Data Driven Supervision and Forecasting for Distributed Energy Resources

Authors

Hugo Riggs, Florida International UniversityFollow

Document Type

Dissertation

Degree

Doctor of Philosophy (PhD)

Major/Program

Electrical and Computer Engineering

First Advisor's Name

Arif Sarwat

First Advisor's Committee Title

Committee Chair

Second Advisor's Name

Fahad Saeed

Second Advisor's Committee Title

Committee Member

Third Advisor's Name

Jean Andrian

Third Advisor's Committee Title

Committee Member

Fourth Advisor's Name

Sumit Paudyal

Fourth Advisor's Committee Title

Committee Member

Keywords

machine learning, photovoltaics, forecasting, ensemble learning, distribution power grid

Date of Defense

3-24-2023

Abstract

Global energy demand continues to grow and many grid operators are setting ambitious goals, in the years to come, to have high integrations of renewable energy sources such as wind and solar. The smart grid is the modernization of the electrical critical infrastructure which incorporates information technologies into the operational architecture over wide management areas. The communications capabilities of the smart grid can be used in managing distributed renewable energy assets. The requirements for managing these distributed resources include supervision and control capabilities, along with forecasting of load and generation for programming schedules.

This dissertation presents the research for the development of a forecaster of PV generation, and a computer cluster to run machine learning processes and create visualizations of the grid’s distributed assets. The PV forecast implements an estimation of AC power generated by a PV array. It applies an estimation model parameterized by the PV site-specific factors and the irradiance that the array will receive. The local irradiance, solar energy, is the independent variable that is forecasted. The ideal irradiance is exactly known by solar position modeling. However, the atmosphere interferes with the irradiance, and the amount of irradiance that reaches a PV array is reduced and often intermittent. The irradiance is forecasted through an ensemble model which performs well at next-day prediction with only 13% relative root mean square error.

The centralized supervision of grid assets is approached by the implementation of a computer cluster that incorporates various software frameworks and manages multiple computers. The computer cluster distributes data stream processing, handling incoming data sources, data storage/retrieval, and visualization. It is validated using the consumer electrical load data from the advanced meter infrastructure of south Florida. A K-means clustering is performed on individual meters. Each meter is assigned to a cluster. In evaluating this computer cluster it is used for calculating the day ahead consumer load. The load is forecasted by a classification and regression tree. The regression model is trained on meter cluster membership and local ambient temperature. Aggregated energy consumption for the neighborhood in a summer week can exceed 250,000 kWh. The 24-hour forecast has a 11.36 kWh mean absolute error. The contextualized AMI data is visualized in an energy dashboard. By forecasting load and distributed generation, an economic dispatch schedule for traditional generation sources is calculated. The dispatch scenario studied resulted in a reduced cost by $10,786 to run the daily dispatch of generators.

Identifier

FIDC011009

ORCID

0000-0001-6469-015X

Recommended Citation

Riggs, Hugo, "Data Driven Supervision and Forecasting for Distributed Energy Resources" (2023). FIU Electronic Theses and Dissertations. 5314.
https://digitalcommons.fiu.edu/etd/5314