Power Market Modeling

Causal Analysis of Renewable Energy Penetration and Electricity Price Forecasting with Transformers

Master’s Thesis - Humboldt-Universität zu Berlin
Author: Emircan Ince
Supervisors: Prof. Dr. Stefan Lessmann · Prof. Dr. Jan Mendling

Causal Impact of Renewable Penetration

This section quantifies the short-run price effects of renewable generation in the German day-ahead market from January 2022 to December 2024 by applying causal machine learning techniques.

Non-Parametric Mean Surfaces (LWPR)

• Solar (left): Steep price drops up to ~25%; strongest around noon (e.g. −3.1 €/MWh at 15%, 12:00).
• Wind (right): Consistent decline up to ~60% share. Smoother and persistent across 24h, especially overnight.

Causal Heterogeneity (DML)

• Solar CATE (left): Strong at low shares (−0.95 €/MWh at 5%). Weakens but remains negative beyond 30%.
• Wind CATE (right): Stable impact (−0.8 €/MWh up to 10%, −0.75 €/MWh around 30%). Persists up to 60%.

Forecasting Results

The forecasting section benchmarks six models, including both deep learning architectures and linear baselines across multiple prediction horizons, focusing on accuracy and stability under a unified experimental protocol.

Horizon	Metric	TimeXer	iTransformer	PatchTST	DLinear	SCINet	Autoformer
24h	MSE	0.208	0.302	0.380	0.209	0.210	0.261
	MAE	0.114	0.212	0.288	0.118	0.117	0.153
48h	MSE	0.243	0.303	0.360	0.234	0.247	0.268
	MAE	0.152	0.217	0.269	0.147	0.154	0.175
96h	MSE	0.274	0.318	0.330	0.254	0.288	0.280
	MAE	0.187	0.236	0.251	0.172	0.205	0.191
168h	MSE	0.278	0.322	0.368	0.263	0.289	0.285
	MAE	0.192	0.240	0.274	0.186	0.204	0.201

Best values per horizon are bolded.

TimeXer performs best at the 24-hour horizon, while DLinear takes the lead from 48 hours onward, reflecting a trade-off between complexity and generalization.

Working with the Repository

Dependencies

Developed with:

• Python 3.9
• PyTorch 2.0
• LightGBM 4.6

All dependencies are listed and pinned in requirements.txt.

Setup

# Clone the repository
git clone https://github.com/emircanince/power.git
cd power

# Create and activate a virtual environment
python -m venv venv
source venv/bin/activate      # for macOS/Linux

# Install dependencies
pip install -r requirements.txt

Reproducing Results

Causal Analysis

To compute the LWPR surface, the overall DML estimate, and compare CATE to the observational mean, run:

python models/nonlinear_regression_models/LWPR_mean_wind.py
python models/double_machine_learning/DML_overall_wind.py
python models/double_machine_learning/dml_vs_mean_plot_wind.py

Replace wind with solar in the filenames to run the analysis on the solar dataset.

Outputs are written to CSV and saved as figures in the png/ directory.

Forecasting Experiments

To train and evaluate TimeXer and the benchmark models:

bash scripts/TimeXer.sh       # optional: add --pred_len 48 / 96 / 168 (default: 24)

The script handles data loading, training, checkpoints, and TensorBoard logging automatically.

Models: TimeXer, iTransformer, PatchTST, DLinear, SCINet, Autoformer

Change TimeXer.sh to the corresponding script (e.g., iTransformer.sh) to run a different model.

Project Structure

.
├── checkpoints/                      # Model checkpoints
├── data/                             # Preprocessed dataset (thesis_data.csv)
├── data_provider/                    # Data loading and preprocessing logic
├── exp/                              # Experiment configurations
├── layers/                           # Custom neural network layers
├── models/                           # Causal analysis + forecasting models
│   ├── double_machine_learning/      # DML-based treatment effect estimation
│   ├── nonlinear_regression_models/  # LWPR models
│   ├── Autoformer.py
│   ├── DLinear.py
│   ├── PatchTST.py
│   ├── SCINet.py
│   ├── TimeXer.py
│   └── iTransformer.py
├── png/                              # Output figures and plots
├── results/                          # Stored outputs
├── scripts/                          # Shell scripts for model training/evaluation
├── test_results/                     # Test logs
├── utils/                            # Utility functions and helpers
├── .gitignore                        # Git exclusion rules
├── README.md                         # ← You are here
├── requirements.txt                  # Dependency specification
└── run.py                            # Main script to launch workflows

Keywords: Causal Machine Learning, Deep Learning, Time Series Modeling, Renewable Energy, Electricity Price Forecasting

References

• Market Impact of Renewables
• TimeXer
• ENTSO-E Transparency Platform