Abstract
Condition-based and predictive maintenance enable early detection of critical system conditions and thereby enable decision makers to forestall faults and mitigate them. However, decision makers also need to take the operational and production needs into consideration for optimal decision-making when scheduling maintenance activities. Particularly in network systems, such as power grids, decisions on the maintenance of single assets can affect the entire network and are, therefore, more complex. This paper proposes a bi-level multi-agent decision support system for the generation maintenance decision (GMS) of power grids in an electricity market in the context of predictive maintenance. The GMS plays an important role in increasing the reliability at the network level. The aim of the GMS is to minimize the generation cost while maximizing the system reliability. The proposed framework integrates a central coordination system, i.e. the transmission system operator (TSO), and distributed agents representing power generation units that act to maximize their profit and decide about the optimal maintenance time slots while ensuring the energy balance. In the proposed bi-level approach, n upper levels and one lower level (i.e. sub-problems) are solved by the independent agents and by the TSO, respectively. We derive the optimal strategy of the agents that are subject to predictive maintenance via a distributed algorithm, through which agents make optimal maintenance decisions and communicate them to the central coordinator, i.e. the TSO. The TSO decides whether to accept the agents’ decisions by considering market reliability aspects and power supply constraints. To solve this coordination problem, we propose a negotiation algorithm using an incentive signal to coordinate the agents’ and the central system’s decisions, such that all the agents’ decisions can be accepted by the central system. We demonstrate the effectiveness of the proposed algorithm with reference to the IEEE 39 bus system.
Introduction
In a modern electricity market, maintenance scheduling is a tug-of-war between two conflicting interests. On one side, decentralized producers (agents) want to maximize their revenue by running their plants when prices are high and performing maintenance only when absolutely necessary—ideally just before a failure occurs. On the other side, the Transmission System Operator (TSO) must ensure the "lights stay on" by maintaining a constant balance between power supply and demand across the entire network.
Traditional maintenance is often "preventive" (fixed intervals) or "corrective" (fix it when it breaks). While Predictive Maintenance (PdM) uses sensors to estimate a machine's Remaining Useful Life (RUL), implementing it in a large-scale power grid is difficult. Centralized approaches require producers to share private financial and health data, which they are often reluctant to do. Decentralized approaches often fail to guarantee that the grid won't crash if too many producers decide to take their plants offline for maintenance at the same time.
Methodology
The authors propose a bi-level multi-agent decision support system that uses an iterative negotiation algorithm to bridge the gap between individual profit and grid reliability.
As illustrated in the framework above, the system operates through several key stages:
- Agent Level (Lower): Each power generation unit uses its own health monitoring data to predict its RUL. They solve a stochastic optimization problem to find the maintenance slot that maximizes their expected reward (revenue from power minus the cost of potential failure).
- TSO Level (Upper): The TSO receives these maintenance "bids." It runs a market-clearing model to see if the grid can still meet demand with those units offline.
- The Negotiation (Incentive Signal): If the TSO cannot accept a bid due to reliability risks, it doesn't force a change. Instead, it sends an incentive signal—essentially a financial penalty—that modifies the agent's objective function.
- Convergence: The agents recalculate their optimal slots based on this new signal. This process repeats until the agents' proposed schedules are compatible with the grid's requirements.
A standout feature of this mechanism is that it is "fully budget balanced" at the point of convergence. The signal acts as a coordination pressure that disappears once a feasible global solution is found. Furthermore, because the coordination happens via an incentive signal rather than direct control, agents do not need to share their private cost structures or detailed health data with the TSO.
Experimental Setup
The framework was evaluated using the IEEE 39-bus New England system, a standard benchmark in power systems research consisting of 39 buses, 29 lines, and 10 generating units. The study used a standard yearly load curve with a peak demand of 6254 MW and simulated 50 different scenario paths for the RUL of each generator to account for prediction uncertainty.
Key Results
- Rapid Convergence: In the case study, the algorithm reached a consensus in just three iterations.
- Conflict Resolution: The system successfully identified and resolved conflicts where specific agents originally wanted to perform maintenance during high-demand periods that would have compromised the grid.
- Superior Rewards: Compared to baseline strategies—like "Condition-Based Maintenance" or "Corrective Maintenance"—the proposed negotiation algorithm resulted in higher expected rewards for almost all agents.
- Uncertainty Handling: The study showed that as RUL prediction uncertainty increases, the total expected reward decreases by about 2%, highlighting the value of high-accuracy prognostic algorithms.
Conclusion and Future Work
While the results are promising, the authors acknowledge a few areas for growth. Current limitations include scalability, as real-world grids are significantly larger than the 10-generator test case. Future work will also need to address renewable integration, accounting for the high volatility of wind and solar sources, and more complex market dynamics where large entities might attempt to influence market prices through their maintenance decisions.
Summary updated