INTELLIGENT ENERGY TRADING AND DYNAMIC LOAD BALANCING IN DISTRIBUTED SOLAR PV NETWORKS USING REINFORCEMENT LEARNING
Faisal Aburub, Ibrahim Mohammad, Asokan Vasudevan, Yang Chunwen, Ulug'bek Sobirov, Linara Karimova
Abstract: The research will focus on the design of an intelligent system combining peer-to-peer trading and dynamic load balancing in distributed solar networks with the use of multi-agent deep reinforcement learning. The presented solution will include two layers – the first one will be a peer-to-peer marketplace operating under the multi-agent Proximal Policy Optimization algorithm with a centralized learning process and decentralized actions which optimize energy trades between energy producers and consumers. The second layer will be an adaptive load balancing controller which will adapt the proposed transactions according to the technological limitations of the grid. The architecture was developed and applied using real-world data for a 32-node power grid that had 18 producers and consumers having photovoltaic and battery storage in the Abdoun Amman area, which is the capital of Jordan, between January and September 2025. According to the results, the MAPPO algorithm proposed in this study performed better than the standalone PPO and the central DQN algorithms by 23.5% and 51.4%, respectively, with a cumulative reward of 4,872. The new setup resulted in an increase in daily producer-consumer profit of 98.6% to 2.84 Jordanian Dinars. A Gini Coefficient value of 0.28 shows the fairly equitable benefit distribution among participants. In terms of a technical approach, the suggested method has led to a reduction in voltage deviation to 1.82 percent, i.e., 42.2 percent lower than the two-way auction approach. The load balancing index also has been improved to 1.14 by 17.4 percent. Peer-to-peer monthly trading volume for the proposed case was estimated to be 14.82 MWh. Through sensitivity analysis, it is shown that the availability of battery energy storage systems plays a significant role in the success of the market and has the potential to lower the payback period from 8.5 to 5.2 years.
Keywords: Peer-to-peer energy trading, Multi-agent reinforcement learning, Dynamic load balancing, Distributed photovoltaic networks, Managing local power supply
DOI: 10.24874/PES08.03B.010
Recieved: 13.05.2026 Revised: 20.08.2026 Accepted: 16.09.2026
UDC:
Reads: 8 
