Day-Ahead Demand Forecast Optimization for a Chemical Plant in the Netherlands

  • Sotirios Tsalikis

Student thesis: Master's Thesis

Abstract

This thesis presents a comprehensive approach to improving Day-Ahead demand forecasting accuracy for a chemical plant in the Netherlands, with an emphasis on streamlining the daily submission process and reducing imbalance costs. By integrating machine
learning techniques with operational data such as maintenance schedules and latest
power flow information, a robust forecasting tool has been developed using Excel and
VBA. The resulting solution not only enhances power consumption predictions but also
significantly boosts the efficiency of the submission process.
One of the main pillars of this thesis is to accurately predict the active power generation of a 40 MW rated Gas Turbine Generator. To achieve this several ML models were
trained. The research identified Polynomial Regression as the most suitable model for
integration into the Excel VBA tool, due to its strong predictive performance and ease
of implementation. The model demonstrated high accuracy, with a Mean Squared Error
of 0.3076 and a Coefficient of Determination of 0.8563. Despite challenges in predicting
the output solely based on ambient temperature, the model provided reliable forecasts,
highlighting though the importance of adding additional factors like fuel quality and
additional ambient conditions.
A major contribution of this thesis is the development of a user-friendly, interactive
Python Dashboard using the Dash framework. This application allows the end-user to
explore two main forecasting methodologies that were developed during this research,
apply several improvements, and visualize the impact of these changes on forecasting
accuracy and imbalance costs. The dashboard proved invaluable in demonstrating the
effectiveness of the proposed “Latest Known Power-flow” methodology, which, when
combined with the GTG model and operational data, reduced imbalances in MWh
by approximately 86.44% and associated costs by 85.40%. This approach is
projected to save the plant an estimated 1.3 million euros annually.
This thesis contributes to the field of Energy Informatics by demonstrating the effectiveness of combining machine learning models with operational data to improve load
forecasting accuracy in industrial settings. Future work should explore the inclusion of
additional variables such as relative humidity, ambient pressure and fuel composition
to further enhance the GTG model accuracy. Additionally, expanding the dashboard
to serve as a Digital Twin for automating the daily forecast submissions could further
streamline processes and reduce manual labor.
Keywords: Load Forecasting, Chemical Plant, Imbalance Settlement, Day-Ahead market, Python, Machine Learning
Date of Award2024
Original languageEnglish (American)
SupervisorStephan Selinger (Supervisor) & Armin Veichtlbauer (Supervisor)

Cite this

'