Time Series Analysis TIS3 Capstone Project – Electricity Price Forecasting for EU AI Datacenter Site Selection

Project Overview

This project forecasts daily wholesale electricity prices in selected European countries to support cost- and risk-aware site selection for a new AI data center.

Electricity is a major driver of datacenter operational expenditure (OpEx), and price uncertainty directly translates into planning and budgeting risk. The goal is therefore not only to forecast prices, but to turn forecasts into a decision tool.

The project delivers an end-to-end forecasting pipeline, model comparison across multiple approaches, and daily forecasts with uncertainty for January 2026.

---

Business Problem

Problem Forecast daily electricity prices to support cost- and risk-aware selection of an AI data center location in the EU.

Why this matters

• Electricity cost is a direct and continuous operational expense
• Forecast uncertainty translates into operational and budgeting risk
• Long-term infrastructure decisions require comparability across locations

---

Data

• Source: Ember dataset from ENTSO-e (European Network of Transmission System Operators for Electricity)

• Data type: Daily wholesale electricity prices

• Time range: 01.01.2015 – 28.12.2025

• Countries:

  • Austria
  • Germany
  • Switzerland
  • France
  • Poland
  • Estonia

• Forecast horizon: Daily prices for January 2026

---

Forecasting Pipeline (CRISP-DM)

The project follows a CRISP-DM-style forecasting pipeline:

1. Business understanding
2. Data preparation and preprocessing
3. Baseline forecasting (naive, weekly seasonal naive)
4. Statistical models (Holt-Winters, AutoARIMA)
5. Machine learning models (Huber Regression)
6. Deep learning benchmark (xLSTM)
7. Model evaluation and selection
8. Final forecasting with uncertainty

---

Data Engineering

• Extracted country-level electricity price series
• Stored each country as a separate tabular dataset for clarity and efficiency
• Aligned all data to a daily frequency
• Imported and assembled datasets dynamically for modeling and evaluation

---

Modeling Approach

The goal was to compare model families, not just individual methods.

Models evaluated

• Baselines

  • Naive
  • Weekly seasonal naive

• Statistical models

  • Holt-Winters
  • AutoARIMA

• Machine learning

  • Huber Regression with lag-based features

• Deep learning (benchmark)

  • xLSTM

Key findings

• Statistical models did not outperform baselines
• Huber Regression consistently improved performance across countries
• xLSTM did not add predictive value under the given setup and data size

---

Evaluation

• Validation strategy: Time-based validation to avoid information leakage

• Metrics:

  • MAE
  • RMSE
  • MAPE (reported, but MAE/RMSE used for selection)

• Selection criterion:

  • Consistent out-of-sample improvement relative to baselines

---

Final Forecast

• The final Huber Regression model was retrained on the full dataset
• Daily electricity prices for January 2026 were forecasted for all countries
• Uncertainty was quantified using residual-based prediction intervals

Forecast output: outputs/jan_2026_electricity_forecasts.csv

---

Decision Support & Insights

Forecasts were translated into decision-oriented insights using:

• Expected average prices
• Prediction interval widths (planning risk proxy)
• Upper-bound cost scenarios
• Monthly accumulated OpEx under a 1 MW, 24/7 load assumption

These outputs enable:

• Cost-efficient location comparison
• Risk-aware planning
• Transparent trade-offs between low cost and high uncertainty

---

Team

• Peter Keszei Data setup, preprocessing, country-level dataset preparation

• Mohamed Haroun Modeling, evaluation, forecasting pipeline, model selection, January 2026 forecasts

• Raphael Suchomel Visualization, interpretation, decision framing, presentation

All team members contributed throughout the project; listed roles indicate areas of primary focus.