 while holding epochs constant improved performance, and extending training to 200 epochs pushed accuracy further. The pattern illustrates a core lesson: there's no free lunch. More parameters and more training can help, but they also increase the risk of overfitting if the data don't support the additional complexity.

The frontier tier reaches into active research areas. The DNN-assisted optimization module embeds a learned model inside a mixed-integer linear program (MILP), replacing a computationally expensive constraint with a fast neural network approximation that preserves solvability. The battery storage module uses deep Q-learning (DQN), a reinforcement learning method where an agent learns a policy for sequential decision-making by maximizing cumulative reward, in this case minimizing energy costs while respecting battery constraints. The PINN module incorporates physical governing equations—the swing equation describing electromechanical dynamics—directly into the training loss, so the network learns dynamics that respect conservation of energy even when measurement data are sparse.

Why This Changes Things

The electricity grid is becoming something our grandparents wouldn't recognize. Solar panels on rooftops, wind farms hundreds of miles from population centers, batteries that charge when the wind blows and discharge when demand peaks—all of it adds flexibility and complexity in equal measure. Keeping the lights on in this world requires tools that can handle nonlinearity, uncertainty, and speed. AI offers those tools. But the people who need them most—power engineers who understand the physics of rotating machines and voltage collapse, who can tell you why a doubly-fed induction generator behaves differently under fault conditions—have been largely left out of the AI revolution.

This isn't a talent problem. It's an access problem. The 92% barrier rate in the survey didn't reflect a lack of intelligence or motivation; it reflected a mismatch between how AI is taught and how power engineers learn. MNIST is a fine tutorial for a computer scientist who wants to understand convolution. It's a poor tutorial for a grid operator who wants to predict tomorrow's load curve so they can plan reserve margin.

The framework Yin and colleagues built addresses this mismatch directly. By pairing AI concepts with power-system tasks at every level—function approximation with load-curve prediction, convolution with power-flow surrogates, reinforcement learning with battery scheduling—it makes the abstract concrete. The learner doesn't have to translate between domains; the materials do the translation for them.

Relevance of Generic AI Examples to Power Systems

Label	Value
Very relevant	12
Somewhat relevant	18
Neutral	15
Slightly irrelevant	8
Not relevant at all	5

Source: Community survey, N=52, single choice. Combined 'Neutral' to 'Not relevant at all': 54%.

The implications extend beyond individual learners. As renewable penetration increases and grid operators lean harder on AI for forecasting, optimization, and control, the field needs practitioners who can speak both languages: who understand the physics of the grid and the mathematics of machine learning well enough to catch when a model produces physically impossible results. The current pipeline produces specialists in one domain who dabble in the other. Engineering-grounded AI, done well, could produce engineers who think in both.

There's also a reproducibility dimension. The modules ship as executable notebooks with pinned environment specifications and one-click Colab execution. A learner in Knoxville and a learner in Seoul can run the same code and get the same results, modulo hardware differences. This matters for education—consistent baselines make it easier to isolate what changed when you changed a parameter—but it also matters for research. Power-systems papers that cite AI results often can't reproduce those results because the code is missing, poorly documented, or dependent on institutional resources. Open, executable baselines lower that barrier.

The webinar numbers suggest hunger for this kind of material. More than 590 live attendees for a technical webinar is substantial for any professional society; ranking it among the ten most-attended IEEE PES webinars ever puts it in rarefied company. The 344 repository visits in two weeks—before any formal announcement beyond the webinar—suggest people were bookmarking the resources, not just spectating.

What's Next

No framework is complete, and the authors acknowledge several limitations. The six modules are deliberately small and stylized: a 5-bus system rather than a 2000-bus transmission model, a single battery rather than a fleet of distributed energy resources. This is a feature, not a bug—the authors designed for entry-level accessibility—but it raises questions about scalability. Can the DNN that learned a cubic polynomial learn a regional load curve with 10,000 nodes? Can the PINN that solved a swing equation solve a full electromagnetic transient? The authors provide guidance on scaling up, but this remains to be tested in practice.

The survey sample, while diverse, is small (N=52) and self-selected. Respondents were presumably already interested enough in AI to complete a survey about AI learning barriers, which means the 92% barrier rate may actually underestimate the challenge in the broader community. Replication with a larger, less engaged sample would strengthen the findings.

The framework also sidesteps some harder questions about what "engineering-grounded AI" means when engineering and AI point in different directions. A physics-informed neural network that respects conservation laws is philosophically appealing, but what happens when the physics model is itself an approximation? What happens when the real system doesn't behave like the swing equation because of unmodeled harmonics or controller saturation? The modules teach the methodology; applying it to messy real-world cases is left to the learner.

These are not criticisms so much as next steps. The authors have laid a foundation that others can build on. The Tier 3 frontier modules—DRL for batteries, PINNs for dynamics—are themselves active research areas where best practices are still emerging. As those areas mature, new modules could follow. The framework's modularity makes that extension natural; the notebook skeleton stays the same, the content evolves.

The deeper contribution may be methodological. By demonstrating that AI education can be both rigorous and domain-grounded—that you don't have to choose between teaching concepts and teaching context—Yin and colleagues have offered a template for other fields facing similar gaps. AI in healthcare. AI in climate modeling. AI in finance. Each has its own jargon, its own data formats, its own physical constraints. The question "what if we taught AI through the problems people actually solve?" is not specific to power systems. The answer, as this work suggests, might be: a lot more people would actually learn it.

The open-source modules, online course, and webinar recordings remain available. For a community that has spent years watching AI transform adjacent fields while wondering when their turn would come, that accessibility matters. The barrier is lower now than it was before. Whether people cross it depends, as it always does, on whether the path ahead looks worth walking.