Accurately forecasting market dynamics, particularly within the electricity market, is essential for maintaining grid stability and preventing power outages. This report presents the methodology employed by the Dresden Database Research Group to forecast day-ahead electricity prices, which determine the energy prices for the next day upon which the companies can plan their production, as part of the BTW 2025 Data Science Challenge. Data from several sources, including weather and market data, was collected. We used Explainable AI methods to analyze the data and identify the most important predictors for day-ahead energy prices. This allowed us to gain deeper insights into the factors influencing electricity price predictions. Afterward, we compared several state-of-the-art time series forecasting methods on an extensive test set to identify the best among them. For this, we considered Chronos, Temporal Fusion Transformer, and AutoGluon, which are methods that were shown to perform very well on different kinds of time series datasets. The models were analyzed, and various approaches were explored to further enhance their performance, including fine-tuning, ensembling, architectural adjustments, and additional feature forecasting. After conducting a six-month benchmark comparing the final models, we concluded that a fine-tuned large Chronos model performed the best among all tested. With this approach, we outperformed the baseline on root mean squared error by a factor of 1.45 and a factor about 1.67 on mean absolute error.

Electricity price forecasting is crucial for market participants to optimize trading strategies and ensure grid stability. This project focuses on predicting next-day hourly electricity prices in Germany using a data-driven approach with machine learning models. The dataset spans the last two years from the current date and includes hourly electricity generation, consumption, and recorded day-ahead market prices.

To capture temporal dependencies, lag features are engineered, incorporating hourly and weekly trends. Extensive data preprocessing, including handling missing values, feature selection, and normalization, ensures model robustness. Various forecasting models, including Gradient Boosting, Random Forest, and Neural Networks, are implemented and evaluated using key performance metrics such as Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). Hyperparameter tuning and cross-validation are applied to enhance predictive accuracy.

The impact of renewable energy sources, market trends, and seasonal variations is analyzed to understand price fluctuations. The findings contribute to improving short-term electricity price forecasts, enabling energy traders, grid operators, and policymakers to make informed decisions, reduce financial risks, and enhance market efficiency.

This project leverages advanced machine learning techniques to support a sustainable and data-driven energy market, aligning with Germany’s transition towards a more renewable-centric energy system.

This project focuses on developing a machine-learning model to predict day-ahead electricity

prices in the German market. We investigate the impact of various factors, including historical

price data, weather conditions, and electricity generation and consumption. A comprehensive

data analysis is conducted to identify relevant features and potential correlations. The datasets

used also contained missing data which we have successfully processed. The dataset includes

electricity generation, consumption and price features from SMARD.DE and we have also in-

cluded weather features from home.openweathermap.org. After preprocessing of the dataset, we

have trained 2 models and compared the accuracy between them and chose the best performing

model to predict the hourly electricity price for 18.Feb.2025.

Challange.zip contains:

1) 2 files (cleaned_merged_data.csv and formatted_CO2_emissions.xlsx) with collected historical data.

2) 1 PDF file (Report_Data_Science.pdf) of our report.

3) saved csv files, which contain predictions (Linear_Regression_forecast_Feb18_2025.csv, XGBoost_Forecast_Feb18_2025.csv, Random_Forest_forecast_Feb18_2025.csv, LSTM_forecast_Feb18_2025.csv, Prophet_Forecast_2025-02-18.csv, SVR_forecast_Feb18_2025.csv)

4) 1 Jupyter Notebook file of our code.

This study presents a systematic comparison of three forecasting models—Linear Regression, Prophet, and XGBoost—for predicting day-ahead energy prices in Germany. Historical price data, weather features, and the renewable energy mix form the foundation of the predictive framework. Model performance, assessed via RMSE, indicates that integrating weather and energy mix features substantially enhances forecast accuracy. Despite Prophet’s aptitude for modeling seasonality, its accuracy declines without up-to-date energy mix forecasts for the day of predictions. Linear Regression provides interpretability but exhibits limited capacity to capture non-linear market interactions. XGBoost offers superior predictive performance, demonstrating robustness in handling complex feature dependencies. These findings underscore the importance of diverse inputs and advanced modelling approaches for reliable day-ahead energy price forecasting.