1. Introduction
With the global transformation of the energy structure and the promotion of low-carbon goals, the proportion of renewable energy in power systems is gradually increasing. However, the intermittency and uncertainty of renewable energy pose significant challenges to the stability and economy of power systems [
1]. Coal-fired units, as conventional base-load power sources, possess strong regulation capabilities. They can quickly respond to load changes, thus becoming an important resource for peak regulation in power systems [
2]. However, coal-fired units operate in off-design conditions during peak regulation, leading to frequent fluctuations in unit parameters. The heat surfaces in the furnace are often subjected to thermal stress, making them more prone to failure [
3]. This results in unplanned outages and the loss of load resources and has a significant impact on the reliability of power system operation. Furthermore, the proportion of adjustable capacity in conventional thermal power units is too low, and the space for renewable energy absorption is severely insufficient, leading to phenomena such as wind and solar curtailment. Therefore, it is of great significance to effectively assess the impact of coal-fired units’ peak regulation characteristics on power system reliability after the integration of renewable energy. This helps to reasonably arrange generation plans and promote the integration of renewable energy.
PPS is a crucial tool for generation planning and unit operation scheduling in power systems. This method enables the calculation of unit output across different operational modes, along with system production costs and reliability indices. It is particularly useful for studies focusing on generation unit output characteristics and system load curves. PPS methods are generally classified into two main categories: simulation-based approaches [
4,
5] and analytical methods [
6,
7,
8]. The simulation methods, such as Monte Carlo simulation, have the advantages of strong adaptability and simple models, but they suffer from issues related to computational efficiency and accuracy. Analytical methods, with clear physical concepts, high mathematical model accuracy, and fast computation speeds, are suitable when many related factors need to be considered. However, they involve complex mathematical modeling, and it is difficult to obtain the probability distribution and characteristic parameters of various technical and economic indices. Currently, analytical methods remain the core of research on PPS in power systems. Scholars worldwide have conducted extensive work to improve the computational efficiency and accuracy of PPS algorithms. For instance, study [
9] integrates transition frequency analysis into the PPS framework, refining the equivalent electrical energy frequency method to assess the impact of wind power integration on dynamic costs. Another study [
10] introduces a PPS method based on the UGF theory, effectively capturing the reliability implications of wind power integration. Furthermore, study [
11] presents a structural expansion method that models state variable distributions in power systems, incorporating clean energy penetration and low-carbon benefit analysis. While extensive research has addressed the impact of renewable energy intermittency [
12] on system reliability, fewer studies have examined the degradation and operational dynamics of coal-fired units during peak regulation and frequency control. A study [
13] proposes a PPS-based evaluation framework for deep peak regulation of thermal power to facilitate renewable energy integration. However, it does not account for failure rates or state transition characteristics during operation. Another study [
14] develops four multi-state unit models considering various operational conditions, base-load constraints, and segmented unit capacities. Nonetheless, its complex model structure limits computational efficiency and fails to accurately capture the degradation effects of peak regulation on system reliability.
UGF serves as a robust mathematical tool for modeling and analyzing multi-state systems by representing system states and their corresponding probabilities in polynomial form. The fundamental principle of UGF lies in mapping discrete system states into a generating function, where each polynomial term characterizes a potential system state with its coefficient denoting the occurrence probability of that state. In the field of reliability modeling for power systems, UGF has emerged as a critical methodology that enhances model clarity and computational efficiency. A study [
15] leverages the UGF model for the reliability analysis of hybrid uncertainty structures in turbine systems, demonstrating a nearly 18-time improvement in computational efficiency compared with Monte Carlo simulations. Another study [
16] integrates UGF with Bayesian methods for the reliability assessment of complex multi-unit power generation systems, achieving significant efficiency optimization. Furthermore, researchers [
17] have employed UGF techniques combined with Weibull distribution analysis to evaluate the operational reliability of wind turbine components, providing a probabilistic framework for failure mode assessment in wind power generation systems.
Probabilistic models serve as critical tools for describing and analyzing system behaviors under uncertain operational conditions, particularly in evaluating key performance indicators such as operational reliability and risk profiles of power systems. The development of component-level state probabilistic models forms the foundation for such analyses. Researchers [
18] have established a probabilistic reliability model for microgrids and develops an analytical framework for assessing the reliability of distribution systems with microgrids. The proposed microgrids probabilistic model can simplify and accelerate computational processes associated with reliability evaluation of distribution systems incorporating multiple microgrids. For coal-fired power plants, researchers [
19] have developed accident probabilistic models to assess operational risks through dynamic likelihood evaluation of turbine generator contingencies. Additionally, in smart grid applications, an Electric Vehicle Virtual Energy Storage System charging/discharging power probabilistic model has been proposed to enhance the reliability assessment of distribution networks incorporating virtual energy storage technologies [
20]. These implementations demonstrate the adaptability of probabilistic modeling in addressing diverse challenges across renewable energy integration, conventional power generation, and modern grid management.
To fully consider the impact of coal-fired unit peak shaving characteristics on the reliability of the power system, an improved PPS method is proposed in this study, inspired by the multi-state unit model in the literature [
21]. This method first rigorously considers the temporal fluctuations of load demand and renewable energy. Secondly, a multi-state model for coal-fired units is established, incorporating their life degradation and failure rates under different output states. The transition rates between unit states are dynamically adjusted based on load demand and operational status. Similarly, a multi-period equivalent multi-state model is developed for wind power generation. Finally, the UGF algorithm is employed to perform joint calculations on the probabilistic models of different units, enabling the PPS. This approach allows for the computation of system reliability indicators, dynamic costs such as unit start-up and shut-down, and the evaluation of the system’s ability to accommodate wind power. This study investigates the influence of different peak shaving capabilities of units on the reliability of the power system. The proposed method is then verified through case analysis using the IEEE-RTS79 test system.
4. PPS Considering the Load-Following Characteristics of Coal-Fired Units
PPS in power systems is an effective tool for calculating the generation output of each unit, system generation costs, and reliability indicators. In this section, by considering the different state transition characteristics of units under varying load-following capabilities, the principle of prioritizing the operation of thermal power units with lower unit fuel costs is applied to arrange the operation of thermal power plants sequentially. Then, using the multi-state unit models established earlier, a joint calculation is performed using the universal production function. After all units are put into operation, the relevant technical and economic indicators are further calculated.
4.1. Load-Following Capability Constraints for Coal-Fired Units
This paper primarily focuses on the limitations of a unit’s load-following capacity when responding to load demand, studying the state transition characteristics under different load-following capabilities. Considering that the equipment of the unit is often subjected to high-temperature environments during actual operation, the reliability of high-temperature heat surfaces generally represents the reliability of the entire unit [
30]. Therefore, this paper derives the relationship between load change rate, peak shaving depth, and the lifetime of high-temperature heat surface pipelines and uses the remaining life of high-temperature equipment to correct the unit’s mean time between failures, thus calculating the failure rate of coal-fired units under different load-following depths and change rates.
4.1.1. Fault Probabilistic Model of Coal-Fired Units
This paper will use a multilayer perceptron (MLP) to fit the relationship between the variable load rate, peak shaving depth, and the remaining life of the high-temperature heat-exposed surfaces in coal-fired units. MLP is a feedforward neural network model, and its hidden layers are capable of extracting nonlinear features from the input data and performing representation learning [
31]. Let the variable load rate and peak shaving depth be denoted as
and
, respectively. The fitting process is as follows:
where
is the input matrix of the
n-th layer network with a dimension of
, where
n is the number of layers in the network;
is the number of neurons in the hidden layer of the
n-th network. Specifically, for the first layer, the input is
;
is the weight matrix of the
n-th layer with a dimension of
;
is the bias matrix with a dimension of
; the actual predicted value of the network is denoted as
;
is the activation function, and
is the weight matrix of
.
The relationship between the variable load rate, peak shaving depth, and the remaining life of high-temperature equipment is fitted into a nonlinear function using MLP, as shown in Equation (22). Based on the remaining life of the high-temperature equipment, the MTTF of the unit is then corrected. Assume that MTTF of unit
i is
, and its correction is given by Equation (23).
where
is the design remaining life,
is the corrected MTTF.
4.1.2. Unit Start-Up and Shut-Down Time Constraints
Assume that the system’s normal operation time and shut-down duration follow exponential distributions with the parameters
and
, respectively. The minimum continuous operation time and minimum economic shut-down time for a generator unit are denoted as
and
, respectively. Then,
where
is the probability that the operation time in state k exceeds the minimum continuous operation time
, and
is the probability that the shut-down time of the generating unit moving to state k exceeds the minimum economic shutdown time
.
Considering the start-up and shut-down time constraints of the unit, the demand
and non-demand rates
in the flexible peak shaving models should be corrected as follows:
4.1.3. Unit Variable Load Time Constraints
The ability of a unit to transition from one state to another during its operation depends on its ramp rate and variable load time. Given that the simulation time step is 1 h, the equivalent output of unit
i at different times during the simulation period must satisfy the following constraint:
where
is the equivalent output of unit
i at time
t, and
is the ramp rate of unit
i, expressed in %/min.
Based on this constraint, the equivalent output of unit
i is corrected as follows:
4.2. UGF Model of Coal-Fired Unit
Based on the three unit models described in
Section 2.2, in actual power system operation, units must flexibly adjust their output levels in response to load demand to accommodate instantaneous load fluctuations while meeting generation requirements as well as facilitating the integration of wind and photovoltaic power. This process requires careful consideration of the unit’s peak shaving capacity and output range to ensure the stable operation of the system.
Furthermore, to ensure the long-term reliability of the units, maintenance personnel typically develop detailed maintenance plans based on the operational status of the units, which is also one of the key factors in ensuring stable unit operation. Therefore, when performing system simulations, the maintenance schedule must be incorporated into the model to enhance the accuracy of the simulation results.
Based on the above factors, assume that coal-fired unit
i has
k states, and its UGF model is as follows:
where
represents the UGF model of unit
i,
and
represent the output and corresponding probability of unit
i at time
t, respectively.
is the parameter reflecting the planned maintenance of the unit. If the unit is in maintenance at time
t, then
; otherwise,
.
4.3. Power System Output State Adjustment
In the PPS of the power system, the entire system’s generation capacity is treated as a multi-state system. By introducing the UGF algorithm into the simulation process, the system’s output state is continuously adjusted as units are put into operation to reflect the simulation process. In particular, wind power, being a clean energy source, is considered with limitations in the PPS, while the remaining conventional units are arranged for operation based on their unit generation cost. Specifically, this paper primarily investigates the impact of coal-fired unit peak shaving characteristics on the reliability of the power system, focusing on the operation of coal-fired units.
Assuming that the unit commissioned in the
i-th operation is unit
i, after it is brought online at time
t, the system output model
is adjusted as per Equation (31).
where
represents the number of output states of the power system after the commissioning of unit
i,
and
represent the output magnitude and output probability of the system’s
s-th state at time
t after the commissioning of unit
i, respectively.
The PPS based on the UGF model can calculate the reliability indices at different times during the simulation period as well as after the commissioning of different units. Assuming the load demand at time
t is
, the loss of load probability (LOLP) and expected loss of load (EENS) for the entire power system after the commissioning of the
i-th unit at time
t can be calculated, along with the expected power generation of the currently operating units. The specific calculations are as follows:
where
represents the probability that the system output is less than the load demand on a given day, and
represents the expected value of the system output falling below the load demand.
is the expected generation of unit
i.
After the commissioning of generator i, if the power supply cannot meet the demand, generator
i + 1 will be commissioned. Therefore, the demand rate for the commissioning of unit
i + 1
can be represented as
:
4.4. PPS Process
A flowchart of the method is shown below. Wind power units are prioritized for commissioning, followed by the sequential commissioning of coal-fired units. The constraints and correction methods required in the simulation process are described in the previous sections. Unlike traditional convolution-based calculations, the simulation method in this paper employs a joint distribution calculation based on the UGF model. The specific process is shown in Algorithm 1.
Algorithm 1: PPS considering coal-fired unit peaking characteristics. |
Input: Number of rounds T, number of unit I, time-series load data L, wind resource data W, MTTF, MTTR, Ramp rate , peak shaving depth Initialization: |
1. | Model wind power output using a multi-state approach | //as per Equation (7) |
2. | Rank units by generation cost and set commissioning order |
3. | Obtain the remaining lifetime fitting model. | //as per Equation (22)Start |
4. | for t = 1 … T do |
5. | Initialize system capacity based on wind power state at t |
6. | for i = 1 … I do |
7. | if unit i under maintenance then | //as per Equation (30) |
8. | skip unit, set i = i + 1 |
9. | else continue |
10. | end if |
11. | if unit i in baseload operation then |
12. | use two-state mode | |
13. | else use multi-state model | |
14. | update coal-fired unit state space | //as per Equation (13) |
15. | compute unit steady-state probability | //as per Equation (16) |
16. | compute equivalent output model | //as per Equations (17)–(19) |
17. | end if | |
18. | while unit output does not meet the ramping constraints |
19. | adjust the unit output |
20. | end while | //as per Equations (28)–(29) |
21. | update the system output distribution | //as per Equation (31) |
22 | compute the expected power generation of the i-th unit//as per Equation (34) |
23.. | end for |
24. | compute the reliability indices at time t. | //as per Equations (32)–(33) |
25. | end for |
26. | compute the system reliability indices. | //as per Equations (36)–(47) |
End | | |
6. Case Study
To further verify the effectiveness of the algorithm in simulating the peak shaving characteristics of coal-fired units and the integration of wind power in large-scale systems, the case study uses the IEEE-RTS79 test system, which consists of 32 units with a total installed capacity of 3405 MW [
32]. The operating parameters and corresponding costs of the generators are shown in
Table 2. All power generation units, except for nuclear plants, are replaced with thermal power units. The maximum annual load is 2850 MW.
6.1. Coal-Fired Unit Peak Shaving Fault Model
The relationship between peak shaving depth, load variation rate, and unit lifespan is fitted using the MLP algorithm. The hyperparameters are set as follows: two hidden layers are set, each with 20 neurons; the activation function is rectified linear unit; the initial learning rate is 0.01; the optimizer is Adam optimizer; the maximum number of iterations is 100; and the batch size is 200. Using these hyperparameters, the simulation data from the literature are fitted, with a final fitting degree of 0.9971.
6.2. Algorithm Verification
When the peak shaving depth of the unit model is set to 0%, the unit model becomes a two-state model. To verify the rationality of the proposed method, three methods are compared for the original system and the original load curve: (1) the equivalent load duration curve method; (2) the simplified model of the proposed method (0% peak shaving depth); and (3) the proposed method (2% ramp rate, 60% peak shaving depth). The simulation results are shown in
Table 3.
According to the results in
Table 3, when the peak shaving depth is set to 0% and internal constraints such as ramp rates are not considered, the algorithm proposed in this paper yields simulation results similar to the traditional equivalent load duration curve method, and the computational efficiency of the proposed method is improved by about two times, which verifies the effectiveness of the proposed algorithm. Further analysis shows that when the peak shaving depth is set to 60%, the ramp rate is 2%, and internal unit constraints are considered, the power generation of the base-load unit is similar to that of the traditional model. However, the power generation of the flexible peak shaving units shows greater fluctuations compared to the traditional simulation methods, while the power generation of the start–stop peak shaving units is generally higher than that of the traditional methods. This is because the flexible peak shaving units, when responding to load demands, take into account their dynamic adjustment characteristics, adjusting the output flexibly according to actual needs. However, due to constraints on the unit parameters, state transitions are limited, leading to larger fluctuations in power generation. Additionally, due to the limitations of ramping capacity, peak shaving units cannot respond quickly to load changes, forcing start–stop peak shaving units to respond more frequently, which results in higher power generation compared to the units in the traditional method.
Furthermore, from the perspectives of LOLP and EENS, the values obtained using the proposed method are higher than those calculated by traditional simulation methods. This can be attributed to the fact that our method incorporates the impact of losses occurring during flexible deep load-following operations, thereby better reflecting the operational constraints faced by generating units in real-world scenarios. Unlike traditional models, which assume that units consistently operate at their rated output, the proposed approach fully captures the dynamic adjustments of units during load-following operations. As a result, the obtained reliability indices more accurately align with actual system behavior.
6.3. Impact of Different Peak Shaving Schemes
This section analyzes the impact of unit peak shaving operation on the system under different peak shaving schemes. The range of the unit’s load change rate is set from 1% to 5%, and the peak shaving depth is set to 90%, 80%, 70%, 60%, 50%, and 40%. Based on the integration of a 300 MW wind turbine, different peak shaving schemes are obtained by pairing the various peak shaving parameters. PPS is then conducted to study the system’s reliability, economic performance, and renewable energy absorption capacity.
Analysis of Simulation Results Based on LOLP, Expected Wind Curtailment, Start–Stop Counts, and Generation Costs: As shown in
Figure 6, under varying ramp rates, the LOLP increases with deeper peak shaving depths. This is due to the fact that as the peak shaving depth increases, the units operate outside of their design conditions, significantly affecting their reliability. Additionally, it can be observed that when the ramp rate is between 2% and 3%, the LOLP decreases. This is because an increase in ramp rate allows the units to respond to load demand more promptly while ensuring reliable operation. However, as the ramp rate rises from 4% to 5%, the LOLP increases again. The larger fluctuations in output at higher ramp rates can severely compromise the unit’s operational reliability, reduce its lifespan, and consequently affect the overall system’s reliability.
As shown in
Figure 7, the expected wind curtailment decreases with an increase in peak shaving depth. This is because a deeper peak shaving depth creates more space for wind power absorption, effectively facilitating the integration of renewable energy. It is also observed that when the ramp rate is between 2% and 3%, compared to 1%, the wind curtailment is relatively alleviated. The flexibility in the ramping capability of the units is an important factor in ensuring wind power absorption. However, when the ramp rate is set to 4% to 5%, the expected wind curtailment increases again. The severe fluctuations in load demand at higher ramp rates significantly impact the unit’s thermal operation, thereby hindering the absorption of wind power.
With an increase in peak shaving depth, the start–stop frequency of the units also rises. The larger the peak shaving depth, the greater the load variation range of the units, making it more difficult for flexible peak shaving units to complete peak shaving tasks within the required time. As a result, start–stop peak shaving units must respond to load demands by starting and stopping promptly. It can be observed from
Figure 8 that the start–stop frequency is minimized when the ramp rate is between 2% and 3%, as the units can maintain both flexibility and reliability at this ramp rate. Moreover, before reaching a 50% peak shaving depth, the start–stop frequency is the lowest at a 4% ramp rate. At this point, the load adjustment range of the units is relatively small, and the 4% ramp rate enables the units to adjust flexibly within a certain range, thus reducing the frequency of starts and stops.
As illustrated in
Figure 9, the generation cost of the system decreases with an increase in the peak shaving depth of the random units. This is because, on one hand, the units can operate at a reduced output to meet load demands, and on the other hand, the higher peak shaving depth leads to an increased failure rate, forcing the units to operate at a lower output. Additionally, before reaching a 65% peak shaving depth, the generation cost increases with the ramp rate, as the units operate in a high-load adjustment state. A higher ramp rate results in increased fuel consumption, thereby raising the generation cost. However, when the peak shaving depth exceeds 65%, an excessively high ramp rate reduces the unit’s reliability, decreases generation output, and increases dynamic costs, ultimately leading to a reduction in the generation cost.
Based on the above analysis, the system operates most optimally when the ramp rate is between 2% and 3% and the peak shaving depth ranges from 55% to 65%. This configuration closely aligns with practical scenarios, thus validating the reasonableness of the proposed algorithm.
7. Conclusions
In this study, we proposed an improved PPS method that integrates coal-fired unit degradation dynamics and wind power fluctuations within a multi-state reliability framework. Unlike traditional methods, our approach dynamically updates state transition probabilities based on operating conditions, ramping constraints, unit degradation, and load demand, leading to a more accurate and computationally efficient reliability assessment. The UGF framework was employed to optimize computational efficiency and overcome state-space explosion issues.
The simulation results of the IEEE-RTS79 test system demonstrate that the proposed method more accurately reflects real-world operational conditions. For conventional generating units, system reliability is improved when the ramp rate is set between 2 and 3% and the load-following depth is maintained at 55–65%. However, if the ramp rate is increased from 1% to 5%, the LOLP increases by 48%, confirming that forcibly enhancing a unit’s load-following capability leads to a higher failure rate, thereby exposing the entire system to greater risk.
While this study focuses on wind power integration, the method can be extended to solar power and hybrid renewable systems. Future research will explore economic factors, electricity market pricing, and further real-world validations. This approach provides a more practical and efficient tool for reliability assessment in modern power systems.