Machine Learning and Data Science for Power System

Course Material

Module Aims

Introduce core ML concepts and methods relevant to modern power system operation, forecasting, security, and data-driven decision making. Students will learn basic supervised + unsupervised ML, performance metrics, model selection, and a first look at time-series foundation models.

Learning Outcomes

By the end of this module, students should be able to:

  • define and distinguish supervised vs unsupervised learning
  • implement and tune basic ML models (linear regression, logistic regression, SVM, PCA, k-means)
  • evaluate ML models using MSE, F1, TPR/FPR, etc
  • understand generalization, overfitting, and regularization
  • conduct hyperparameter tuning with a proper validation pipeline
  • understand high-level concepts of time-series foundation models (e.g. Chronos) (New in 2025-2026)
  • use pretrained Time-Series Foundation Models (e.g. Amazon Chronos) for energy forecasting (New in 2025-2026)

Indicative Topics

  1. Introduction
    • why ML is needed in modern power systems
    • data-driven vs model-driven approaches
  2. Supervised Learning
    • linear regression + polynomial basis
    • logistic regression + cross entropy
    • support vector machine (hard/soft margin intuition)
  3. Unsupervised Learning
    • principal component analysis (dimension reduction)
    • k-means clustering
  4. Performance Assessment
    • regression: MSE
    • classification: TPR, FPR, FNR, TNR, F1
  5. Model Capacity / Overfitting
    • bias–variance trade-off
    • regularization (weight decay)
  6. Hyperparameters & Validation
    • train/validation/test split
    • grid search
  7. Special Topics (New in 2025-2026)
    • No Free Lunch Theorem
    • Foundation Models (definition & transferability)
    • Time-Series Foundation Models (e.g. Amazon Chronos)
    • where to obtain real power system data (ENTSO-E, ERA5, etc.)

Outcomes of the course work

Based on the lecture notes, the course work trains students to apply modern machine learning and data science methods to power system operation, forecasting and security. The module integrates theory from lecture notes (data-driven modelling, probabilistic forecasting, state estimation, attack modelling) with hands-on coursework tasks on real grid data.

After completing the coursework, students will be able to:

  • Construct complete ML pipelines for power system applications.
  • Understand zero-shot learning, autoregression, and covariate-informed forecasting (New in 2025-2026).
  • Use pretrained Time-Series Foundation Models (e.g. Amazon Chronos) for energy forecasting (New in 2025-2026).
  • Collect and preprocess power system datasets (including missing time stamps / irregular sampling) (New in 2025-2026).
  • Build power network measurement matrices under DC power flow assumptions (A matrix, Bpf, Bpi).
  • Perform DC state estimation using weighted least squares and correctly handle reference buses.
  • Design chi-square based Bad Data Detection (BDD) with controlled false positive rate.
  • Train and evaluate machine learning classifiers for FDI attack detection on measurement data.
  • Present solutions using pseudo-code, equations and clear experimental methodology.