This project develops and evaluates time series models for forecasting solar power generation in Germany. We use both univariate (Moving Average, NBEATS, NHITS, Random Forest, CatBoost) and multivariate (LSTM) models. Our datasets, available on Hugging Face, include:
Key findings: CatBoost performed best for univariate analysis (R² .8457), while the original LSTM model excelled in multivariate analysis (R² .8826).
Package Info: A python package was created to ease the imports of the various helperfunctions when working with colab, for instance. The package can be installed via:
pip install timeseriesfcstA local version of the package can also be found in the repository direcotry tsfcst_local_version.
All required libraries can be installed by running pip install -r requirements.txt.
The data exploration, statistics and modeling was done in separate notebooks for univariate and multivariate analyses. This was done to maintain a clear structure and to avoid confusion, as different models may require different processing steps. The notebooks are structured as follows:
- Notebook 0 contains an overview of the target feature, solar power, and various visualisations.
- Notebook 1 contains the exploratory data analysis for all features, including weather data and correlation analyses.
- Notebook 2 contains the feature engineering and dataset creation.
- Notebook 3 contains the univariate time series models.
- Notebook 4 contains the multivariate LSTM model.
The rising demand for renewable energy and the variability in solar power generation due to weather conditions necessitate accurate prediction models for energy supply. This project outlines a comprehensive approach to forecasting solar power generation, utilising both univariate and multivariate time series models. Key methodologies include preprocessing techniques such as stationarity adjustments, Fourier analysis, and anomaly detection using autoencoders. Various models, including Moving Average, NBEATS, NHITS, Random Forest, and CatBoost for univariate data, and LSTM for multivariate data, are evaluated. The study leverages weather data from six German cities and highlights the challenges and potential improvements in the prediction models.
The necessity for a consistent and affordable energy supply is critical for industrial expansion, particularly as renewable and weather-dependent energy sources such as solar power become more widely used. As part of the Opencampus Course "Advanced Time Series Prediction," this group project intends to test alternative time series algorithms for forecasting solar power supply in Germany, taking into account the high volatility and seasonality associated with solar power output.
The solar power dataset was obtained from Energy Charts and historical weather data from Open-Meteo. Regarding the weather data, we considered six German cities (Templin, Kastellaun, Gütersloh, Ingolstadt, Erfurt, and Neumünster) that are recognised for their prominent solar energy centres while simultaneously covering a broad geographical range within Germany. Three datasets were created from the data, using different preprocessing, feature engineering, and feature selection techniques. The datasets have been made available on Hugging Face: Original Dataset, PCA Dataset, and Feature-Selected Dataset.The goal is to develop a reliable forecasting model that can accurately predict Germany's solar power supply, ultimately contributing to the advancement of renewable energy technologies.
To prepare the data for analysis, the following preprocessing steps were undertaken:
- Stationarity: The data was standardised, differentiated to remove trends, and normalised by annual volatility.
- Fourier Analysis: Various frequency cutoffs were tested to optimise the performance and quality of fit.
- Anomaly Detection: Data was sliced into smaller vectors, and an autoencoder was trained to detect anomalies based on reconstruction errors.
Several univariate models were tested, including:
- Moving Average: A simple mean of the past hours, yielding an R² score of .6150.
- NBEATS and NHITS: Advanced models with varying levels of computational complexity and performance.
- Autocorrelation Features: Lags with the highest autocorrelation values were identified to enhance model accuracy.
- Random Forest and CatBoost: Ensemble methods showed promising results, with R² scores of .8430 and .8457, respectively.
The Long Short-Term Memory (LSTM) model was employed for multivariate analysis, incorporating additional weather data and engineered features, including lagged features, rolling features, and time-dependent features such as sine and cosine transformations of the datetime values. Despite promising initial results (R² score of .8826 for the original LSTM model), the introduction of feature engineering and Principal Component Analysis (PCA) did significantly impede performance (R² scores of .4499 and .2834, respectively).
| Type | Model | MAE | MSE | RMSE | R² |
|---|---|---|---|---|---|
| Univariate | Moving Av. | 1.4802 | 5.9520 | 2.4397 | .6150 |
| Univariate | Rdn. Forest | .81939 | 2.4157 | 1.5542 | .8430 |
| Univariate | CatBoost | .8177 | 2.3737 | 1.5407 | .8457 |
| Multivariate | LSTM (Orig) | .2031 | .1739 | .0417 | .8826 |
| Multivariate | LSTM (FE) | 1.0324 | 2.5386 | 1.5933 | .4499 |
| Multivariate | LSTM (PCA) | 1.0554 | 3.0770 | 1.7541 | .2834 |
The greatest challenge we faced was related to the need for more GPU resources. Future efforts will focus on optimising the models in terms of efficiency and performance, as well as exploring transformer models on different data subsets.
This research underscores the importance of selecting appropriate preprocessing techniques and models for accurate solar power prediction. While univariate models like CatBoost showed robust performance, multivariate models require more careful feature selection and optimisation. This is indeed a trade-off, but with some additional effort, these models outperform the traditional, univariate methods.