Swiss Grid Load Forecasting: How Machine Learning Reshapes Energy Management

A bachelor's thesis project demonstrates how to use machine learning models to predict energy demand for the Swiss grid. By combining historical data with weather forecasts, it achieves an 8-week rolling forecast. It compares classic time series models with supervised learning models and provides actionable technical solutions for the energy industry. The project was published on GitHub by Omar Abdesslem in 2025. Online demo link: https://energy-forecast.netlify.app

The core challenge of power systems is balancing supply and demand. Traditionally, this relied on dispatchers' experience, but the intermittency of renewable energy increases prediction complexity. As a key node in the European grid, Switzerland's energy consumption is affected by seasons, weather, and economic activities. Accurately forecasting load for the next few weeks is crucial for grid stability.

Data sources include historical Swiss grid data (Swissgrid weekly aggregated consumption records) and temperature data from the European Centre for Medium-Range Weather Forecasts (ECMWF). Feature engineering constructs a multi-dimensional feature set: lagged consumption values, rolling statistics, calendar seasonal features, and temperature variables to capture periodic patterns and weather sensitivity.

The model system covers classic time series models (seasonal naive, AR, ARIMA/SARIMA, SARIMAX) and supervised learning models (Ridge, Random Forest, XGBoost). The final model used is SARIMAX(1,0,0)(1,0,0,52), which integrates temperature exogenous variables. The rolling forecast mechanism retrains the model weekly, fits it with the past two years of data, predicts the next 8 weeks, and residual diagnostics verify the model's effectiveness.

The frontend uses Next.js + Tailwind CSS for responsive layout, and Plotly.js to render interactive forecast charts. The backend uses Python for offline model training and data processing. The separated frontend-backend architecture ensures computational efficiency and user experience.

Model selection requires trade-offs: classic statistical models (e.g., SARIMA) are robust when data volume is limited, while machine learning models (e.g., XGBoost) capture non-linear relationships when features are abundant. Integrating weather data significantly improves accuracy, and this approach can be extended to industries like retail and transportation. The rolling update mechanism is key to maintaining prediction accuracy.

This project transforms academic research into an interactive practical tool, providing a complete forecast system paradigm (data integration, feature engineering, model selection, continuous update). It offers a reference implementation for time series learners and energy industry developers.