The installation of large amounts of distributed photovoltaic and other distributed energy resources injects additional uncertainty to distribution networks, which poses challenges to the reliable and economic operations of distribution systems. To proactively address these challenges, this project brings together a team of experts in meteorology, solar power forecasting, machine learning, and big data analytics, to jointly address the challenges in multi-timescale probabilistic net load forecasting in distribution systems that have high solar penetration, thereby closing the technology gaps through the novel and transformational technical developments. In this project, we seek to develop a software tool for commercialization that further enhances the probabilistic solar forecasting skills on behind-the-meter solar generation, thereby helping the operation of a low-carbon grid and improve the reliability, efficiency, and resiliency of the nations power grid. This work provides an advanced reinforced machine learning-based physical-and datadriven forecasting application, which provides accurate, robust, and hierarchically consistent multiple look-ahead (e.g., from day-ahead, hours-ahead, to intra-hour) probabilistic solar and netload forecasts for distribution systems with high photovoltaic penetration and multi-hierarchical structures. The project will: (1) Implement the reinforced unsupervised/supervised machine learning-based physical and data-driven forecasting methodology; (2) Design and develop required interface and core models of the forecasting software tool; (3) Test and validate the software with a large-scale distribution system test bed using practical datasets. The project aims to tackle emerging solar forecasting challenges in distribution systems, including the complicated input space, the lack of robust and best-performing models, and the aggregate inconsistency. With the help of these data-driven algorithms, the forecasting and reliability of the system will be greatly enhanced. The advanced machine learning algorithms will be implemented in online and offline applications, with data preprocessing and visualization modules for the proposed project. The algorithms and software packages to be developed in this project are crucial in enhancing the monitoring, visualization, operation, and control applications.
