Developing machine learning algorithm for energy supply optimization in virtual power plants

Articles in Press

Document Type : Article

Authors

1 Department of Industrial Engineering, Faculty of Engineering, Yazd University

2 Department of Industrial Engineering, Yazd University

3 Department of Electrical Engineering, Yazd University, Yazd, Iran

10.24200/j65.2025.66704.2433

Abstract

The day-by-day increasing need for the generation and distribution of electric power all over the world creates the power market restructuring. Virtual power plants as a set of supply and demand units have a growing role in the power markets. The demand management as a new concept in the energy supply chain helps to improve the network reliability, reduce the operational costs, protect the environment, and increase the consumer satisfaction. In this research, MILP and machine learning approaches are proposed for the optimal unit commitment in a virtual power plant using a demand-side storage to reduce the gap between energy supply and demand. The proposed model has two objective functions: Maximizing the profits and minimizing the energy loss of hydroelectric power plant as well as the operational costs of pumps. Multiple scenarios based on the demand-sid storage, weather conditions, network grid status is studied. The numerical results for the IEEE standard 24-bus system show that although the use of demand-side storage increases the related cost as well as the generation fluctuation in the case of virtual power plant, the related costs are about 1.5 times lower than those of the traditional power plant without storage. In addition, comparing the traditional grid with a virtual power plant shows a 1.7-fold reduction in costs. However, comparing a virtual power plant including demand-side storage with a virtual power plant without demand-side storage shows an increase in costs along with an increase in energy production fluctuations. To solve the model in the real-world dimensions, a ELM-GA hybrid algorithm is proposed which provides a near optimal solution but much faster than the MILP method because the training and learning accelerate the convergence to the optimal point. The experimental results for a large-sized problem of 60 units demonstrate the small optimality gap of about 4.4% with the MILP but 16 times superiority in the solution time.

Keywords

Main Subjects

Sharif Journal of Industrial Engineering & Management

Articles in Press, Accepted Manuscript
Available Online from 19 October 2025

Files

Share

How to cite

Statistics