Deep Learning Based Forecasting Studies for Real VPP Applications

Abstract

As renewable energy gains prevalence and machine learning technology advances, the need for accurate forecasting in Virtual Power Plants (VPPs) has become increasingly urgent, especially given the enhanced computational capabilities that make high-accuracy deep learning algorithms feasible, underscoring its importance for grid stability and efficiency. Considering these developments, this dissertation focuses on three key forecasting areas essential for VPP operations: solar photovoltaic (PV) output, load demand, and electricity prices. This thesis employ various deep learning algorithms to make these forecasts, each tailored to the specific characteristics of the data and the operational needs. For both solar PV and load demand, this thesis provide forecasts at two different time intervals: 1-hour and 5-minute. The 1-hour solar PV forecasts aim to provide a 24-hour outlook for day-ahead markets. The 5-minute forecasts are further divided into two categories: immediate next 5 minutes and up to 4 hours into the future. The rationale for these different forecasting horizons is twofold: The 5-minute forecasts are designed for real-time grid management, while the 4-hour forecasts align with the typical charge and discharge cycles of battery storage systems, thereby facilitating more effective battery scheduling. Electricity price forecasts are also made at 5-minute intervals for both immediate and medium-term scenarios. This research has been applied in a production simulation for a retail electricity company, demonstrating its practical utility. The integrated forecasting and optimization model developed in this study offers tangible benefits, enabling consumers to save on electricity costs while increasing profitability for the company.