Explainable Deep Learning for Analysis and Forecasting of Multivariate Time Series: Applications to Electricity Markets
Type: Master's Thesis
Author: Ali Burak Tosun
1st Examiner: Prof. Dr. Stefan Lessmann
2nd Examiner: Prof. Dr. Benjamin Fabian
| Length | Metric | NATMFeature | DNN | NATM | LSTM | SCINet | NATMTime |
|---|---|---|---|---|---|---|---|
| 8 | MAE | 2.6 | 5.8 | 3.6 | 5.2 | 1.0 | 2.8 |
| RMSE | 2.8 | 5.6 | 3.8 | 5.4 | 1.0 | 2.4 | |
| SMAPE | 3.2 | 5.8 | 3.2 | 4.6 | 1.0 | 3.2 | |
| R2 | 3.0 | 5.4 | 3.4 | 5.6 | 1.0 | 2.6 | |
| 16 | MAE | 2.4 | 6.0 | 3.6 | 5.0 | 1.0 | 3.0 |
| RMSE | 2.2 | 5.8 | 3.6 | 5.2 | 1.0 | 3.2 | |
| SMAPE | 3.0 | 6.0 | 3.0 | 4.6 | 1.0 | 3.4 | |
| R2 | 2.4 | 5.8 | 2.8 | 5.2 | 1.0 | 3.8 | |
| 24 | MAE | 2.6 | 6.0 | 3.0 | 5.0 | 1.0 | 3.4 |
| RMSE | 2.4 | 5.6 | 3.2 | 5.4 | 1.0 | 3.4 | |
| SMAPE | 2.2 | 6.0 | 3.0 | 4.6 | 1.2 | 4.0 | |
| R2 | 2.4 | 5.8 | 3.0 | 5.2 | 1.0 | 3.6 | |
| Overall Avg. Rank | 2.6 | 5.8 | 3.3 | 5.1 | 1.0 | 3.2 |
Electricity Price Forecasting (EPF) is crucial for renewable energy stakeholders. This thesis explores Neural Additive Time-series Models (NATMs)1 for forecasting electricity prices in Germany, France, Belgium, Spain, and the Netherlands. NATMs offer accurate predictions with interpretability, identifying contributions of different input variables. Using datasets from the ENTSO-E Transparency Platform2 (2020-2023), NATMs are compared with Long Short-Term Memory (LSTM) networks, Deep Neural Networks (DNNs), and SCINet3 using metrics like MAE, RMSE, SMAPE, and R². Results show SCINet achieves the highest predictive accuracy, but NATMs perform well with added interpretability. Contribution maps from NATMs show significant insights into the impact of historical prices and generation forecasts on electricity prices. This research demonstrates NATMs' potential in enhancing EPF accuracy and transparency.
Keywords: Electiricity Price Forecasting, Explainable Artificial Intelligence, Deep Learning, Multivariate Time-series Prediction, Neural Additive Time-series Models
Full text: The full text for this work is available here.
The environment was set up using Python 3.8 and Conda 4.12, though it should be compatible with similar versions.
-
Clone this repository:
git clone https://github.com/abtosun/xAI_electricity_prices.git cd xAI_electricity_prices -
Create and activate a virtual environment:
conda env create -f environment.yml conda activate myenv
The reproduction of results is divided into two main parts: Training & Evaluation, and Visualization. Follow the steps below to replicate the results for each part:
The Learning.py script includes both the training and evaluation phases of the models. You can run the Learning.py script with different methods, datasets, and input lengths. Alternatively, you can run the Learning.ipynb notebook, but you will need to change parameters in config.py. Here are the details:
| method | dataset_name | input_length |
|---|---|---|
Independent |
germany |
8 |
Feature |
netherlands |
16 |
Time |
belgium |
24 |
DNN |
spain |
|
SCINet |
france |
|
LSTM |
-
Make sure the Conda environment is activated or selected as a kernel
-
Prepare the data
Ensure the sample data is correctly placed in the
sample_datadirectory. This directory should contain subdirectories for each dataset (e.g.,germany,netherlands, etc.) with the respective CSV files. -
Run the
Learning.pyscript
Ensure that you have the necessary permissions and access to the GPUs on your servers to execute the experiments efficiently. To run the code for different combinations, use the following commands:
# Method: Independent, Dataset: Germany, Input Length: 24
python Learning.py --method Independent --dataset_name germany --input_length 24
# Method: Feature, Dataset: Spain, Input Length: 24
python Learning.py --method Feature --dataset_name spain --input_length 24
# Method: SCINet, Dataset: Belgium, Input Length: 24
python Learning.py --method SCINet --dataset_name belgium --input_length 24 In each run of Learning.py, the models will be trained and evaluated for 5 different folds. After each run, a results folder will be created, and the evaluation metrics will be saved in a CSV file.
The visualization code is used to generate visualizations for the results of the NATMs (Independent, Feature, Time) in EPF. Contribution maps for NATMs are created to provide insights into the model's behavior and feature importance.
The visualizations are generated using a Jupyter Notebook:
-
Make sure the Conda environment is activated or selected as a kernel
-
Unzip the checkpoints and logs:
unzip ckpt.zip -d ckpt
-
Run the Jupyter Notebook:
In the Notebook:
-
Change the
base_pathto the desired method, dataset, and input length. For example:base_path = 'ckpt/test_Feature_netherlands_8'
-
Set a valid date in the validation set for
input_date. For example:input_date = '2023-07-17 09:00:00'
-
These changes will ensure that contribution maps are generated for the specified method, dataset, and input length.
You can access the pretrained model weights and logs of NATMs in the ckpt.zip file. This archive are essential for visualization and understanding the model performance.
All evaluation metrics of the experiments can be found in the results folder named as evaluation_metrics_results.csv. This file provides detailed metrics for each experiment, allowing for easy reproduction and analysis of the results.
Additionally, the results folder contains a hyper-parameter_analysis subdirectory with detailed analyses of various hyper-parameters used in the experiments.
The project has a following structure:
.
├── ckpt.zip # Model checkpoints
├── environment.yml # Environment configuration file
├── Learning.ipynb # Jupyter notebook for learning experiments
├── Learning.py # Python script for learning experiments
├── README.md # The main documentation file for the project
├── results # Directory for storing result files
│ ├── evaluation_metrics_results.csv # CSV file containing evaluation metrics results
│ └── hyper-parameter_analysis # Directory for storing hyper-parameter analysis results
│ ├── dropout_regularization.csv # CSV file for dropout regularization results
│ ├── Hidden_sizes.csv # CSV file for hidden sizes analysis
│ └── layer_weight_combinations.csv # CSV file for layer weight combinations
├── sample_data # Directory for storing sample data files
│ ├── belgium
│ │ └── aggregated_dataframe_belgium.csv # Aggregated data for Belgium
│ ├── france
│ │ └── aggregated_dataframe_france.csv # Aggregated data for France
│ ├── germany
│ │ └── aggregated_dataframe_germany.csv # Aggregated data for Germany
│ ├── netherlands
│ │ └── aggregated_dataframe_netherlands.csv # Aggregated data for Netherlands
│ └── spain
│ └── aggregated_dataframe_spain.csv # Aggregated data for Spain
├── src # Source code directory containing scripts and modules
│ ├── config.py # Configuration file for the project
│ ├── data_prepare.py # Data preparation script
│ ├── __init__.py # Init file for the src module
│ └── models # Directory for model-related scripts
│ ├── activation
│ │ ├── exu.py # EXU activation function
│ │ ├── __init__.py # Init file for activation functions
│ │ └── relu.py # ReLU activation function
│ ├── base.py # Base model script
│ ├── DNN.py # Deep Neural Network model script
│ ├── featurenn.py # Feature neural network script
│ ├── __init__.py # Init file for the models module
│ ├── LSTM.py # Long Short-Term Memory model script
│ ├── model_selector.py # Model selection script
│ ├── natm.py # NATM model script
│ └── SCINet.py # SCINet model script
├── structure.txt # File containing the project structure
└── Visualization.ipynb # Jupyter notebook for visualization experiments