A compliance lead walks into a vendor briefing and hears: "Our platform uses a transformer-based model with RAG grounding, reducing hallucination risk through ensemble validation and continuous drift monitoring." She nods. She has no idea what any of that means. Six months later, her organization signs a contract and discovers the "drift monitoring" was a dashboard metric that nobody was assigned to watch, the "ensemble validation" was a marketing term for running the same prompt twice, and the "RAG grounding" relied on a document database that had not been updated in four months. Understanding AI vocabulary is how energy professionals protect their organizations from the gap between what vendors say and what their products actually do.

Why Vocabulary Matters in a Regulated Industry

Every technical field has its own vocabulary, and AI is no exception. But in a regulated, safety-critical industry like energy, vocabulary carries particular weight: the terms you use in a contract, a rate-case exhibit, a compliance filing, or an audit response have legal and financial consequences. A utility professional who cannot distinguish "model training" from "model fine-tuning" is vulnerable to a vendor who uses those terms interchangeably. A compliance lead who does not know what "drift" means is not equipped to require drift monitoring as a contractual deliverable. A planner who does not understand what "uncertainty quantification" means cannot evaluate whether an AI forecast gives them the information they need for a procurement decision.

This lesson defines the AI terms you will encounter most often in utility work, in the order that makes them easiest to build on. Each definition includes a "why you care" sentence written in grid terms, because the goal is not to make you a data scientist: it is to make you a literate, skeptical professional who can participate in AI decisions with your organization's interest in mind.

Model Training and Inference: The Two Modes

Training is the process of building an AI model from data. During training, the model is shown a large dataset and adjusts its internal parameters (numerical weights) to minimize its prediction errors on that data. Training is computationally intensive and time-consuming. It happens in advance of deployment, not in real time. Why you care: the training data and training methodology determine what the model knows and what it can do. When you ask "what was this model trained on?" you are asking the most important question in AI evaluation.

Inference is the process of using a trained model to produce outputs. When you submit a load forecast request, when you run a completeness check on an interconnection application, or when you ask a generative AI to draft a report section, you are doing inference. Inference is fast. It happens in real time. Why you care: the model during inference is not learning from the new inputs you give it. It is applying the weights learned during training. Giving an AI model new information during inference does not update its underlying knowledge; it only provides context for that specific interaction.

Fine-tuning is a form of additional training applied to a pre-trained model, using a smaller, domain-specific dataset to improve the model's performance on a specific task. A general-purpose large language model fine-tuned on NERC standards text and utility compliance documents will produce better compliance language than the base model. Why you care: when a vendor claims their model is "fine-tuned on utility data," ask what specific data, from what time period, and how the fine-tuning validation was conducted. Fine-tuning on outdated or narrow data can make a model worse, not better, on specific regulatory content.

Prompt is the input you provide to a generative AI model at inference time. The quality of the prompt significantly affects the quality of the output. Why you care: learning to write effective prompts (specifying jurisdiction, units, required citation format, uncertainty flags) is a professional skill that determines whether you get reliable output or unreliable output from the same underlying model.

Accuracy, Uncertainty, and Drift: The Performance Vocabulary

MAPE (Mean Absolute Percentage Error) is the standard accuracy metric for load forecasts. It measures the average percentage difference between forecast and actual values across the evaluation period. AI day-ahead load forecasting achieves roughly 1 to 2 percent MAPE on stable load regimes, compared to 3 to 5 percent for traditional methods. Why you care: MAPE is the number a planner defends in a rate case. Know what it means, know how it was measured (training vs. holdout, what time period, what load conditions), and never cite a vendor's MAPE benchmark as your utility's expected MAPE without validating it on your own data.

Holdout validation is the practice of reserving a portion of historical data that the model was not trained on, then measuring its accuracy on that reserved set. Holdout MAPE is more meaningful than training MAPE because it measures how the model performs on unseen data, which is the relevant question for future performance. Why you care: a model with excellent training MAPE but poor holdout MAPE is overfitting, learning the noise in the training data rather than the true underlying pattern. Always ask for holdout MAPE.

Uncertainty quantification refers to the model's ability to produce not just a point forecast but also a range or probability distribution around that forecast (the confidence interval, or prediction interval). A load forecasting model that outputs "peak load = 4,850 MW" is less informative than one that outputs "peak load = 4,850 MW with a 90 percent confidence interval of 4,700 to 5,000 MW." Why you care: uncertainty quantification is what allows a planner to make appropriate reserve procurement decisions. A point forecast with no uncertainty estimate gives the appearance of precision without the information needed to act on it safely.

Model drift (also called concept drift or data drift) refers to the degradation in model accuracy that occurs when real-world patterns change and diverge from the patterns the model was trained on. For load forecasting, drift is caused by new large loads, new DER penetration, changes in customer behavior, or structural shifts in the economy. Why you care: a load forecasting model that was accurate eighteen months ago and has not been monitored since may be significantly less accurate today. Drift monitoring (tracking model accuracy over time and triggering a review when accuracy degrades) is an operational requirement for any production AI deployment, not an optional feature. Require it contractually.

Overfitting occurs when a model learns the noise and idiosyncratic patterns in its training data rather than the true underlying patterns. An overfit model will have excellent training MAPE and poor holdout MAPE. It has essentially memorized the training examples rather than learning the general rule. Why you care: an overfit load forecasting model will appear highly accurate in internal testing and degrade immediately on real-world deployment, producing high forecast errors in the first months of use.

Underfitting is the opposite: the model has learned a relationship that is too simple to capture the true complexity of the data. An underfit model will have poor accuracy on both training and holdout data. Why you care: an underfit model is not extracting the available signal from its inputs. If a model shows similar MAPE to traditional regression methods on a modern dataset with rich weather and DER features, it may be underfitting (the AI advantage is not materializing because the model is too simple for the available data).

Architecture Terms: What the Model Is Built From

Neural network is the broad class of AI models inspired loosely by biological neurons. A neural network consists of layers of connected nodes (neurons), each of which applies a mathematical transformation to its inputs. The "deep" in deep learning refers to models with many layers. Why you care: knowing the architecture type helps you understand the model's data requirements and failure modes. You do not need to understand the math; you need to know enough to ask the right questions.

Transformer architecture is the neural network design that underlies both large language models (for text generation) and increasingly, state-of-the-art load forecasting models. The key innovation is the attention mechanism, which allows the model to weigh the relevance of different parts of the input when producing each part of the output. Why you care: when a vendor says "our forecasting model uses a transformer architecture," they are describing a specific family of models that tends to be highly competitive on complex, long-range sequence tasks but requires substantial training data and computational resources. This is relevant to understanding the model's data requirements and the vendor's infrastructure costs.

Gradient boosting is an ensemble method that builds a strong predictive model by combining many simple decision tree models, each of which corrects the errors of the previous one. Implementations like XGBoost and LightGBM are widely used for load forecasting, outage prediction, and customer segmentation. Why you care: gradient boosting models are relatively interpretable (you can get feature importance rankings showing which variables drive the forecast most), require less training data than deep neural networks, and degrade more gracefully on limited data. They are often the right choice for utilities with moderate-sized historical datasets.

Ensemble model is a model that combines the predictions of multiple individual models to produce a final output, typically more accurate and more robust than any single model. A load forecasting ensemble might combine a gradient-boosted model, an LSTM, and a regression baseline. Why you care: when a vendor describes their product as an "ensemble," ask what models are in the ensemble and how they are combined. "Ensemble" can mean a sophisticated multi-model architecture with proven accuracy benefits, or it can mean running two similar models and averaging the outputs. The distinction matters for evaluating the claimed accuracy.

Data Infrastructure Terms: Where the Model Lives

Training data is the dataset the model learned from. Training data quality, representativeness, and recency are the primary determinants of model quality. Why you care: garbage in, garbage out applies more strictly to AI models than to traditional regression. A model trained on incomplete or biased historical load data will systematically reproduce those biases in its forecasts.

Feature is an input variable used in model training. For a load forecasting model, features include temperature, humidity, hour of day, day of week, month, holiday indicator, and increasingly, behind-the-meter solar output, EV load estimates, and large-load customer activity. Why you care: the feature set determines what information the model can use to make predictions. A model without temperature as a feature cannot respond to weather changes. A model without large-load customer features cannot account for data-center demand. When evaluating a forecasting platform, ask for the complete feature list.

Data pipeline is the automated system that collects, cleans, transforms, and delivers data to the AI model for training or inference. Why you care: a broken or delayed data pipeline is a common cause of AI operational failures. A load forecasting model that depends on real-time weather data delivered via an API will fail if that API is unavailable. Data pipeline reliability is an operational requirement, not just a technical detail, and it should be addressed in the vendor contract.

RAG (Retrieval-Augmented Generation) is the architecture where a generative AI retrieves relevant documents from a controlled database before generating a response. RAG grounds model output in current, specific documents rather than training memory alone. Why you care: for utility regulatory work, a RAG-connected tool that draws from your current filed tariffs, current NERC standards, and current asset records is substantially more reliable for compliance drafting than a base model working from training memory that may be months or years out of date.

Digital twin is a virtual model of a physical asset or system (a substation, a distribution feeder, a generating unit) that is updated in real time with operational data and used for simulation, optimization, or predictive analysis. Why you care: digital twins are increasingly used as the data layer for AI applications in grid operations. An AI topology optimization tool that draws on a real-time digital twin of the transmission network can incorporate current equipment status, ratings, and configurations. A tool that draws on a static network model may be optimizing against outdated system information.

Reliability and Governance Terms

IT vs. OT (Information Technology vs. Operational Technology) is a critical distinction for AI in utilities. IT systems handle business processes: billing, customer information, document management, regulatory filing. OT systems control physical equipment: EMS/SCADA, ADMS, protection relay settings, substation control. AI in IT environments faces standard enterprise security and data governance requirements. AI in OT environments faces additional NERC CIP requirements, real-time performance requirements, and reliability standards that IT AI deployments do not. Why you care: a vendor who tells you their AI platform is "enterprise-grade secure" may be describing IT security, not OT security. Ask specifically which NERC CIP standards the OT components of their platform have been evaluated against.

ADMS (Advanced Distribution Management System) is the modern distribution control system that integrates OMS, DMS, and increasingly DER management functions. AI is increasingly integrated into ADMS for outage prediction, restoration optimization, and DER orchestration. Why you care: the ADMS is where distribution AI applications live in production. Understanding the ADMS integration requirements is essential for evaluating any distribution AI platform.

Model governance refers to the organizational processes and documentation that ensure AI models are developed, validated, monitored, and retired in a controlled, auditable way. Why you care: NERC reliability standards and commission ratemaking requirements place accountability on humans for AI-assisted decisions. Model governance is the framework that demonstrates that accountability: who approved the model, what validation was done, who reviews the accuracy metrics, and who can authorize a change to the model. Without model governance, an AI deployment in a utility context is an undocumented dependency on an opaque system.

Explainability (also called interpretability) refers to a model's ability to produce human-understandable explanations for its outputs. A load forecasting model with explainability features can show which input features drove the forecast (temperature was the dominant driver, followed by hour of day). Why you care: explainability is increasingly required for AI deployed in rate-regulated contexts. A commission that asks "why did your AI forecast predict a 5 percent load growth?" needs an answer grounded in documented inputs and model behavior, not "the model said so."

A Terminology Reference: Grid for the Working Professional

The table below summarizes the key terms from this lesson with the one-sentence "why you care" specific to utility work.

Key Takeaways

• Training is when the model learns from data; inference is when it applies that learning. Giving a model new information during inference does not update its underlying knowledge; it only provides context for that specific interaction.
• MAPE (Mean Absolute Percentage Error) is the primary load forecasting accuracy metric. Always ask for holdout MAPE measured on a period that includes any significant load-regime changes, not training MAPE or vendor reference benchmarks.
• Model drift is the silent failure mode: a model's accuracy degrades as real-world patterns diverge from its training history. Require continuous drift monitoring as a contractual deliverable for any production AI deployment.
• RAG (Retrieval-Augmented Generation) grounds generative AI output in current, controlled documents. For regulatory drafting, a RAG-connected tool is substantially safer than a base model relying on a potentially stale training corpus.
• The IT vs. OT distinction determines which regulatory requirements apply to an AI deployment. OT-side AI in control systems is subject to NERC CIP standards; IT-side AI in business processes is not. Vendors who cannot specify which side their product operates on cannot be evaluated for OT deployment.
• Model governance (documented processes for training, validation, monitoring, and change management) is how a utility demonstrates human accountability for AI-assisted decisions to regulators and auditors.
• Explainability is increasingly required for rate-regulated AI applications. A commission or auditor will expect a human-understandable account of why the AI forecast or recommendation produced a specific output; "the model said so" is not an acceptable answer.