Short-term load forecasting under EU-AI Act Require... | AI Research

Short-term load forecasting (STLF) is a critical component of the electric power industry, essential for balancing energy supply and demand. This paper presents the results of a 41-day live challenge evaluating a forecasting pipeline for the aggregated German transmission-grid load. The study addresses the challenge of meeting modern regulatory standards—specifically the EU AI Act—while maintaining high forecast accuracy in a safety-critical environment.

A Compliance-First Forecasting Pipeline

The research introduces a forecasting pipeline built on the open-source Python library spotforecast2-safe. Unlike general-purpose machine learning tools, this library is designed to meet the requirements of the EU AI Act by prioritizing determinism, reproducibility, and auditability. The pipeline processes data from the European Network of Transmission System Operators for Electricity (ENTSO-E) to predict 24-hour load values. It incorporates anomaly detection, gap-aware data preparation, and a recursive multi-step forecasting algorithm based on a Light Gradient-Boosting Machine (LightGBM).

Methodology and Regulatory Context

The study treats STLF as both a software-engineering and a compliance problem. Because day-ahead forecasts inform decisions for critical infrastructure, the author argues that these systems must be auditable. The spotforecast2-safe library enforces this through strict development rules, including:

• Code-development rules: Ensuring deterministic outputs, fail-safe error handling, and a minimal attack surface.

• Process rules: Implementing traceability, documented threat models, and supply-chain integrity checks.
  The pipeline uses surrogate-model hyperparameter tuning via SpotOptim and compares its performance against the Optuna optimizer.

Performance Results

The spotforecast2-safe pipeline outperformed the official ENTSO-E day-ahead forecast baseline during the 41-day challenge. The study also compared the performance of these local, transparent models—referred to as macl2l—against large, energy-intensive foundation models like Chronos-2. The findings indicate that the smaller, auditable local models remain competitive with foundation models containing over 100 million parameters, while requiring significantly fewer computational resources.

Regulatory Implications

The paper notes that while the EU AI Act classifies AI systems used in electricity supply as high-risk, a day-ahead load forecaster typically sits alongside critical infrastructure rather than within its direct safety perimeter. However, because these forecasts influence operational decisions, the author maintains that they should be held to the same standards of determinism and auditability. The study also acknowledges upcoming regulations, such as the Cyber Resilience Act and the revised Product Liability Directive, which will further define the responsibilities of software suppliers in safety-critical sectors.