Bridging Artificial Intelligence and Power Systems Education Using a Hands-On Executable Framework

Technical Summary: Bridging Artificial Intelligence and Power Systems Education Using a Hands-On Executable Framework

Problem Statement
While Artificial Intelligence (AI) is increasingly central to power and energy systems for modeling, forecasting, optimization, and control, a significant gap exists in educational resources for newcomers and interdisciplinary learners. Existing academic works often emphasize specialized applications or high-level research, leaving beginners with fragmented, context-less code or workflows that are too abstract. Furthermore, a community survey conducted by the authors reveals that 92% of researchers and practitioners report at least one barrier before running an AI model (e.g., insufficient background knowledge, hardware limitations, or environment configuration issues). Additionally, 94% of respondents expressed a desire for a power-specific, hands-on course. Current generic tutorials, often based on image recognition (e.g., MNIST), are perceived as having low relevance to power system problems, which are dominated by time series, waveforms, and power-flow quantities.

Methodology
To address these gaps, the authors propose a framework for "Engineering-Grounded AI" (EGAI), where AI workflows adhere to established engineering and power-system domain rules rather than acting as task-agnostic black boxes. The methodology consists of three core components:

1. Community-Driven Design: A targeted survey (N=52) of power-and-energy researchers and practitioners identified specific learning barriers and resource preferences, directly informing the design of the educational materials.
2. Progressive Knowledge Mapping: The framework organizes teaching modules along a "progressive difficulty ladder" that explicitly pairs core AI concepts with representative power-system tasks. This mapping avoids isolation by coupling:
   • Nonlinear regression with load-curve prediction.
   • Convolutional Neural Networks (CNNs) with power-flow surrogates.
   • Learned constraints with security-constrained decision-making.
   • Deep Reinforcement Learning (DRL) with battery storage control.
   • Physics-Informed Neural Networks (PINNs) with the swing equation.
3. Executable Module Library: The authors developed six open-source, hands-on modules released as Jupyter notebooks. These run locally or on Google Colab, utilizing a unified template structure (Settings, data generation, model construction, training, evaluation, visualization) to ensure reproducibility and ease of modification. The modules are categorized into three tiers:
   • Tier 1 (Foundational): Generic Deep Neural Network (DNN) templates for function approximation (analytic functions) and load-curve fitting (measured data).
   • Tier 2 (Domain-Coupled): A CNN-based power-flow surrogate for a modified PJM 5-bus system, predicting bus voltages and line power flows from operating set-points using Monte-Carlo generated data.
   • Tier 3 (Frontier): Advanced applications including DNN-assisted optimization (embedding learned constraints in Mixed-Integer Linear Programming), DRL for battery energy storage system (BESS) scheduling, and PINNs for solving the swing equation with sparse data.

Key Contributions

• Needs Analysis: Empirical validation of learning barriers and the demand for domain-specific AI education through a community survey.
• Design Framework: A structured approach to mapping AI concepts to power system tasks, moving beyond generic examples to physically interpretable problems.
• Open Module Library: The release of six executable modules covering a spectrum from basic regression to physics-informed learning, all built on a shared, configurable notebook template.
• Reproducibility and Validation: Delivery of these resources through an IEEE online course and IEEE Power & Energy Society (PES) webinar series, ensuring the materials are accessible and reproducible.

Results
The framework was validated through early deployment and community engagement:

• Webinar Attendance: Part I of the associated webinar attracted over 590 live attendees, ranking among the top ten most-attended IEEE PES webinars.
• Repository Engagement: The open-source repository received over 344 visits within two weeks of release.
• Technical Performance:
  • The foundational DNN templates successfully generalized across different function families (trigonometric, polynomial) and real load data.
  • The CNN power-flow surrogate demonstrated that model capacity must be paired with sufficient training length; a model trained for 200 epochs with 240 filters achieved a Mean Absolute Error (MAE) below $10^{-3}$ pu for bus voltages, significantly outperforming undertrained variants.
  • The DRL module for BESS control achieved a 5.9% cost reduction compared to a no-storage baseline and outperformed a simple cycle rule, validating the state-action-reward coupling.
  • The PINN module successfully reproduced rotor angle and angular speed trajectories more accurately than an ordinary neural network when trained on sparse data, demonstrating the value of embedding governing equations.

Significance
The paper claims that this work lowers the entry barrier for applying AI in power systems by providing a "minimal yet extensible coding baseline." By shifting learners from passive reliance on Large Language Models (LLMs) to active experimentation with parameters, architectures, and problem formulations, the framework fosters a deeper understanding of how AI interacts with physical systems. The authors posit that this approach promotes "Engineering-Grounded AI," ensuring that AI tools in power systems are developed with inputs, workflows, and outputs that conform to established engineering principles. The framework is designed not only for power systems but also as a reusable template for other data-driven engineering disciplines, bridging the gap between foundational AI concepts and advanced, domain-specific applications.