Modified Antipredatory Particle Swarm Optimization for Dynamic Economic Dispatch with Wind Power

,


Introduction
As environmental issues become more prominent, wind energy, as an important renewable energy source, has been paid more and more attention.However, the volatility and randomness of wind power make the dynamic economic dispatching (DED) model more and more complicated, which also brings many difficulties to the solution of the model.erefore, the method for solving the dynamic economic dispatching model has always been a research hotspot [1,2].
In the intelligent optimization algorithm, particle swarm optimization is favored by scholars because of its advantages of low parameter setting, fast convergence, robustness, and easy implementation.Some scholars have studied the acceleration coefficients c 1g and c 2g of particle swarm optimization.For example, Snganthan [59] indicated that a better solution can be obtained when the two acceleration coefficients are constant.Trelea [60] further used the convergence factor to improve the convergence ability of the basic PSO and proves the correlation between the two acceleration coefficients.On this basis, Venter and Sobieszczanski-Sobieski [61] proved through experiments that a small cognitive coefficient and a large social coefficient can improve the performance of the PSO algorithm.However, the acceleration coefficients of the abovementioned literatures for particle swarm optimization are all the fixed value, and the influence of the acceleration coefficient of the variation strategy on the performance of the algorithm is not considered.In response to this, Ratnaweera et al. [62] proposed timevarying acceleration coefficients (TVACs), which is to increase the global search ability of the algorithm by using larger individual cognitive coefficients and smaller social coefficients in the early iteration (c 1g is larger, and c 2g is smaller), and as the number of iterations increases, the individual cognitive coefficient decreases linearly and the social coefficient increases linearly so that the algorithm can obtain smaller individual cognitive coefficients and larger social coefficients to speed up the convergence rate of PSO (c 1g is smaller, and c 2g is larger).is change process allows the particles to be distributed in every corner of the search space at an early stage and can quickly converge to the global best at a later stage.However, the linear change strategy makes the value of the acceleration coefficient too uniform in the whole change interval, and the speed of change does not change according to the number of iterations, which cause the particles to fall into prematureness in advance and difficult to jump out in the later stage.
erefore, this linear change strategy still cannot effectively raise the performance of the algorithm.Another part of the literature studies the movement characteristics, displacement, and velocity update formula of particles.Kheshti et al. [63] proposed a double-inertia weighted particle swarm optimization algorithm which uses two different inertia weights at the beginning and the end of the algorithm.Both of inertia weights enhance the local and global search ability of the particles and improve the accuracy of the algorithm.Zou et al. [64] used the new displacement update formula to guide the particle's search activity, expanding the particle's search space and reducing the possibility of particles falling into local optimum.Elsayed et al. [65] added a crossover operation followed by a greedy selection process while replacing the mean best position of the particles with the personal best position of each particle in the velocity updating equation of random drift particle swarm, which increases the diversity of population and improves the performance of the algorithm.Yao et al. [66] introduced quantum computing theory and used quantum bit and angle to depict the state of particles rather than using particle position and velocity of the classical PSO, which shows a stronger search ability and quicker convergence speed.However, these studies all make improvements from the aspects of the optimal solution of particles and do not consider the influence of the worst solution on particle optimization.us, Selvakumar anushkodi [67] added the worst solution idea to the particle swarm optimization and proposed the antipredatory particle swarm optimization.
e algorithm likens the local worst solution and the global worst solution found by the particles in the iteration to the natural predator.e particles will dodge the predator and avoid the worst solution during the optimization process.However, in each iteration, the particles cannot judge whether to make evasive behavior or the degree of evasion according to the distance between the particles and the worst solution at each iteration; that is to say, no matter how far the particles is from the worst solution, it will make the same degree of evasive behavior, which inevitably increases the complexity of the algorithm and the limitation of search space.Synthesizing above all researches, this paper proposes a modified antipredatory particle swarm optimization (MAPSO) based on the evasive adjustment behavior of the particles.Firstly, a nonlinear change strategy is introduced to solve the problem that the linear change strategy of acceleration coefficient makes the particles prone to precocity.en, before the global worst position part of the antipredation particle swarm velocity formula, the avoiding inertia weight is added to adjust the degree of dodge of the particles.
e equation of avoiding inertia weight considers the monotonicity of the sigmoid function and introduces the linear decreasing coefficients to control the avoiding inertia weight within the ideal range of variation.Finally, the effectiveness of the proposed algorithm is verified by test functions and standard power system cases.

Mathematical Model for Economic Dispatch considering Wind Power
2.1.Cost Function.As a renewable and clean energy source, wind power does not generate fuel costs.erefore, on the premise of not considering wind power fuel cost, the goal of dynamic economic dispatching of power system with wind farms is to reasonably distribute the output of each generator set in the power grid to minimize the total cost of power generation during dispatching under the condition of meeting the constraints of load and unit constraints.Usually, the objective of economic dispatch of power systems with wind power is usually to minimize the sum of fuel cost of thermal power units, valve-point effect cost, overestimation cost, and underestimation cost of wind power uncertainty.e total cost expression is as follows: where f(P, w) is the total cost of dispatching in single period, F is the fuel cost, R is the cost of wind power uncertainty, n is the number of thermal power units, N is the total number of thermal power units, m is the number of wind farms, M is the total number of wind farms, P n is the active power output of the nth thermal power unit in a single period, and w m is the actual output of the mth wind farm.e objective function of dynamic economic dispatch needs to consider the sum of costs of each period.e expression is as follows: where h is the dispatching periods, H is the total number of dispatching periods, and C is the total cost of dynamic economic dispatch in H period.
2.1.1.Fuel Cost.e total fuel cost includes fuel cost of thermal power units and energy consumption cost of steam turbine valve-point effect, and the expression is set as ( e power balance constraint equation considering the network loss and wind power is expressed as where P D is the predicted value of active power in single load period and P L is the network loss in single load period.e network loss calculation formula is where B i,j , B i,0 , and B 0,0 are network loss parameters.

Unit Active Output Constraints
P n,min < P n < P n,max , (8) where P n,max and P n,min are, respectively, the upper and lower limits of the output of the nth unit and w m, max is the maximum installed capacity of the mth wind farm.

Unit Ramp Rate Constraints
where U Rn and D Rn are the up and down ramp rates of the nth unit, respectively, and ΔT is the dispatch time interval and takes 1 h.

System Spinning Reserve Constraints
(1) Positive spinning reserve capacity constraint with wind power system: (2) Negative spinning reserve capacity constraint with wind power system: where w u % and w d % are, respectively, the demand coefficients of wind power prediction errors for positive and negative spinning reserve; L u % and L d % are, respectively, the demand coefficients of the system load prediction errors for positive and negative spinning reserve; U n and D n are, respectively, the positive and negative spinning reserve capacity provided by the nth thermal power unit; and T 10 is the response time of spinning reserve, taking 10 minutes.

Antipredation Particle Swarm Optimization (APSO).
Antipredation particle swarm optimization (APSO) algorithm regards the particle optimization process as the foraging process of birds.Birds will try their best to avoid the attack of natural enemies while searching for food in nature.APSO adds the position of the natural enemies as the worst position to the PSO model and divides the cognitive experience in the PSO algorithm into two parts: good cognitive experience and bad cognitive experience.e former is the best location memory of the particle's own experience, and the latter is the memory of the worst position (predator's position) experienced by the particle itself.Similarly, social Mathematical Problems in Engineering experience is divided into two parts: good social experience and bad social experience.e former is the best position memory experienced by the group, and the latter is the worst position memory experienced by the group.us, the speed and displacement update formula of APSO algorithm is obtained as follows: where V i t is the velocity of the ith particle at the tth iteration; ω is the inertia weight, indicating the degree of succession of the particle to the tth iteration speed; c 1g is the acceleration coefficient of the particle flying to its own best position; c 1b is the acceleration coefficient of the particle flying away its own worst position; c 2g is the acceleration coefficient of the best position of the particle flying to the group; c 2b is the acceleration coefficient of the worst position of the particle flying away from the group; r 1 , r 2 , r 3 , and r 4 are random numbers uniformly distributed on [0, 1]; pbest t i is the individual optimal value of the ith particle at the tth iteration; pworst t i is the individual worst value of the ith particle at the tth iteration; gbest t is the global optimal value of the particle swarm at the tth iteration; gworst t is the global worst value of the particle swarm at the tth iteration; and x t i is the displacement of the ith particle at the tth iteration.
In equation ( 13), the velocity update formula is composed of five parts: the first part reflects the influence of the current velocity on the particle, which is related to the current state of the particle; the second part reflects that the particle is affected by its own best position, namely, the particle's good cognitive experience; the third part reflects the influence of its own worst position, i.e., the bad cognitive experience of particles; the fourth part reflects that particles are affected by the best position of the group, that is, the good social experience of the particle; and the fifth part means the influence of particles on the worst position of the group, i.e., the bad social experience of particles.

Nonlinear Time-Varying Acceleration Coefficient.
In order to improve the performance of the algorithm, Ratnaweeera et al. proposed a linear strategy to adjust the acceleration coefficients c 1g and c 2g and experimentally obtained that c 1g is linearly decreasing from 2.5 to 0.5 and c 2g is linearly increasing from 0.5 to 2.5 (the black dotted part in Figure 1), and the algorithm has the best optimization effect.e reason for adopting of linear strategy is that, at the beginning of iteration, particles need to occupy every corner of the search space by relying on their own cognitive experience and maintain the diversity of particles, so a larger c 1g and a smaller c 2g are set.At the later stage of iteration, the global optimal value approaches the convergence value, and each particle needs to approach the optimal solution and search around it.At this time, c 1g is small, while c 2g is large.However, the linear time-varying acceleration coefficient changes too fast in the early iteration so that the particles have not yet undergone too many local searches and have to move closer to the global optimal solution on a large scale.In general, the global optimal value of the population in the early search is often the local optimal value, especially for the multipeak problems, which lead to the premature convergence of the particles to the local optimum; while the speed of change of the acceleration coefficient slows down in the late iteration, which makes the particles learn too much from themselves, and is rarely affected by other particles, further aggravating the premature algorithm.Based on this, this paper proposes a nonlinear time-varying acceleration coefficient method as a new adjustment strategy for c 1g and c 2g (the solid line part of Figure 1).e improved expression is as follows: where the nonlinear parameters η and δ are used to limit the range of variation of c 1g and c 2g .Since the nonlinear strategy has the same trend as the linear strategy, only the rate of change is different, so this paper also uses the same range of variation as the linear variation strategy, and in order to satisfy it, the η and δ are supposed to be 2 and 0.5, respectively; θ is the current number of iterations; and θ max is the total number of iterations.Figure 1 intuitively shows the change process of linear acceleration coefficient and nonlinear acceleration coefficient.
As can be seen from Figure 1, at the initial iteration stage of the algorithm, c 1g and c 2g change slowly so that the 4 Mathematical Problems in Engineering particles can have more time to learn from themselves at the beginning of the iteration and are less affected by other particles.In this way, the particles can avoid blindly moving closer to other particles, which can effectively ameliorate the precocity phenomenon; in the later stage of iteration, c 1g and c 2g change faster, and the social experience of particles accumulates rapidly, which accelerate the convergence speed of particles to the global optimal solution.In summary, the setting of the nonlinear acceleration coefficient can improve the premature phenomenon of the particles in the early stage and speed up the convergence of the particles to the optimal solution in the later stage.

Adjustment of Avoid Inertia
Weight ω gw .c 2b in APSO is used to determine the extent to which particles avoid the global worst solution.It is defined as a constant in this algorithm, which may cause particles to avoid the global worst solution even when they are far away.Figure 2(a) shows the optimization process of two particles i and j in APSO.e superscript t represents the optimization result of the tth iteration, and the superscript t + 1 represents the optimization result of the (t + 1)th iteration; black circle represents the global worst solution.It can be seen from the optimization result of the tth iteration that the particle i is closer to gworst and the particle j is farther away from gworst.In APSO, in the process of avoiding the global worst solution gworst, the particle does not adjust the avoidance degree according to the distance between itself and gworst, so the avoidance extent of i and j is the same.In fact, the particle j is far from gworst and does not need to make the same avoidance extent as particle i, or it may cause the problem that the particle j in Figure 2(a) oversteps the global optimal solution due to its excessive avoidance extent.
In order to solve this problem, the inertia weight ω gw is defined based on the distance between the particle and the global worst solution gworst to adjust the evading degree of the particle.As shown in Figure 2(b), when the particle j is far away from gworst, the particle's evasive inertia weight ω gw is small, and the degree of evasive behavior of the particle is small, avoiding the problem that the particle j is over the optimal solution in Figure 2(a); when the particle i is closer to gworst, the particle's evasive inertia weight ω gw is large, and the degree of evasive behavior of the particle is large, thereby increasing the search range of the particles i.
e expression of ω gw is shown in the following equation: where where α and β are linearly decreasing coefficients; μ is the inclination coefficient; l i is the distance between the particle i and the global worst solution gworst; l min and l max are the minimum and maximum distances of the global worst solution to the particle swarm, respectively; l normal i is the normalized distance of l i ; and I is the number of particles.
When the particles are moving, ω gw and l i should exhibit an inverse relationship, i.e., the distance between the particles and gworst is too close (l i is small), then ω gw is larger, which makes the particles are completely away from gworst; on the contrary, ω gw is smaller, making the particles slightly away from gworst.However, due to the monotonous increment of the sigmoid function itself, the inverse relationship cannot be met.erefore, the linear decreasing strategy is added to the sigmoid function in equation (17) and the values of α and β are defined, which not only makes ω gw show an inverse ratio with the change of the distance but also ω gw is controlled within an ideal range related to α and β.
e fractional part in equation ( 17) is the sigmoid function, and the variation of ω gw depends on l i , which is the global worst value and the distance between the particles.However, if l normal i is directly expressed in l i , then in the actual process, the difference in dimension and order of magnitude caused by different optimization models will cause l i to be too large and the fractional part can only be taken to 1.In order to avoid this phenomenon, this paper normalizes l i and controls it between [− 1, 1], e normalized value is expressed by l normal i , as shown in equation (18).
Since the value of l normal i is between [− 1, 1], the value range of sigmoid function is the red bold curve (AB ⌢ ) of Figure 3.However, this red part is only a small part of the whole value range of sigmoid function (CABD ⌢ ).As can be seen from equation (17), this phenomenon will lead to a very small range of values of ω gw , so it is necessary to multiply a coefficient μ before l normal i to expand the range of the domain of the sigmoid function so that the range of the sigmoid function can be taken as wide as possible.
e red bold curve (AB ⌢ ) in Figure 3 is the range of the sigmoid function before the coefficient μ is defined, and the region (CABD ⌢ ) represented by μ • l normal i in the figure almost covers the entire value range of the sigmoid function, so it is reasonable to define the coefficient μ.
Based on the analysis in this section, the velocity and displacement expressions of MAPSO are given by the following equations: Equation ( 20) changes the trend of c 1g and c 2g into a nonlinear strategy based on the speed update formula of APSO and adds ω gw before the global worst position, which not only improves the premature of the particles but also expands the search scope, allowing the algorithm to more effectively find the optimal solution to the problem.

Constraint Processing and Algorithm Solving Steps
4.1.Constraints Processing.When dealing with DED model of power system, it is necessary to comprehensively consider the output distribution of units and the dynamic adjustment of units in different periods.erefore, the upper and lower power bounds of thermal power units are dynamically adjusted before the dispatching calculation in each period to satisfy the unit ramp and power limitation constraints.
Combined with equations ( 8) and ( 10), the following treatment is carried out:

Algorithm Steps and Flow Chart.
According to the dynamic economic dispatch model of wind farms in this paper, the steps of MAPSO are as follows: Step 1: initialize the particle population and randomly generate the initial velocity and displacement of each particle, input system, unit parameters, and algorithm control parameters. is paper uses linear decreasing inertia weight (LDIW) [68]: where ω max is set to 0.9 and ω min is set to 0.4.
Step 2: limit the initial velocity and displacement of each particle according to the constraints processing of Section 4.1 so that the particle swarm can search in the feasible region.
Step 3: calculate the fitness value of the initial position of the particle and retain the individual optimal and worst positions pbest and pworst, global optimal, and worst positions gbest and gworst.
Step 4: calculate ω gw according to equation (17) and update the displacement and velocity of each particle by equations (20) to (21).
Step 5: calculate the fitness value, update pbest, pworst, gbest, and gworst, and compare with the initial fitness value of step 3, and if the updated fitness value is dominant, it is retained.
Step 6: repeat steps 4 to 5 until the iteration terminates.
Step 7: output the optimal result.
Flow chart is shown in Figure 4.

Case Analysis
Benchmark functions and economic dispatch cases were used in this section to test and verify the MAPSO.e algorithm runs on the personal computer (Windows 7, 64 bits, Intel Core i5 2.6 GHz, 4G RAM) version of MATLAB 2014a.

Benchmark Functions.
In order to verify the effectiveness of the MAPSO algorithm, five test functions are used to prove the search performance of the MAPSO.
(1) Sphere function: 6 Mathematical Problems in Engineering (2) Schwefel function: (3) Rastrigin function: (4) Griewank function: (5) Styblinski function: Among these test functions, the minimum value of f 1 ∼f 4 is 0 and the minimum value of f 5 is − 39.16599n, where n is the dimension of the function (for example, when n � 30, the minimum value is about − 1170).e algorithm parameters are set as follows: the total number of particles is 40, the maximum number of iterations is 1000, and the function dimension is 30.
e time-varying acceleration coefficients particle swarm optimization (TVAC-PSO), the nonlinear time-varying acceleration coefficients particle swarm optimization (NTVAC-PSO), the time-varying acceleration coefficients antipredator particle swarm optimization (TVAC-APSO), the time-varying acceleration coefficients antipredator particle swarm optimization (NTVAC-APSO), and the proposed MAPSO are used to calculate the benchmark functions 50 times independently and compared the optimal value, average value, and standard deviation of the results, as shown in Table 1.
It can be seen from Table 1 that, in terms of the average value and the optimal solution, the results of the MAPSO are superior to the other four algorithms, even multimodal functions such as f 3 and f 4 , which have multiple local minimum values, perform equally well.In terms of standard deviation, the standard deviation of f 1 and f 2 is 0, and in the function f 3 -f 5 , the results of MAPSO are also smaller than the other four algorithms.Taking f 5 as an example, the standard deviation of MAPSO is 34.7.Compared with TVAC-PSO, NTVAC-PSO, TVAC-APSO, and NTVAC-APSO, the standard deviation of MAPSO decreases by 6.8, 3.8, 0.3, and 2.5, respectively.is shows the stability of the MAPSO algorithm and proves the superiority of the MAPSO algorithm in the global optimal solution search ability.
In order to visually demonstrate the convergence characteristics of MAPSO and PSO, Figure 5 shows the convergence curves of two PSO variants and MAPSO.
As can be seen from Figure 5 that the iteration times of the MAPSO algorithm are the least when it converges to the optimal solution, which means that the proposed algorithm can not only improve the convergence speed but also increase the accuracy.In addition, during the iteration process, the fluctuation of the MAPSO convergence curve is relatively more frequent than other PSO variants, which shows that the MAPSO algorithm can expand the search space of particles and improve the activity of the particles so that the particles are not easy to fall into local optimum.

Economic Dispatch.
In order to further prove the superiority of the MAPSO algorithm, the performance of the algorithm is validated by analyzing and simulating the standard 5 units and standard 10 units of IEEE 39-bus in this section.e parameters of the algorithm are set as follows: the number of particles is 30; the total number of iterations is 100; c 1g and c 2g are set according to formula (15)∼( 16); c 1b and c 2b are set to 0.4 and 0.2, respectively; ω is updated according to formula (23); α and β are taken to 0.4 and 0.2, respectively; and μ is taken to 9.

MAPSO Test of 5 Generators Unit System without Wind
Power.Taking the five-unit power system without wind power as an example, the dispatching model considers fuel cost (3) and constraints ( 5), ( 7), (8), and (10)  Mathematical Problems in Engineering period H � 24 h, and time interval is 1 h.e parameters of units, loads, and network losses in the system are obtained from [10].e dynamic economic dispatching model is optimized 30 times by using the MAPSO algorithm, and the optimal total cost is 43439 USD. e output, network loss, and cost distribution of each unit within 24 hours are shown in Table 2.
In order to highlight the performance of the MAPSO algorithm, Table 3 lists the optimal results of the proposed algorithm compared with those of other algorithms.
It can be clearly seen that the proposed algorithm is superior to other algorithms in terms of average, maximum, and minimum, which further illustrates the effectiveness of the algorithm for model solving.

MAPSO Test of 10 Generators Unit System without Wind Power.
e dispatching model is the same as the tenunit model.e parameters of units, loads, and network losses in the system are obtained from [69].e dynamic economic dispatching model is run 30 times by using the MAPSO algorithm, and the optimal total cost is 2474831 USD. e output, network loss, and cost distribution of each unit within 24 hours are shown in Table 4.
e MAPSO algorithm is compared with the algorithms used in the same model and data in recent years [19,37,57,[69][70][71]. e specific results are shown in Table 5.
From Table 5, it can be seen that, compared with the IMOEA/D-CH algorithm, MAMODE algorithm, HMO-DE-PSO algorithm, and RCGA/NSGA-II algorithm, the cost of MAPSO is reduced by 5387 USD, 17638 USD, 9187 USD, and 4187 USD.Similarly, compared with traditional algorithms such as AIS, PSO, and EP, the cost savings are also considerable.In terms of computing time, MAPSO only takes 12.26 seconds, which is much lower than MAMODE and RCGA/NSGA-II.Compared with traditional algorithms, the computing time of MAPSO is also significantly reduced.Compared with IBFA, the computing time of both algorithms is close, but the cost is reduced by 6920 USD.Synthesizing the above data analysis., it can be seen that the proposed MAPSO algorithm is more effective and superior for solving high-dimensional and multiconstrained dynamic economic dispatching models.It can not only save time but also find more economical and reasonable dispatching schemes for decision makers in the process of dispatching without losing accuracy while guaranteeing speed.

MAPSO Test of 10 Generators Unit System with Wind
Power.In this test system, a wind farm connected to the grid is considered.e total installed capacity of the wind farm connected to the grid is 100 MW, with a total of 50 wind turbines.e dispatching model considers fuel cost (3) and constraints ( 6)- (10), dispatching period H � 24 h, and time interval is 1 h.e parameters of units and network losses in the system are obtained from [70].e predicted values of wind power and system load in each period are shown in Table 6 w u % and w d % are 0.2 and 0.3, respectively, while L u % and L d % are 0.05, respectively.e dynamic economic dispatching model is optimized 30 times by using the MAPSO algorithm, and the optimal total cost is 2355671 USD. e output, network loss, and cost distribution of each unit within 24 hours are shown in Table 7.
e power  Mathematical Problems in Engineering distribution of each unit is shown in Figure 6, which verifies that it satisfies the power balance constraints.
Similarly, in order to verify the search performance of the MAPSO algorithm, the optimal results are compared with those of IMOEA/D-CH and NSGA-II-CH algorithms in [70], as shown in Table 8.
It can be seen from Table 8 that the MAPSO algorithm is still superior to the other two algorithms in solving the dynamic economic model with wind power.Compared with the IMOEA/D-CH algorithm and NSGA-II-CH algorithm, the cost is reduced by 4129 USD and 20229 USD, respectively.erefore, the advantages of the MAPSO algorithm are reflected in the results.
In order to embody the rationality of wind power integration into the power system, Table 9 gives the comparison of the optimization results of the dispatching model before and after the application of the MAPSO algorithm to wind power integration.
From Table 9, it is clear that the dispatching model with wind farms is lower than that without wind farms, regardless of network loss or cost.In this paper, Tables 4 and 7 are used to extract three periods from 24 periods for comparative analysis: for example, in the first moment, the network loss of the model without wind farms is 19.54 MW, and that of the model with wind farms is 17.75 MW, which is reduced by 1.79 MW and 9.2%.In the 16th period, the network loss of  [53] 49216.81--SA [10] 47356 --IRCGA [9] 47185 --PS [54] 46530 --DE [9] 45800 --GA [35] 44862.4244921.7645893.95AIS [36] 44385.4344758.8445553.77HS [55] 44367.23 --PSO [35] 44253.2445657.0646402.52ABC [35] 44045 .4%,and the total network loss of the 24 periods was reduced by 95.1 MW.Similarly, in these three periods, the cost savings of wind power grid-connected models are 2726 USD, 7271 USD, and 3633 USD, respectively.e total cost of 24 periods decreased by 119142 USD.It can be seen that the addition of wind power is very beneficial to the efficiency and cost of the system operation.

MAPSO Test of 10 Generators Unit System considering
Wind Power Uncertainty.Considering the randomness of wind power, the model adds overestimation and underestimation of penalty costs on the basis of Section 5.2.3; therefore, the model considers ( 3) and ( 4) and constraints ( 6)- (10), dispatching period H � 24 h, and time interval is 1 h.e parameters of units and network losses in the system are obtained from [70].e total installed capacity of the -2480200 MAMODE [19] 505 2492451 IBFA [57] 5.2 2481733 HMO-DE-PSO [71] -2484000 RCGA/NSGA-II [69] 1080 2516800 AIS [37] 53.56 2519700 PSO [37] 68.47 2572200 EP [37] 72.68 2585400 e dynamic economic dispatching model is run 30 times by using the MAPSO algorithm, and the optimal total cost is 2368764 USD. e thermal power and wind power output, network loss, and cost distribution of each unit within 24 hours are shown in Table 10.
It can be seen from Table 10 that the optimal output, network loss, and cost of each unit in each dispatching period considering wind power uncertainty.e total optimal output of ten thermal generators (G1-G10) and one wind farm (W) under each load is equal to the system load, which satisfies the power balance constraints and the upper and lower power limits of each unit, so the results obtained by MAPSO are reasonable.In order to reflect the rationality of considering the randomness of wind power, Table 11 lists the data comparison results considering and not considering the randomness of wind power dispatching models.
From Table 11, it is obviously that the cost of model considering wind power uncertainty is 13093 USD higher than that of model not considering wind power uncertainty.
is is because the former model adds overestimation and underestimation costs to the original model.Although the cost increases, the impact of wind power stochastic characteristics on the grid cannot be neglected in the actual wind power grid operation.erefore, the slight increase in costs is reasonable.On the contrary, it also shows that the MAPSO algorithm has good search performance in solving the problem of the wind power stochastic model.

Conclusions
Dynamic economic dispatching models are built.Fuel cost of thermal power units, valve-point effect cost, and overestimation and underestimation costs of wind power uncertainty are considered in the objective function of the model.Network loss and positive and negative spinning reserve demand added by uncertainty of wind power grid connection are considered in the constraints.For solving the model, a modified antipredator particle swarm optimization algorithm is proposed, and the following work is done: (1) e algorithm considers the distance between the global worst particle and other particles, introduces the inertia weight ω gw to control the degree of particle movement, and constructs the formula by using the characteristics of sigmoid function, normalization, and linear decreasing.In view of the shortcomings of time-varying acceleration coefficient, a nonlinear time-varying acceleration coefficient is proposed, which improves the local search and global search ability of the particle.(2) In order to verify the effectiveness of the improved algorithm, several benchmark functions and power grid system models are used to analyze the proposed algorithm.e simulation results show that the MAPSO algorithm is practical and superior in model solving.Inertia weight of the particle ω max and ω min : e maximum and minimum inertia weight of the particle c 1g and c 1b : e acceleration coefficients of the particle flying to its own best position and flying away its own worst position, respectively c 2g and c 2b : e acceleration coefficients of the best position of the particle flying to the group and the worst position of the particle flying away from the group, respectively r 1 , r 2 , r 3 , and r 4 : e random number uniformly distributed on [0, 1], respectively pbest t i and pworst t i : e individual optimal value and worst value of the ith particle at the tth iteration, respectively gbest t and gworst t : e global optimal value and worst value of the particle swarm at the tth iteration x t i : e displacement of the ith particle at the tth iteration η and δ: e nonlinear parameters θ and θ max : e current number and total number of iterations, respectively ω gw : e particle's evasive inertia weight A and β: e linear decreasing coefficients of ω gw μ: e inclination coefficient of ω gw l i : e distance between the particle i and the global worst solution gworst l min and l max : e minimum and maximum distances of the global worst solution to the particle swarm, respectively l normal i : e normalized value of l i I: e number of particles.
14 Mathematical Problems in Engineering

Figure 1 :
Figure 1: Change trend chart of c 1g and c 2g .
2n + b n P n + c n + d n sin e n P n,min − P n , b n , and c n are the fuel cost coefficient, d n and e n are the valve-point effect cost coefficients, and P n,min is the lower limit of the active power output of the nth thermal power unit.
L + P D .

Table 1 :
Comparison of five algorithms for benchmark functions.

Table 2 :
Unit output, network loss, and cost during each period of no wind five-unit system.

Table 3 :
Comparison of simulation results.
wind farms is 43.9 MW, and that of the model with wind farms is 38.15 MW, which is reduced by 5.75 MW and 13.1%.In the 24th period, the network loss of the model without wind farms was 25.58 MW, and that of the model with wind farms was 22.15 MW, reduced by 3.43 MW and 13

Table 4 :
Unit output, network loss, and cost during each period of no wind ten-unit system.

Table 5 :
Algorithm optimization results and comparison in 24 h.

Table 6 :
Wind power and system load forecast at different periods.to the grid is 100 MW, with a total of 50 wind turbines.epredictedvalues of wind power and system load in each period are shown in Table6.w u % and w d % are 0.2 and 0.3, respectively, while L u % and L d % are 0.05, respectively.C ow,m and C uw,m are 30 and 5, respectively.

Table 7 :
Unit output, network loss, and cost during each period of wind ten-unit system.

Table 8 :
Comparison of optimization results of each algorithm.

Table 9 :
Data comparison of two models.

Table 11 :
Data comparison of two models.

U
Rn and D Rn : e up and down ramp rates of the nth unit, respectively ΔT: e dispatch time interval L u % and L d %: e demand coefficients of the system load prediction errors for positive and negative spinning reserve w u % and w d %: e demand coefficients of wind power prediction errors for positive and negative spinning reserve, respectively U n and D n : e positive and negative spinning reserve capacity provided by the nth thermal unit, respectively T 10 : e response time of spinning reserve k and c: Shape factor and scale factor of Weibull distribution, respectively w r : e rated capacity of the wind farm V, v i , v r , and v o :