1. Introduction
Renewable energy power generation based on wind and solar energy has attracted considerable attention around the world [
1]. Nevertheless, the randomness, intermittence and periodicity of wind and solar energy challenge the stability of the power grid and power quality [
2,
3]. Recently, microgrids possessing advantages of flexible, modular and localized attributes emerged as a solution [
4,
5], but the output power of renewable energy generation and load consumption are inherently dynamic, making continuous balance impossible, which will influence the safe operation of microgrids. The application of energy storage systems (ESS) can effectively mitigate the variability of wind and solar power generation [
6,
7,
8]. However, single ESS often have inherent limitations [
6,
9]. For instance, vanadium redox flow batteries (VRBs) have high energy density but struggle with rapid charge/discharge cycles [
9]. Supercapacitors (SCs) excel in high-power, but their energy density is low [
10]. Hybrid energy storage systems (HESS) combining VRBs and SCs create a synergistic solution for applications requiring both long-duration storage and rapid power response [
11,
12].
To fully leverage the advantages of different energy storage components, optimal power allocation and capacity configuration are the cornerstones of HESS to ensure grid stability, efficiency and longevity [
13]. Scholars have extensively studied power and capacity allocation strategies to optimize performance, cost and reliability [
14,
15,
16]. The primary approaches include filtering decomposition [
17,
18], Fourier decomposition [
19], wavelet decomposition [
20,
21] and mode decomposition [
22,
23,
24]. Filtering decomposition such as low-pass filters and high-pass filters are generally used to allocate power in HESS for smoothing renewable energy fluctuations [
25,
26], but there is a delay in the low-pass filter process, which will distort power allocation in HESS, and lead to inefficient operation or even instability. Jiao et al. [
27] found that power allocation based on low-pass filters resulted in excessively large capacity values. The Discrete Fourier Transform (DFT) approach for HESS power allocation has a key limitation: loss of temporal information due to its frequency-domain nature [
28]. Yun et al. [
29] employed Wavelet Packet Decomposition (WPD) to split renewable energy output power into high-frequency and low-frequency components for optimal allocation between batteries and SCs. Nevertheless, the performance of WPD is highly sensitive to the choice of wavelet basis function [
21]. Additionally, wavelet analysis is computationally intensive, which can be a critical bottleneck for real-time HESS power allocation [
20]. Empirical mode decomposition (EMD) proposed by Huang could self-adjust according to the time-scale attributes of data for signal decomposition [
30]. Tian et al. [
31] and He et al. [
32] applied EMD to achieve the active power distribution, but EMD has a critical limitation of mode mixing, where components with similar frequencies are improperly separated during recursive decomposition [
33]. Furthermore, researchers developed ensemble-based EMD variants such as EEMD [
34], CEEMD [
35] and CEEMDAN [
36]. These methods can improve frequency separation for HESS by introducing controlled noise during decomposition, but EMD and the improved methods rely entirely on data-driven decomposition.
To solve the above problems, variational mode decomposition (VMD) excelling in local optimization was proposed by Dragomiretskiy [
37], which makes it particularly effective for handling anti-noise robustness and non-stationary burst signals while mitigating mode aliasing [
6,
38]. However, VMD is not model-adaptive, and the decomposition results heavily rely on the number of mode decomposition (K) and the penalty factor (α), which should be set in advance [
39]. This means that the performance of VMD heavily depends on the parameters of K and α. The traditional methods such as the maximum central frequency criterion were generally used to optimize VMD parameters [
38]. However, there are no clear measurement criteria; they are subjective. Several researchers proposed entropy-based optimizati [
6,
38] to automatically determine the optimal parameter for VMD [
40]. This approach minimizes the entropy measures that quantify mode compactness, sparsity, or information uniformity, leading to more adaptive and accurate signal decomposition [
40,
41,
42,
43]. Xiao et al. [
38] employed adaptive VMD to determine the decomposed mode number and grid-connected modes and allocated power to SCs and lead-carbon batteries accordingly. Zhang et al. [
44] used a genetic algorithm (GA) to optimize the selection of K and
α and improve the decomposition efficacy. Huang et al. [
45] and Gao et al. [
46] applied the whale optimization algorithm (WOA) and the sparrow search algorithm (SSA) to optimize the parameters of VMD for HESS power decomposition. Zhang et al. [
47] applied the improved pelican optimization algorithm (IPOA) to adjust the parameter of VMD, enabling accurate distribution of power for different storage types. However, algorithms like WOA, SSA, and IPOA often face challenges such as long computation times, convergence to local optima, or even failure to find viable solutions, prompting the need for more efficient approaches.
In contrast, the grey wolf optimizer (GWO) algorithm, proposed by Mirjalili et al. [
48], is less prone to falling into local optima and possesses high global search capabilities. With fewer parameters, GWO is easier to adjust and control, making it more convenient and efficient for practical applications. Elsaid et al. [
49] utilized GWO to address the challenges of Intrusion Detection Systems in efficiently handling redundant features, significantly boosting detection accuracy and reducing computational overhead. Nevertheless, basic GWO may suffer from premature convergence and sensitivity to parameter settings. To eliminate the limitations, Shaikh et al. [
50] used a chaotic map to improve the ability of GWO to find the best solutions and achieve faster convergence. The Improved Chaotic Grey Wolf Optimization (ICGWO) algorithm was applied to enhance wireless sensor networks coverage and connectivity. The findings confirmed that ICGWO efficiently improved the coverage and connectivity. Dagal et al. [
51] applied the Hierarchical Multi-Step Gray Wolf Optimization (HMS-GWO) algorithm to specific energy systems optimization problems. The results demonstrated that HMS-GWO effectively optimized the operation of the IEEE 30-bus system, demonstrating superior performance compared to other algorithms.
The random walk grey wolf optimizer (RW-GWO) algorithm combines the robust exploitation–exploration balance of GWO with enhanced global search capabilities through random walk dynamics, mitigating stagnation in local optima [
52]. Therefore, RW-GWO outperforms basic GWO and several established optimization techniques including Particle Swarm Optimization (PSO), Aquila Optimization Algorithm (AOA), Ant Colony Optimization (ACO), Spider Monkey Optimization (SMO), and Whale Optimization (WO) in terms of solution accuracy as well as convergence speed for complex multimodal problems, motivating its selection for this work.
In this paper, a RW-GWO-optimized VMD approach was developed for power allocation to mitigate wind–solar fluctuation-induced instability in wind–solar complementary microgrids. HESS capacity was determined based on the allocation strategy, with the optimization objective being whole life cycle cost reduction. To validate the proposed strategy’s feasibility, we conducted comparative analyses against EMD, PSO-VMD, AOA-VMD and GWO-VMD approaches. The principal contributions of this work are as follows:
- (1)
A power allocation strategy based on RW-GWO-VMD was proposed. Compared with PSO, AOA and GWO, RW-GWO demonstrated superior search speed and optimization accuracy, improving the mode-matching accuracy of VMD.
- (2)
Using the whole life cycle cost of HESS as the objective function, the dividing point N was used as the optimization variable to obtain the optimal capacity configuration results. The configuration costs for HESS are significantly reduced.
- (3)
An optimization model was established to determine HESS capacity requirements using normalized daily operational data. Case study analysis showed that the proposed approach reduced the whole life cycle cost of HESS by 7.44%, 1.00% and 0.72% compared with EMD, PSO-VMD and AOA-VMD, respectively, providing valuable engineering insights for HESS capacity planning.
The paper is structured as follows:
Section 2 presents the structure of microgrids,
Section 3 introduces the proposed power allocation strategy for HESS,
Section 4 establishes the capacity optimization configuration model for HESS,
Section 5 introduces the model solution methods,
Section 6 presents a comparative analysis of HESS power allocation strategies and capacity configuration results, and
Section 7 is the conclusion.
6. Example Analysis
A wind–solar hybrid farm in Northwestern China serves as a case study to validate the effectiveness of the RW-GWO-VMD strategy. The historical data from the wind–solar farm with the capacity of a 2.0 MW wind turbine and 1.5 MW solar as well as typical electrical load are employed as test data. The dataset comprises 1440 samples collected at an interval of 1 min over a 24-h monitoring period. Measurement data over 15 consecutive days during the latter half of June are shown in
Figure 4. The statistical analysis for the measurement data over 15 consecutive days is displayed in
Table 2.
As shown in
Table 2, for photovoltaic power generation, the average power remained at about 305 kW on D4, D8, D10 and D14, with stable maximum power (around 1050–1176 kW). However, the average power was similarly low on D12 (69 kW) and D15 (95 kW), possibly due to cloudy conditions. In addition, power data with high standard deviation of 417 on D1 indicated unstable sunlight conditions. For wind power generation, the average power was 1067 kW and 994 kW on D11 and D13, respectively, with maximum power approaching the limit of 1640 kW. Nevertheless, the average power was lower on D2 and D9 than that of other days, likely due to low wind speeds. Taking the maximum photovoltaic power generation and fluctuation into account, data collected on the eighth day were representative, but D13 was chosen as a typical day considering the maximum wind power generation. In this paper, the choice of typical day mainly referred to the maximum photovoltaic power generation data.
The wind and solar power generation as well as the power consumption curves of electrical load are shown in
Figure 5. As depicted in
Figure 5, the data exhibit significant fluctuation, indicating that the unstable wind and solar conditions are characterized by frequent wind speed variations and illumination variation. The relevant system parameters are shown in
Table 3.
6.1. Results of Power Allocation
A comparative evaluation of RW-GWO against AOA, PSO and standard GWO was conducted to demonstrate its superior performance in VMD parameter optimization. The fitness convergence curves of PSO, AOA, GWO and RW-GWO are shown in
Figure 6. From
Figure 6, it can be observed that RW-GWO has rapid convergence ability, achieving a fitness value of 9.9564 within 5 iterations, while the fitness function of GWO, AOA and PSO converges to 9.9564 (after 27 iterations), 9.9577 (after 8 iterations) and 9.9562 (after 9 iterations), respectively, indicating that RW-GWO could achieve fast convergence and find better results relative to GWO, AOA and RW-GWO.
To empirically demonstrate the superior performance of RW-GWO-VMD, methods such as EMD, AOA-VMD, PSO-VMD and GWO-VMD are used to decompose the power curve of HESS synthesized by the electricity load and the wind–solar power generation data for comparison. The variational nature of VMD intrinsically demands presetting the parameters of K and
α for its constrained optimization framework [
56]. Therefore, the parameters of VMD are optimized by PSO, AOA, GWO and RW-GWO in this paper, and the results are shown in
Table 4.
As shown in
Table 4, the RW-GWO optimization yielded the following VMD parameters: K = 11 and
α = 432.14. The IMF components decomposed by RW-GWO-VMD are shown in
Figure 7. From
Figure 7, it can be seen that the frequency of each mode function gradually decreases from IMF1 to IMF11, while the amplitude gradually increases. The decomposition results from EMD, PSO-VMD, AOA-VMD and GWO-VMD have similar rules, as depicted in
Figure 8,
Figure 9,
Figure 10 and
Figure 11. Every IMF component is capable of fully capturing the fluctuation features present in the original unbalanced power. Furthermore, the low-frequency modal component obtained by RW-GWO-VMD is more prominent and less volatile than EMD, PSO-VMD and AOA-VMD, which is suitable for the actual reference value of VRB energy storage frequency modulation.
The components obtained through EMD decomposition include multiple overlapping frequency elements (see
Figure 12), and the considerable spectral overlap makes it difficult to differentiate low-frequency from high-frequency elements. Conversely, as illustrated in
Figure 13 and
Figure 14, the VMD parameters optimized by PSO and AOA yield appropriate K and
α values that can partially mitigate modal aliasing. As a result, the IMFs exhibit clear frequency separation and uniform distribution. However, PSO and AOA have a higher tendency to converge to local optima during optimization compared with GWO and RW-GWO. Therefore, PSO-VMD and AOA-VMD decomposition lacks thoroughness, which may affect the HESS configuration results [
35]. The spectrum of IMFs from the GWO-VMD and RW-GWO-VMD algorithms showed clearly delineated frequency components (see
Figure 15 and
Figure 16), indicating successful mode separation.
VRBs as energy-type storage devices are suitable for low-frequency fluctuations, which exhibit small amplitudes but long durations. In contrast, SCs as power-type storage devices are suitable for high-frequency fluctuations, which possess larger instantaneous amplitudes but short durations. An inappropriate division point between high- and low-frequency components in the unbalanced power will impact the charge/discharge commands for VRBs and SCs and thus affect the results of capacity configuration. In this paper, the selection of the frequency-dividing point is based on the optimal whole life cycle cost of HESS as the objective, and the dividing point as the optimization variable.
As depicted in
Figure 17, the whole life cycle costs vary depending on the selected frequency boundary. Especially, the whole life cycle cost of HESS is the lowest when N = 1. Combined with the analysis of modal components obtained by RW-GWO-VMD, GWO-VMD, PSO-VMD, AOA-VMD and EMD in
Figure 7,
Figure 8,
Figure 9,
Figure 10 and
Figure 11, the amplitude of IMF1 vibration is much larger than that of the other modal components. Additionally, the frequency of IMF1 is low, which is in line with the storage characteristics of VRBs. For the IMF components obtained by RW-GWO-VMD, the center frequency of IMF1 is only 4.04% of that of IMF2. Therefore, the sub-modal IMF1 is reconstructed as the compensation power of VRBs, and IMF2~IMF11 are reconstructed as the compensation power of SCs.
Based on the previous HESS power distribution strategy, the power fluctuation characteristics of VRBs and SCs are determined through RW-GWO-VMD. Additionally, the results are further compared with those from GWO-VMD, AOA-VMD and PSO-VMD methods, as illustrated in
Figure 18 and
Figure 19.
The combination of modal decomposition and the whole life cycle cost optimization provides an excellent method for establishing frequency separation points and allocating HESS power. From
Figure 18, it can be seen that the power curves of VRBs among the four algorithms are rather similar. However, the power curves of VRBs obtained by GWO-VMD and RW-GWO-VMD exhibit greater fluctuation compared with the power curves based on AOA-VMD and PSO-VMD. The reason may be that RW-GWO-VMD and GWO-VMD yield less high-frequency residual in IMF1 compared to other decomposition methods. From
Figure 19, it can be seen that SCs, functioning as power-based devices, primarily absorb high-frequency power components. Compared with AOA-VMD and PSO-VMD, the power curves of SCs based on RW-GWO-VMD and GWO-VMD show low fluctuation amplitude and fast charging and discharging, indicating that RW-GWO-VMD and GWO-VMD could suppress mode aliasing and enable SCs to smooth high-frequency fluctuations.
6.2. Results of Configuration
Based on the results in
Section 4 and
Section 6.1, the capacity configuration of HESS was carried out.
Table 5 and
Figure 20 show the results of the configuration scheme with different methods, with reference to the economic parameters detailed in
Table 3.
Table 5 demonstrates that the rated power and rated capacity configuration of SCs derived from optimized VMD methods (PSO-VMD, AOA-VMD, GWO-VMD and RW-GWO-VMD) are reduced compared to those obtained by EMD. However, all the algorithms produce minimal variation in rated power and the rated capacity configuration of VRBs. The explanation lies in the enhanced ability of VMD to prevent modal aliasing compared to EMD. The bandwidth and frequency range of the modal function can be constrained through the introduction of the regularization term, which can differentiate the high- and low-frequency signals. As a result, more reasonable power allocation and accurate HESS capacity configuration can be realized. The whole life cycle cost of RW-GWO-VMD is 7.44% lower than that of EMD, 1.00% lower than that of PSO-VMD, 0.72% lower than that of AOA-VMD, and 0.27% lower than that of GWO-VMD.
As observed in
Figure 20, EMD and PSO-VMD exhibit longer error bars with significant fluctuations in objective function values. In contrast, AOA-VMD and GWO-VMD show shorter error bars, while RW-GWO-VMD demonstrates the shortest error bars and the most stable objective function values.
The convergence curve of the capacity optimization configuration model is shown in
Figure 21. Compared with the methods of PSO-VMD, AOA-VMD and GWO-VMD, RW-GWO-VMD exhibits the fastest convergence speed and the highest accuracy.
It is well known that EMD suffers from mode mixing issues, failing to adequately accommodate the characteristics of different energy storage devices. AOA exhibits premature convergence, resulting in local optimal solutions. PSO demonstrates weak global search capability and similarly tends to become trapped in local optima. GWO lacks sufficient search capacity, leading to slow convergence speed. In contrast, the random walk mechanism in RW-GWO significantly accelerates search capability. Therefore, RW-GWO-VMD delivers more accurate HESS power distribution than PSO-VMD, AOA-VMD and GWO-VMD approaches.
6.3. Comparative Analysis
The running time can reflect the computational complexity of the algorithm, and the reconstruction error indicates the accuracy. The running time and reconstruction error are shown in
Table 6. It can be seen that RW-GWO-VMD has the least running time and the smallest reconstruction error, which indicates its accuracy and low complexity.
Based on CEC 2022 functions (dimension = 10), a Wilcoxon signed-rank test between RW-GWO and GWO was conducted at a significance level of 5%. The results were evaluated based on the following criteria: (A) The observed difference is significant if
p-value < 0.05; (B) The observed difference is marginally significant if 0.05 <
p-value < 0.10. (C) The observed difference is not significant if
p-value > 0.10. The results are shown in
Table 7. When RW-GWO is superior to GWO, the conclusions are A+, B+ and C+. When GWO is superior to RW-GWO, the conclusions are A, B and C. From
Table 7, it can be seen that RW-GWO is better than GWO in 10 out of 12 functions and significantly better than GWO in 6 out of 12 functions.