Capacity Configuration Optimization of Wind–Light–Load Storage Based on Improved PSO

Abstract

To improve the economy and stability of data center green power direct supply, the capacity configuration optimization of wind–light–load storage based on improved particle swarm optimization (PSO) is conducted. According to wind speed, the Weibull distribution of wind output is established, while the Beta distribution of solar output is established according to light intensity. Furthermore, by conducting the correlation analysis, it is indicated that there is a negative correlation between wind and solar output, which is helpful to optimize the mix of wind and solar output. To minimize the yearly average cost of wind–light–load storage, the capacity configuration optimization model is established, where the constraints include wind and solar output, energy storage capacity, balance between wind and solar output and data center load. To solve the capacity configuration optimization model, the improved PSO is adopted, compared to other optimization algorithms, like differential evolution (DE), genetic algorithm (GA) and grey wolf optimizer (GWO); by adjusting the inertia weight factor dynamically, the improved PSO is more likely to escape the local optimal solution. To validate the feasibility of data center green power direct supply with wind–light–load storage, a case study is conducted. By solving the capacity configuration optimization model of wind–light–load storage with the improved PSO, the balance rate between wind–solar output and data center load is improved by 12.5%, while the rate of abandoned wind and solar output is reduced by 17.5%, which is helpful to improve the economy and stability of data center green power direct supply.

1. Introduction

With the development of data processing technologies, data centers are becoming more and more significant. However, to reduce the carbon footprint, data centers face enormous pressure to save energy and reduce emissions. To ensure data centers are economically viable and environmentally sustainable, China launched the strategic project East Data West Calculation. To utilize abundant renewable resources, developing data center green power direct supply is straightforward. However, to ensure the economy and stability of data centers, there are two issues to be considered. Firstly, due to the randomness of wind speed and light intensity, the fluctuation of wind and solar output is strong. Secondly, data center load is affected by computing tasks significantly; namely, during peak load periods, wind and solar output is insufficient, while during low-load periods, extra wind and solar output is abandoned. To weaken the imbalance between wind–solar output and data center load, energy storage devices are adopted to adjust the imbalance, improving the economy and stability of data centers.

To improve the economy and stability of data centers, it is necessary to analyze the randomness of wind and solar output and the fluctuation of data center load [1,2,3,4,5,6,7,8]. Li et al. [1] adopted particle swarm optimization to reconstruct load-storage systems, analyzing the randomness of wind and solar output. Liang et al. [2] compared different indicators and methods to evaluate the confidence level of solar output, presenting measures to improve the confidence level of solar output. Based on Copula theory, Zhang et al. [3] conducted correlation analysis between wind and solar output. Zhou et al. [4] predicted power flow of load-storage systems with the adaptive support vector machine model.

Due to the development of distributed energy technologies, studying the capacity configuration optimization of wind–light–load storage is necessary [9,10,11,12]. However, the randomness of wind and solar output is strong, so correlation analysis between wind and solar output is absent. Therefore, to improve the economy and stability of data centers, the present work conducts correlation analysis between wind and solar output, which is helpful to optimize the mix of wind and solar output. To solve the capacity configuration optimization model, different optimization algorithms are developed, including differential evolution (DE), genetic algorithm (GA), grey wolf optimizer (GWO) and particle swarm optimization (PSO). The GA is dependent on the computationally expensive crossover and mutation operation, leading to slower convergence. Though the DE is highly robust for the complex problem, careful parameter tuning is required, which is difficult to implement. As a more recent nature-inspired algorithm, the GWO demonstrates strong exploration; however, it suffers from premature convergence. Comparatively, due to its computational efficiency and straightforward implementation, the PSO offers a unique balance between convergence and robustness, making it particularly suitable to solve the capacity configuration optimization model. Besides traditional optimization algorithms, the data-driven method has potential to optimize the operation of the multi-energy microgrid. For example, Jia et al. [13,14] designed a reinforcement learning-based method to obtain the optimal solution of the multi-energy mircrogrid. The present work is organized as follows. Firstly, we introduce characteristics of wind and solar output and data center load, and the correlation analysis between wind and solar output is conducted. Based on historical data, we establish probability distribution models of wind and solar output and data center load. Then, the capacity configuration optimization model of wind–light–load storage is established; to solve the capacity configuration optimization model, the improved PSO is adopted. Finally, to validate the feasibility of data center green power direct supply with wind–light–load storage, a case study is conducted.

2. Characteristics of Wind–Light–Load Storage

Firstly, we introduce the characteristics of wind and solar output and data center load; then, the correlation analysis between wind and solar output is conducted. Due to the randomness of wind speed and light intensity, the fluctuation of wind and solar output is strong. Based on historical data of wind speed and light intensity, probability distribution models of wind and solar output are established [15,16,17,18].

Generally, statistical characteristics of wind speed follow Weibull distribution [15,16], namely,

$$W\left(v\right)=w_1{w}_2{v}^{w_2-1}exp(-w_1{v}^{w_2})$$

(1)

where v is wind speed, and $w_1$ and $w_2$ are parameters of Weibull distribution. Due to its climate characteristics, Qinghai Province, China, is abundant in wind energy resources. With the continuous development of wind power generation bases, Qinghai Province is accumulating a large amount of wind power operating data. To balance the accuracy and computational cost, the sampling interval of wind speed data is set to be 15 min. We divide wind speed data into 10 sets of wind speed levels, and the frequency of each set is calculated. Adopting the maximum likelihood method to estimate parameters of Weibull distribution, Figure 1 shows the probability distribution of wind speed, namely,

$$W\left(v\right)=0.0414v^{1.74}exp(-0.0151v^{2.74})$$

(2)

According to wind speed, wind output is calculated by

$$P_W=\left\{\begin{array}{cc}0,\hfill & v<v_{ci}\hfill \\ (v-v_{ci})/(v_{co}-v_{ci})P_W^{\ast },\hfill & v_{ci}\le v\le v_{co}\hfill \\ 0,\hfill & v>v_{co}\hfill \end{array}\right.$$

(3)

where v is wind speed, $v_{ci}$ is cut-in wind speed, $v_{co}$ is cut-out wind speed, and $P_W^{\ast }$ is rated wind output. Furthermore, two typical distributions of wind speed are compared, including Weibull distribution and Lognormal distribution. To evaluate the fitting accuracy,

Table 1. Two typical distributions of wind speed and light intensity are compared.

Distribution	RMSE of Weibull	RMSE of Lognormal	Best Fit
Wind Speed	0.015	0.028	Weibull
Distribution	RMSE of Beta	RMSE of Lognormal	Best Fit
Light Intensity	0.012	0.021	Beta

lists the root mean square error (RMSE) of the two typical distributions. The RMSE of Weibull distribution is 0.015, while the RMSE of Lognormal distribution is 0.028. Compared to Lognormal distribution, the RMSE of Weibull distribution is lower; namely, the fitting accuracy of Weibull distribution is better.

Generally, statistical characteristics of light intensity follow Beta distribution [17,18], namely,

$$B\left(I\right)=b_{1}exp\left[-{\left({\displaystyle \frac{I-b_2}{b_3}}\right)}^2\right]$$

(4)

where I is light intensity, and $b_1$, $b_2$, $b_3$ are parameters of Beta distribution. The sunshine hours of Qinghai Province are long; therefore, Qinghai Province is advantageous to develop solar power generation. With the continuous construction of solar power generation bases, Qinghai Province is accumulating a large amount of solar power operating data. Similar to wind speed, the sampling interval of light intensity is set to be 15 min. Adopting the maximum likelihood method to estimate parameters of Beta distribution, Figure 2 shows the probability distribution of light intensity, namely,

$$B\left(I\right)=1.626exp\left[-{\left({\displaystyle \frac{I+7.982}{7.23}}\right)}^2\right]$$

(5)

According to light intensity, solar output is calculated by

$$P_V=I[1+(T-25)\gamma ]P_V^{\ast }$$

(6)

where I is light intensity, T is temperature, $\gamma$ is photoelectric conversion coefficient, and $P_V^{\ast }$ is rated solar output. Like wind speed, two typical distributions of light intensity are compared, including Beta distribution and Lognormal distribution. Table 1 lists the root mean square error (RMSE) of the two typical distributions. The RMSE of Beta distribution is 0.012, while the RMSE of Lognormal distribution is 0.021. Compared to Lognormal distribution, the RMSE of Beta distribution is lower; namely, the fitting accuracy of Beta distribution is better.

To improve the economy and stability of data center green power direct supply, analyzing the correlation between wind and solar output is necessary [19,20,21,22]. To analyze the correlation between wind and solar output, the Kendall rank correlation coefficient is calculated, namely,

$$\tau =P[(x_1-x_2)(y_1-y_2)>0]-P[(x_1-x_2)(y_1-y_2)<0]$$

(7)

where $\tau =1$ indicates that changes in x and y are completely consistent, $\tau =-1$ indicates that changes in x and y are completely opposite, $\tau =0$ indicates that x and y are independent of each other, $0<\tau <1$ indicates that there is a positive correlation between x and y, and $-1<\tau <0$ indicates that there is a negative correlation between x and y. Figure 3 shows wind and solar output on a typical day in Qinghai Province. The Kendall rank correlation coefficient between wind and solar output is −0.5539, indicating that there is a negative correlation between wind and solar output; namely, with decreasing wind output, solar output increases, and vice versa. The negative correlation between wind and solar output has profound implications for the optimal mix of wind, light and energy storage capacity, reducing system costs and enhancing reliability through complementary generation profiles. Firstly, when the correlation between wind and solar output is negative, the peaks and troughs of the net load are flattened, and the energy storage requirement and cost are reduced. Secondly, the negatively correlated wind and solar output is conducive to enhancing the system reliability; for example, when the wind output decreases due to an occasional event, the solar output increases to compensate. Lastly, the negatively correlated wind and solar output guides the optimization of the wind–light mix; a more balanced mix of wind and light is favored, instead of a system heavily tilted toward one resource.

The fluctuation of data center load is strong, which is affected by computing tasks significantly. Figure 4 shows the daily variation in data center load levels in different seasons. In different seasons, the daily variation in data center load levels is different. The daily variation in data center load levels in the summer is similar to that in the autumn, while the daily variation in data center load levels in the winter is similar to that in the spring. To establish the probability distribution model of data center load, the data center load data of Qinghai Province is sampled. As with wind speed and light intensity, the sampling interval is set to be 15 min. Adopting the function approximation method to fit statistical characteristics of data center load, Figure 5 shows the probability distribution model of data center load, namely,

$$E\left(L\right)=0.1412L^{1.249}exp(-0.0628L^{2.249})$$

(8)

where L is data center load. The present study utilizes a typical, annualized load profile derived from historical operation data, instead of the worst-case scenario. Firstly, the objective of the present work is to analyze the long-term energy and economic performance of the integrated system under normal operating conditions; a typical profile allows for a representative assessment of annual energy consumption. Secondly, the function approximation method is fitted to power consumption data over a calendar year, where the extreme short-term spikes are smoothed out and the central tendency of the load is captured. The seasonal variations are critically important and taken into account, as shown in Figure 4. We select a typical day from each season, and the overall data is representative of the annual data center load.

To weaken the imbalance between wind–solar output and data center load, energy storage devices are adopted to adjust the imbalance, improving the economy and stability of data centers. To indicate charging and discharging capabilities of energy storage devices, state of charge (SOC) is adopted; the relationship between SOC at time $t+1$ and SOC at time t is

$$\mathrm{SOC}(t+1)=\mathrm{SOC}\left(t\right)+P_c\left(t\right){\eta }_c\Delta t-\lambda Q$$

(9)

$$\mathrm{SOC}(t+1)=\mathrm{SOC}\left(t\right)+P_d\left(t\right){\eta }_d\Delta t-\lambda Q$$

(10)

where $P_c\left(t\right)$ and $P_d\left(t\right)$ are charging and discharging power of energy storage devices at time t, ${\eta }_c$ and ${\eta }_d$ are charging and discharging efficiency of energy storage devices, $\lambda$ is the discharging rate of energy storage devices, and Q is the capacity of energy storage devices. Battery degradation is not a static process, but rather a complex function of multiple stress factors, including cycle aging and calendar aging. In reality, the usable capacity of the battery will shrink annually, reducing its ability to provide capacity services, diminishing future revenue streams. In addition, without considering battery degradation, the replacement cost of the battery is understated. However, the capacity of the battery is usually not used up; the maximum SOC is about 80%, so slight battery degradation does not affect its ability to provide capacity services. Of course, battery degradation will increase the replacement cost of energy storage devices; compared to the purchase cost of energy storage devices, the replacement cost is minor [23,24]. To simplify the present study, the degradation of energy storage devices’ capacity and efficiency is not considered during the evaluation period. We should point out that the above simplification may lead to optimistic results.

3. Capacity Configuration Optimization Model of Wind–Light– Load Storage

To improve the economy and stability of data centers, the capacity configuration optimization model of wind–light–load storage is established. The objective of the capacity configuration optimization model is the yearly average cost of wind–light–load storage, namely,

$$minC(N_W,N_V,N_S)$$

(11)

where $N_W$ is the number of wind output devices, $N_V$ is the number of solar output devices, and $N_S$ is the number of energy storage devices. The yearly average cost of wind–light–load storage is

$$C=C_1+C_2-C_3-r{C}_1$$

(12)

where $C_1$ is the purchase cost, $C_2$ is the maintenance cost, $C_3$ is the electric fee of data center load, and r is the residual value rate. The purchase cost of wind–light–load storage is

$$C_1=(N_W{C}_W+N_V{C}_V+N_S{C}_S)p$$

(13)

where $C_W$ is the price of a set of wind output devices, $C_V$ is the price of a set of solar output devices, $C_S$ is the price of a set of energy storage devices, and p is the present value rate. The maintenance cost is

$$C_2=(Q_W+Q_V+Q_S)t$$

(14)

where $Q_W$ is the unit working time maintenance cost of wind output devices, $Q_V$ is the unit working time maintenance cost of solar output devices, $Q_S$ is the unit working time maintenance cost of energy storage devices, and t is the working time. The electric fee of data center load is

$$C_3=\sum _{i\in t}[P_W\left(i\right)C_E\left(i\right)+P_V\left(i\right)C_E\left(i\right)]\Delta t$$

(15)

where $P_W\left(i\right)$ is wind output at time i, $P_V\left(i\right)$ is solar output at time i, $C_E\left(i\right)$ is the unit electric price at time i, and $\Delta t$ is the time interval. The constraints of the capacity configuration optimization model include wind and solar output constraints, energy storage capacity constraints, and constraints on the balance between wind–solar output and data center load. Firstly, wind and solar output constraints are

$$P_W^{min}\le P_W\le P_W^{max}$$

(16)

$$P_V^{min}\le P_V\le P_V^{max}$$

(17)

where $P_W^{min}$ and $P_W^{max}$ are minimum and maximum values of wind output, respectively, and $P_V^{min}$ and $P_V^{max}$ are minimum and maximum values of solar output, respectively. Secondly, energy storage capacity constraints are

$$P_S^{min}\le P_S\le P_S^{max}$$

(18)

where $P_S^{min}$ and $P_S^{max}$ are minimum and maximum values of energy storage capacity, respectively. Lastly, the balance between wind–solar output and data center load constraints is

$$P_L=P_W+P_V+P_S$$

(19)

where $P_L$ is data center load; $P_W$, $P_V$, $P_S$ are wind output, solar output, and power of energy storage devices.

4. Improved Particle Swarm Optimization (PSO)

To solve the capacity configuration optimization model, particle swarm optimization (PSO) is adopted, as also seen elsewhere [25,26]. However, there are some shortages of the traditional PSO, including falling into the local optimal solution and convergence speed slowing down [27,28]. The convergence speed of PSO is affected by both local search capability and global search capability, which is related to inertia weight factor closely. To balance the local and global search capabilities of PSO, the improved PSO algorithm can adjust inertia weight factor dynamically, namely,

$$v_i^{j+1}=\omega v_i^j+c_1{r}_1(B_1-x_i^j)+c_2{r}_2(B_2-x_i^j)$$

(20)

$$x_i^{j+1}=x_i^j+v_i^{j+1}$$

(21)

where $\omega$ is inertia weight factor, $x_i^j$ is the position, $v_i^j$ is the velocity, i is the particle index, j is the iteration index, $c_1$ and $c_2$ are learning factors, $r_1$ and $r_2$ are random numbers between $[0,1]$, $B_1$ is the local optimal solution, and $B_2$ is the global optimal solution. The search capability of PSO is affected by inertia weight factor directly; namely, when $\omega$ is small, the local search capability is strong while the global search capability is weak, whereas when $\omega$ is large, the local search capability is weak while the global search capability is strong. To adjust inertia weight factor, based on the periodic oscillation of the cosine function, the population diversity is improved, helping particles escape the local optimal solution; the optimized inertia weight factor is

$$\omega =a\left[cos\left(\pi {\displaystyle \frac{i-1}{n}}\right)+1\right]+b$$

(22)

where $a=0.1$ and $b=0.2$ are parameters and n is the number of iterations. The parameters a and b are specific to the problem and are not constant. Generally, the parameters a and b are in the range $[0,1]$ and related to the problem. Adopting the improved PSO to solve the capacity configuration optimization model is carried out as follows: Firstly, set parameters of wind and solar output, energy storage devices, and constraints; adopt the improved PSO to generate a solution space randomly. Then, calculate the fitness of particles in the solution space; namely, simulate the operation of wind–light–load storage based on the number of wind output devices, solar output devices, and energy storage devices, and calculate the yearly average cost. Finally, determine whether the capacity configuration optimization model converges. If it converges, output the optimal solution, namely, the number of wind output devices, solar output devices, and energy storage devices; otherwise, update the solution space and calculate the fitness of particles in the solution space.

5. Case Study

To validate the feasibility of data center green power direct supply with wind–light–load storage, a case study is conducted.

Table 2. Parameters of wind and solar output, energy storage devices, and constraints.

	Wind Output	Solar Output	Energy Storage Devices
Purchase cost/kW	6000	4500	1800
Maintenance cost/kW	30	25	18
Unit electric price/kW·h	0.1	0.2	0.9
Minimum power (kW)	0	0	360
Maximum power (kW)	5000	6000	1800

lists parameters of wind and solar output, energy storage devices, and constraints. To make the present work realistic, the parameters of wind and solar output and energy storage devices are justified by the previous studies [29,30]. Generally, the cost parameters, including purchase cost, maintenance cost and unit electric price, are relatively stable over a long period of time; the fluctuation of the cost parameters is insignificant. The capacity configuration optimization model is solved on a dedicated workstation with the following configuration: CPU, Intel Core i9-13900; memory, 64 GB DDR5 RAM; Operating System, Microsoft Windows 10 Pro; MATLAB Version, MATLAB R2021a. The iteration step of the improved PSO is set to 150; the population size is set to 500. The parameters a and b affect inertia weight factor, which affects the local and global search capabilities of the improved PSO. To show the impact of parameters a and b on the convergence, a sensitivity analysis is conducted, where three different values of the parameters a and b are selected, namely, $a=0.1,b=0.2$; $a=0.2,b=0.8$; and $a=0.5,b=0.9$. If the parameters $a=0.1,b=0.2$ are adopted, the local and global search capabilities are balanced, showing consistent convergence with good solution quality. Comparatively, if the parameters $a=0.2,b=0.8$ and $a=0.5,b=0.9$ are adopted, though the early convergence is enhanced, the final convergence is slower. Therefore, the parameters $a=0.1,b=0.2$ are selected. Generally, PSO is an unconstrained optimizer; to guide the swarm toward the feasible region and to penalize the infeasible solution, the penalty function method is adopted, where a penalty term is added to the objective function, the constrained problem is transformed into an unconstrained problem. The penalty term is calculated based on how much the solution violates the constraints; for the energy storage SOC limit,

$${\mathrm{SOC}}^{min}\le \mathrm{SOC}\le {\mathrm{SOC}}^{max}$$

(23)

the violation of the energy storage SOC constraint might be

$$\mathrm{Violation}\left({\mathrm{SOC}}^{min}\right)=max(0,{\mathrm{SOC}}^{min}-\mathrm{SOC})$$

(24)

$$\mathrm{Violation}\left({\mathrm{SOC}}^{max}\right)=max(0,\mathrm{SOC}-{\mathrm{SOC}}^{max})$$

(25)

The total penalty could be a weight sum of all violations across all time steps,

$$\mathrm{Penalty}=K\times \sum {(\mathrm{Violation}\left({\mathrm{SOC}}^{min}\right)+\mathrm{Violation}\left({\mathrm{SOC}}^{max}\right))}^2$$

(26)

where the penalty coefficient $K={10}^6$ is a large constant to heavily discourage infeasible solutions. Compared to other optimization algorithms, including genetic algorithm (GA) and gray wolf optimizer (GWO), the convergence speed of the improved PSO is improved greatly by adjusting the inertia weight factor dynamically to balance local and global search capabilities, where the computational time of the improved PSO is 1466 s and 2448 s less than GA and GWO, respectively. Figure 6 shows wind–solar output and data center load on a typical day. Due to the randomness of wind speed and light intensity, the fluctuation of wind and solar output is strong. To weaken the imbalance between wind–solar output and data center load, energy storage devices are adopted to adjust the imbalance, improving the economy and stability of data centers. As shown in Figure 7, energy storage devices discharge to compensate for insufficient wind and solar output during peak load periods; energy storage devices charge to reduce abandoned wind and solar output during low-load periods, where charging power is negative and discharging power is positive. Meanwhile, the SOC of energy storage devices is between 20% and 80%, satisfying energy storage device capacity constraints.

Furthermore, the balance rate between wind–solar output and data center load and the rate of abandoned wind and light output are defined as

$$\eta =1-{\displaystyle \frac{{\sum }_{t=1}^T[P_W\left(t\right)+P_V\left(t\right)-P_L\left(t\right)]}{{\sum }_{t=1}^T{P}_L\left(t\right)}}$$

(27)

$$\theta ={\displaystyle \frac{{\sum }_{t=1}^T[P_W\left(t\right)+P_V\left(t\right)-P_L\left(t\right)]}{{\sum }_{t=1}^T[P_W\left(t\right)+P_V\left(t\right)]}}$$

(28)

where $\eta$ is the balance rate between wind–solar output and data center load, $\theta$ is the rate of abandoned wind and solar output, $P_L\left(t\right)$ is data center load at time t, $P_W\left(t\right)$ is wind output at time t, $P_V\left(t\right)$ is solar output at time t, and T is the evaluation period. Obviously, the balance rate between wind–solar output and data center load is closer to 1; the imbalance between wind–solar output and data center load is smaller, so the stability of data center green power direct supply is improved. The rate of abandoned wind and solar output is smaller, and less wind and solar output is abandoned, so the economy of data center green power direct supply is increased. Furthermore,

Table 3. Optimized capacity configuration with different optimization algorithms.

Algorithm	Wind Devices	Light Devices	Energy Storage Devices	$\mathit{\eta }\uparrow$	$\mathit{\theta }\downarrow$
Improved PSO	15	316	619	12.5%	17.5%
DE	16	322	640	10.9%	14.8%
GA	15	323	635	11.3%	15.3%
GWO	15	319	635	11.9%	15.7%

lists the optimized capacity configuration with different optimization algorithms, including the improved particle swarm optimization (PSO), differential evolution (DE), genetic algorithm (GA) and grey wolf optimizer (GWO). Compared to DE, GA and GWO, by adjusting the inertia weight factor dynamically, the improved PSO is more likely to escape the local optimal solution; the yearly average cost of wind–light–load storage optimized with the improved PSO is at the minimum, and the improvement in the balance rate between wind–solar output and data center load is at its maximum. By solving the capacity configuration optimization model of wind–light–load storage with the improved PSO, the balance rate between wind–solar output and data center load is improved by 12.5%, while the rate of abandoned wind and solar output is reduced by 17.5%, which is helpful to improve the economy and stability of data center green power direct supply.

6. Discussion

There are some limitations of the present study. Firstly, the calculated wind and solar output and data center load with probability distribution models are not equal to actual wind and solar output or data center load strictly, which affects the capacity configuration optimization of wind–light–load storage. Secondly, though energy storage devices are adopted to adjust the imbalance between wind–solar output and data center load, the decaying rate of energy storage devices is not taken into account, which affects the balance between wind–solar output and data center load.

Meanwhile, some following studies are forecast. To improve the economy and stability of data center green power direct supply, predicting wind–solar output and data center load is essential. With the development of artificial intelligence, combing the improved PSO and machine learning for time series prediction is potential. In addition, the goal of the capacity configuration optimization model of wind–light–load storage is the economy and stability of data center green power direct supply, namely a balance between wind–solar output and data center load and the rate of abandoned wind and solar output, which is incomplete. More factors should be taken into account, like environmental cost, latency, and thermal load. In the future, the multi-objective optimization of wind–light–load storage will be conducted, to represent the capacity configuration optimization model completely.

7. Conclusions

The capacity configuration optimization of wind–light–load storage is conducted to improve the economy and stability of data center green power direct supply. The objective of the capacity configuration optimization model is to minimize the yearly average cost of wind–light–load storage, and the constraints include wind and solar output, energy storage capacity, and the balance between wind–solar output and data center load. To solve the capacity configuration optimization model effectively, the improved PSO is adopted. Compared to other optimization algorithms, like DE, GA and GWO, by adjusting the inertia weight factor dynamically to balance the local and global search capabilities, the improved PSO is more likely to escape the local optimal solution. To validate the feasibility of data center green power direct supply with wind–light–load storage, a case study was conducted. Compared to other optimization algorithms, like DE, GA and GWO, the balance rate between wind–solar output and data center load is improved by 12.5%, which is the maximum, while the rate of abandoned wind and solar output is reduced by 17.5%, which is the maximum as well.

By conducting the capacity configuration optimization of wind–light–load storage, the balance of the resource distribution is improved while the carbon footprint is reduced. Firstly, the massive computational tasks are processed with green power direct supply, which is economically viable and environmentally sustainable. Secondly, by intelligently partitioning workloads between data centers in the west and low-latency edge processing units in the east, large-scale batch processing is efficiently routed to powerful data centers while latency-sensitive applications receive immediate processing, providing a practical blueprint for operational execution.