Forecasting GAS-CHP power plant generation: a comparative analysis of time series and machine learning models

Forecasting conventional power plant generation is crucial for enhancing the operational efficiency and sustainability of modern energy systems. This dissertation investigates the application of time series and machine learning models to predict the generation of a gas-fired Combined Heat and Power (CHP) plant in Germany, chosen due to the critical role of this kind of power plants in the transition toward a greener energy sector thanks to their high fuel efficiency. An aspect which distinguishes this study from the others present in literature is its focus on leveraging more accessible data, such as economic and weather variables, rather than highly technical or engineering-based features, which are often difficult to obtain. Sourcing data from the ENTSO-E Transparency platform, the analysis investigates forecasting techniques on historical generation data for a 206 MW gas-CHP plant. Traditional models, such as SARIMAX, are compared with advanced machine learning methods like Random Forest, LightGBM, and hybrid approaches combining Facebook’s Prophet and LightGBM. The forecasting methods are, then, tested on various temporal setups (e.g., weekdays versus weekends) and evaluated using metrics such as Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE). The results highlight the effectiveness of incorporating features such as electricity prices and renewable energy generation data in forecasting models. Additionally, the analysis demonstrates the value of splitting the data into weekday-weekend subsets compared to modeling the entire time series, revealing improved performance in capturing generation patterns. Key challenges, including data preprocessing, handling outages, and addressing generation intermittency, are also addressed in the dissertation.