KR102540822B1 - Real-time power optimization method for wind farm and apparatus performing the same - Google Patents

Abstract

A method for real-time power optimization of a wind farm (WF) according to the present invention includes collecting parameters for the plurality of WTs, which are associated with the output power of the wind farm; Generating a wake weight coefficient matrix numerically representing the wake effect acting on each other of the plurality of WTs and a unique wake direction graph representing the wake effect acting on each other with the WTs as nodes and direction lines. ; performing a pruning process on the unique wake-directed graph using a predefined adaptive pruning algorithm to generate a pruned wake-directed graph in which subsets are completely separated; and applying a predefined decentralized optimization solution to independently control clustered subsets according to the network topology according to the pruned wake-directed graph.

Description

Real-time power optimization method of wind farm and apparatus performing the same

The present invention relates to a real-time power optimization technique capable of maximizing the power generation and calculation speed of a wind farm composed of a plurality of wind turbines.

Wind farms composed of a plurality of wind turbines have limitations in securing maximum power due to the influence of a wake effect. More specifically, wind turbines located downstream can produce less power than when operating in a free stream due to wind speed deficits with upstream wind turbines. This wake effect has a great influence on the output power of wind farms.

Various optimization techniques to cope with this wake effect have been proposed. However, in the case of the prior art, there are limitations in terms of power output and calculation speed, and in particular, it is impossible to perform optimal control in real time in response to fluctuating wind conditions.

Korean Patent Publication No. 10-2013-0039156

The present invention was derived to solve the above-described problems, and a real-time power optimization method of a wind farm, which can maximize power output and calculation speed, and enable optimal control in real time in response to fluctuating atmospheric conditions, and It is intended to provide a device that performs this.

According to one aspect of the present invention, a method for real-time power optimization of a wind farm (WF) comprising a plurality of wind turbines (WTs) includes: Collecting parameters for; Generating a wake weight coefficient matrix numerically representing the wake effect acting on each other of the plurality of WTs and a unique wake direction graph representing the wake effect acting on each other with the WTs as nodes and direction lines. ; performing a pruning process on the unique wake-directed graph using a predefined adaptive pruning algorithm to generate a pruned wake-directed graph in which subsets are completely separated; and applying a predefined decentralized optimization solution to independently control clustered subsets according to the network topology according to the pruned wake-directed graph.

In one embodiment, the generating of the pruned wake direction graph may include changing a predefined threshold by a predetermined value and performing a pruning process on the wake direction graph, but without a shared turbine between the subsets. The pruning process may be performed until they are completely separated.

In an embodiment, generating the pruned wake directed graph may include deriving a plurality of threshold values capable of completely separating subsets; calculating power generation information and calculation speed information for each pruned wake direction graph according to the derived threshold values; determining an optimal threshold value among the derived threshold values based on the calculated power generation result information and calculation speed result information; and generating the pruned wake direction graph using the determined optimal threshold.

In an embodiment, in the determining of the optimal threshold value, among the derived threshold values, a threshold value exceeding a predetermined reference power generation value and exhibiting the highest calculation speed may be determined as the optimal threshold value.

In one embodiment, the decentralized optimization solution includes optimization values for each of the yaw angle and axial coefficient of the WT that maximizes the output power of the WF, and is derived by a predefined optimization algorithm, wherein the optimization The algorithm may correspond to an MC-BAS algorithm in which a Monte Carlo law (MC) law is embedded in a beetle antennae search (BAS) algorithm.

According to another aspect of the present invention, an apparatus for real-time power optimization of a wind farm for optimizing a wind farm (WF) including a plurality of wind turbines (WTs) includes wake between the plurality of WTs. a network topology determination unit generating a wake direction graph according to interactions and determining a network topology of the plurality of WTs based on the wake direction graph; and a power adjustment optimization unit applying a predefined non-centralized optimization solution to independently control the clustered subsets according to the determined network topology.

Here, the network topology determination unit collects parameters for the plurality of WTs associated with the output power of the wind farm, and wake weighting that numerically represents a wake effect acting on each other among the plurality of WTs. A coefficient matrix and a unique wake direction graph representing the WT as a node and the wake influence acting on each other as a direction line are generated, and a branch for the unique wake direction graph is generated using a predefined adaptive pruning algorithm. A pruning process is performed to generate a pruned wake directed graph in which subsets are completely separated, and a network topology for the plurality of WTs is determined based on the generated pruned wake directed graph.

The present invention determines a network topology in which subsets are completely separated by generating a pruned wake-directed graph using an adaptive pruning algorithm, and independently controls the clustered subsets according to the determined network topology. By applying a centralized optimization solution, it is possible to maximize power generation and calculation speed, while enabling optimal control in real time in response to fluctuating atmospheric conditions.

1 to 11 are reference views for explaining a real-time power optimization method of a wind farm according to the present invention.
12 is a block diagram illustrating an apparatus for optimizing real-time power of a wind farm according to the present invention.

Since the present invention can apply various transformations and have various embodiments, specific embodiments will be illustrated in the drawings and described in detail in the detailed description. However, it should be understood that this is not intended to limit the present invention to specific embodiments, and includes all transformations, equivalents, and substitutes included in the spirit and scope of the present invention. In describing the present invention, if it is determined that a detailed description of related known technologies may obscure the gist of the present invention, the detailed description will be omitted.

Terms such as first and second may be used to describe various components, but the components should not be limited by the terms. These terms are only used for the purpose of distinguishing one component from another.

Terms used in this application are only used to describe specific embodiments, and are not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly dictates otherwise. In this application, the terms "include" or "have" are intended to designate that there is a feature, number, step, operation, component, part, or combination thereof described in the specification, but one or more other features It should be understood that the presence or addition of numbers, steps, operations, components, parts, or combinations thereof is not precluded. Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1. Introduction

Wind farm (WF) control strategies can be used to achieve a variety of goals, such as increasing production power, improving the lifetime of turbines, and tracking power reference signals to improve energy grid integration. Recently, research on power generation optimization and control for large-scale WF is being conducted. However, there are limitations in real-time implementation of nonlinear optimization algorithms and in responding to fluctuating atmospheric and turbine conditions. As the scale of WF increases, conventional centralized optimization techniques are difficult to utilize. Therefore, efficient non-centralized nonlinear optimization algorithms are needed to perform optimization and minimize online computation time.

In the case of a large-scale WF, the communication burden is very high. To reduce this burden, an efficient decentralized clustering approach is required, through which large WFs can be partitioned into smaller groups or multiple independent subsets of turbines. Here, each group is connected to a local controller that communicates only to a subset of the turbines. The communication structure can be extended to neighboring groups, but in this case, an effective turbine clustering methodology is required. As a prior study on clustering methodologies, various methodologies such as a clustering method considering physical distance and a clustering method based on a downstream velocity coefficient induced by a turbine have been proposed. However, these conventional studies could only solve a specific wind direction problem and could not effectively deal with a complicated wind direction problem. Accordingly, a more efficient clustering methodology considering the wake effect is required.

Another important aspect is to propose an efficient optimization algorithm for nonlinear and non-convex problems. WF control and optimization problems are non-convex problems. Thus, the gradient-based optimization method has difficulty due to local optima convergence, and the solution is affected by the initial guess. In this regard, various algorithms such as evolutionary algorithms (EA), particle swarm optimization (PSO), and genetic algorithm (GA) have been proposed, but these conventional studies have limitations in performing real-time optimization.

The present invention has been made to solve this problem, and proposes a more improved decentralized power optimization scheme. In particular, the present invention proposes an intelligent algorithm named BAS (beetle antennae search) to improve calculation performance and maximize power generation output. To this end, we define a sparse wake digraph to cluster subsets composed of communicating neighbor turbines, and based on the sparse communication network topology to solve the beaconvex power optimization problem. We propose an MC-BAS optimization algorithm with

Hereinafter, the tuning and optimization technology of a wind farm according to the present invention will be described in more detail with reference to drawings and equations.

2. Gaussian-based wake model considering yaw angle

It is important to accurately determine the wake distribution by turbines in order to generate adaptive wake digraphs. Hereinafter, a wind turbine (WT) wake model considering an axial factor and a yaw angle used for optimizing WF power for wake direction control will be described. The WT wake model can be configured including wake deficit, wake deflection, wake combination and turbulence models, including wake overlap area, downstream turbine inlet wind speed, power production and thrust load. A model that calculates may also be included.

Wake's rate deficit can be calculated assuming a Gaussian wake model, which is based on self-similarity theory and is often used for free shear flows. The analytical equation for the 3-dimensional velocity deficit at the rear of the turbine at the far wake can be derived through simplified Navier-Stokes equations.

[Equation (1)]

는 WF에서의 자유 스트림 유입 풍속, x는 유동 방향, y는 가로 방향,

는 웨이크 중심선, z는 수직 방향,

는 허브 높이,

는 y 방향에서의 웨이크 확장(expansion),

는 z방향에서의 웨이크 확장, C는 웨이크 중심에서의 속도 결손이다.where V is the velocity at wake,

is the free stream inlet wind speed in the WF, x is the flow direction, y is the transverse direction,

is the wake centerline, z is the vertical direction,

is the hub height,

is the wake expansion in the y direction,

is the wake extension in the z direction, and C is the velocity deficit at the center of the wake.

는 상류 터빈 i의 요 각도,

는 편향 각도(deflection angle),

는 웨이크 편향(wake deflection)을 나타낸다. 한편, 검은 색 점선은 비 기울임 상태(non-yawed conditions)에서 상류 터빈의 웨이크를 나타내고, 붉은 선은 기울임 상태(yawed conditions)에서 상류 터빈 j의 웨이크를 나타낸다.1 is an example of wake direction control of two turbines. From here,

is the yaw angle of the upstream turbine i,

is the deflection angle,

represents a wake deflection. On the other hand, the black dotted line represents the wake of the upstream turbine in non-yawed conditions, and the red line represents the wake of the upstream turbine j in the yawed condition.

Wake deflection is utilized to redirect wakes from downstream turbines. The relationship between the wake deflection angle and yaw misalignment can be defined as follows.

[Equation (2)]

는 터빈의 요 각도,

는 추력 계수와 발전기 토크,

는 초기 웨이크 편향을 나타낸다.

와 근거리 웨이크(near wake)의 거리

와

간의 관계는 다음과 같이 정의될 수 있다.From here,

is the yaw angle of the turbine,

is the thrust factor and generator torque,

denotes the initial wake deflection.

and the distance of the near wake

and

The relationship between them can be defined as:

[Equation (3)]

와 추력 계수(thrust coefficient)

로 구성된다. 본 발명에서,

와

곡선은 FAST와 NREL(National Renewable Energy Laboratory)의 5MW 터빈에 대한 적합 데이터(fitting data)를 형성하는데 사용되었다.The turbine model provides a power coefficient based on wind speed, tip speed ratio, and blade pitch angle.

and thrust coefficient

consists of In the present invention,

and

The curves were used to form fitting data for FAST and the National Renewable Energy Laboratory (NREL) 5 MW turbine.

[Equation (4)]

는 발전기 효율,

는 공기 밀도,

는 로터 스윕 영역(swept area),

는 요 비정렬의 보정 계수,

는 요각,

는 식 (1)을 통해 계산되는 유효 풍속을 나타낸다. 식 (1) 내지 (4)에 표현된 바와 같이, 생산 전력은 축 계수

와 요각

을 조정함으로써 최적화될 수 있다.From here,

is the generator efficiency,

is the air density,

is the rotor sweep area,

is the correction factor for yaw misalignment,

is the yaw,

represents the effective wind speed calculated through equation (1). As expressed in equations (1) to (4), the produced power is the axial coefficient

with yaw

can be optimized by adjusting

3. Decoupled topology of WF

간 강도의 크기이다. 대규모 WF는 여러 개의 분리된 하위 집합으로 정의될 수 있다. 이하에서는 두 가지 항목으로 분류하여 설명한다. 웨이크 상호작용이 있는 WF에 대한 적응형 웨이크 방향 그래프(adaptive wake digraphs)를 얻는 방법과, 출력 전력 최적화를 위한 축 유도 및 요 제어 조치 기반의 비중앙 집중형 제어 전략(decentralized control strategy)이다. 이러한 제어 전략은 도 2와 같이 요약될 수 있다.A large-scale WF may be separated into at least one subset based on a weight. Methods for determining subsets include the connection of nearest physical neighbors, K-means algorithms, and k-Median clustering algorithms. In the present invention, turbines whose clustering index can be partial or full.

is the magnitude of the liver strength. A large WF can be defined as several disjoint subsets. Hereinafter, two items are classified and described. A method for obtaining adaptive wake digraphs for WF with wake interaction and a decentralized control strategy based on axis guidance and yaw control measures for output power optimization. This control strategy can be summarized as shown in FIG. 2 .

3.1. WF's original wake graph

를 구성하는 방법에 대하여 상세하게 설명한다.Assuming there are N turbines in WF, the wake distribution can be described as an original weighted directed graph. In the following, the eigenwake directed graph

The method of configuring is described in detail.

을 정의한다. 여기에서, 꼭지점

은 터빈을 나타내고, 에지

는 상류 터빈과 하류 터빈간 웨이크 분포를 나타내며,

는 두 노드 사이 상호작용의 강도를 나타내기 위한 가중치(weight)로 사용될 수 있다. (Definition 1) eigenwake direction graph

define Here, the vertex

represents a turbine, and the edge

Represents the wake distribution between the upstream and downstream turbines,

may be used as a weight to indicate the strength of an interaction between two nodes.

[Equation (5)]

는 음수가 아닌 값이다. 상류 터빈

이 하류 터빈

에 웨이크 효과를 발휘하는 경우,

는 아래와 같은 수학식으로 표현될 수 있다.in equation (5)

is a non-negative value. upstream turbine

this downstream turbine

In the case of exerting a wake effect on

can be expressed in the following equation.

[Equation (6)]

는 도 2에 표현되어 있으며, x는 상류 터빈과 하류 터빈 사이의 물리적 거리를 나타내고, D는 터빈 로터의 직경을 나타낸다. 한편, 제어 속도가 변화하는 웨이크 분포에 대응할 수 있을 만큼 충분히 높도록, 웨이크 분포는 제어 기간 동안 일정하게 유지되어야 한다는 점에 유의해야 한다.in equation (6)

is represented in FIG. 2, where x represents the physical distance between the upstream and downstream turbines, and D represents the diameter of the turbine rotor. On the other hand, it should be noted that the wake distribution must remain constant during the control period so that the control speed is high enough to correspond to the changing wake distribution.

의 통신 이웃은

로 표현된다.Vertex (Turbine)

communication neighbors of

is expressed as

와

사이의 공유 통신 이웃은

로 표현되며, 여기에서, Ti는 WF에서의 터빈 번호를 의미한다.

and

Shared communication between neighbors is

where Ti denotes the turbine number in WF.

3.2. Pruned adaptive wake direction graph

를 도출할 수 있다. For most turbines, the degree of coupling between one turbine and another can be high or low. Therefore, the pruned (or sparse) wake-directed graph by pruning the unique wake-directed graph.

can be derived.

를 정의한다. 여기에서, 가지치기된 에지

는

로 정의될 수 있다. (Definition 2) pruned wake directed graph

define Here, the pruned edge

Is

can be defined as

[Equation (7)]

로 정의되며, 여기에서, k는 하이퍼 파라미터이며,

는 기본 임계값을 나타낸다. The adaptive threshold is

is defined as, where k is a hyperparameter,

represents the default threshold.

은 전체 웨이크 가중 계수의 기하학적 중위값(geometric median)으로 정의된다. 여기에서, 기하학적 중위값은,

개의 포인트(

)가 주어졌을 때 유클리드 거리(Euclidean distance)의 합을 최소화하는 값(

)을 의미한다. default threshold

is defined as the geometric median of the total wake weighting coefficients. Here, the geometric median is

points (

) is given, the value that minimizes the sum of Euclidean distances (

) means.

[Equation (8)]

3.3. Separation topology of WF by turbine clustering

A decoupling technique is required for common communication topologies. However, prior research on the methodology for determining the degree of pruning has not yet been actively conducted. In this regard, the present invention proposes an adaptive pruning algorithm for deriving an appropriate threshold value capable of separating into optimal subsets.

각각에 대한 희소 웨이크 방향 그래프

에 기초하여 클러스터링 하위 집합

이 정의될 수 있다. 주어진 풍향

은 이에 대응하는 클러스터 하위 집합

을 갖는다. 여기에서, M은 하위 집합의 수이다. WF의 레이아웃에 따라 고유 웨이크 방향 그래프

를 구축하고, 터빈의 통신 이웃을 결정하기 위한 웨이크 가중 매트릭스

를 계산한다. 여기에서, 도 3을 참조하여 가지치기된 적응형 웨이크 방향 그래프를 이용한 터빈 클러스터링 알고리즘에 대해 설명한다. wind direction

Sparse wake directed graph for each

Clustering subsets based on

this can be defined. given wind direction

is the corresponding subset of clusters

have Here, M is the number of subsets. A unique wake direction graph according to the layout of the WF

and the wake weighting matrix to determine the communication neighbors of the turbine

Calculate Here, a turbine clustering algorithm using a pruned adaptive wake direction graph will be described with reference to FIG. 3 .

, 풍속

, 터빈 간 거리 x, 반경 D,

등 터빈과 관련된 파라미터를 수집 또는 결정한다.Step 1: Based on WF's layout (X,Y), wind direction

, wind speed

, the distance between the turbines x, the radius D,

Collect or determine parameters related to the turbine.

로 재학습(re-train)한다. 희소화 작업은 일부 에지를 제거할 수 있도록 특정 구조적 패턴이 있는 일부 파라미터에 대해 수행된다. Step 2: sparseness degree relative error

Re-train with The sparsization operation is performed on some parameters with certain structural patterns so that some edges can be removed.

를 도출한다.Step 3: Pruning the pruned wake-directed graph by performing a pruning process on the unique wake-directed graph.

derive

를 경험하는 터빈으로 결정된다. 또한, 웨이크 방향 그래프

의 연결 정보를 통해 리드 터빈의 통신 이웃을 결정하며, 여기에서, BFS(depth-first tree search) 알고리즘을 통해 동일한 하위 집합

로 클러스터링되는 꼭지점을 찾을 수 있다. Step 4: Cluster the wake directed graph. Here, the lead turbines of each subset are free-stream velocities.

is determined by the turbine experiencing Also, the wake directed graph

The communication neighbors of the lead turbine are determined through the connection information of the same sub-set through the depth-first tree search (BFS) algorithm.

You can find vertices clustered with .

사이에 공유 셋

가 존재하면, k의 값을 증가(

)시켜 스텝 2로 회귀하고, 존재하지 않으면 스텝 6을 진행한다. Step 5: If the clustered subset

shared between three

If exists, increase the value of k (

) to return to step 2, and if it does not exist, proceed to step 6.

Step 6: Calculate the output power and calculation time based on the k value of step 2, and save each parameter.

의 계수들이 모두 0이 아니라면,

하여 스텝 2로 회귀하고, 다른 k 값에 따른 모든 파라미터들과 비교하여 최적 값을 도출한다. 만약

의 계수들이 모두 0이라면, 스텝 8을 진행한다.Step 7: If

If the coefficients of are all non-zero,

and returns to step 2, and compares all parameters according to other values of k to derive an optimal value. if

If the coefficients of are all 0, proceed to step 8.

에 기초하여, 터빈 클러스터링 하위 집합

을 결정하고, 다른 k값에 따른 모든 파라미터들을 분석하여 최적 k 값을 선택한다.Step 8: Pruned Wake Direction Graph

Based on, the turbine clustering subset

is determined, and an optimal k value is selected by analyzing all parameters according to different k values.

Referring to FIG. 3, there is no shared turbine in the subset, and there is a parameter k for optimal control (on the other hand, the optimal value of k will be described later). An adaptive value k may be determined offline, and a lookup table may be configured to update k whenever the wind speed and direction change.

4. WF control strategy

4. 1. MC-BAS (Monte Carlo law with the BAS) controller

The total power output can be maximized by optimizing the yaw setting and axial induction coefficient setting of the WF using the BAS algorithm approach according to the present invention. The WF cost function is presented in equation (4) as a nonlinear and non-convex optimization problem. Here, the present invention proposes a BAS algorithm to solve this optimization problem. The BAS algorithm according to the present invention does not require prior knowledge of a specific shape of a function or gradient information to optimize efficiency, and can significantly reduce calculation time and speed up optimization compared to other algorithms.

The BAS algorithm has the advantages of fast optimization and convergence, but it can fall into the optimal local value during the iterative process and cannot determine the optimal global value. This is because BAS is overly dependent on the step size. In order to solve and improve these BAS problems, the present invention proposes the Monte Carlo (MC) law of the annealing algorithm (SA), which can greatly improve the repeatability and stability of the algorithm. The improved algorithm according to the present invention can be applied to wake steering control to maximize the power generation of the WF. Simulate Anneal is a Monte Carlo based optimization problem solution. The optimal target value may be determined by simulating the annealing process of the target and the lowest energy of the target value. Simulated annealing incorporates random variables into the retrieval process. That is, the probability of ending the local optimization increases as it contains a solution that is worse than the current solution with a certain probability.

The modified BAS algorithm according to the present invention is as follows.

1) Random direction vector

To simulate the search behavior of the beetle, we define the direction vector as:

[Equation (9)]

Here, rand(k,1) represents a random function, and k represents the dimension of a position.

2) The right and left coordinates of the beetle antennae are defined as follows.

[Equation (10)]

과

은 각각 t 반복에서 딱정벌레의 오른쪽과 왼쪽 더듬이에 대한 공간 위치를 나타내고, d는 왼쪽과 오른쪽 더듬이 사이의 거리를 나타낸다.where t represents the number of iterations,

class

denotes the spatial location for the beetle's right and left antennae at iteration t, respectively, and d denotes the distance between the left and right antennae.

3) fitness value

[Equation (11)]

와

는 현재 공간 위치에서 오른쪽 수염과 왼쪽 수염의 적합도 값을 나타내고, f(-)는 적합성 함수를 나타낸다.From here

and

denotes the fitness values of the right and left whiskers at the current spatial location, and f(-) denotes the fitness function.

4) Pre-update position

[Equation (12)]

는 t 반복에서 알고리즘의 스텝 크기 계수(step size factor)이다.Pre-update the beetle's position based on iterations. Here, sign(-) is a symbol function,

is the step size factor of the algorithm at t iterations.

5) Accepted solution using the Monte Carlo law

The Monte Carlo laws of the SA algorithm can be embedded in BAS. In the iterative process, the probability P is used to accept sub-solutions to improve the global optimization ability of BAS.

[Equation (13)]

는 사전 업데이트 위치를 나타내고,

는 마지막 반복에서 최적 위치를 나타내며, exp(-)는 지수 함수를 나타내며, T는 최고 온도를 나타낸다. From here,

represents the pre-update position,

denotes the optimal position at the last iteration, exp(-) denotes the exponential function, and T denotes the maximum temperature.

6) Step size

[Equation (14)]

는 더듬이의 길이와 스텝 사이즈를 나타내며,

와

는 초기값을 나타내고,

와

는 t 반복에서 알고리즘의 스텝 사이즈 계수이며, 두 파라미터

와

는 사용자에 의해 설정된다. here d and

represents the length of the antennae and the step size,

and

represents the initial value,

and

is the step size coefficient of the algorithm at t iterations, and the two parameters

and

is set by the user.

The Monte Carlo law can be mixed with the BAS algorithm, and the basic process of the grouped MC-BAS algorithm can be summarized as Algorithm 1 below.

[Algorithm 1]

4. 2. Decentralized MC-BAS controller

에 대한 비중앙 집중형 전력 제어는 다음과 같은 단일 목적 최적화 문제(single-objective optimization problem)를 기반으로 한다. The present invention proposes a non-centralized control method for optimal power control of a large-scale WF. A WF is split into several independent WT clusters, and their communication neighbors can be determined by the network topology of the sparse wake digraph. Therefore, we propose a non-centralized control strategy for WF to realize power control of the host computer. Each WT cluster consists of several neighboring WTs distributed in different communication network topologies in an adaptive pruning wake-directed graph. The non-centralized control scheme according to the present invention is shown in FIG. 2 . given clustering subset

Decentralized power control for is based on the following single-objective optimization problem.

을 최대화하는 요(yaw)에 대한 최적화 값

과 축 계수(axial factor)에 대한 최적화 값

을 찾는다. 여기에서, 전체 WF 비중앙 집중형 최적화 함수

는 다음과 같이 표현될 수 있다. A single-purpose optimizer for nonlinear programming problems (NLP) uses the BAS algorithm to maximize the power production of the WF and simultaneously

Optimization value for yaw that maximizes

Optimization values for axial factor and

look for Here, the full WF decentralized optimization function

can be expressed as

[Equation (15)]

는

내지 와

의 범위를 가지며, 축 계수

는

내지

의 범위를 갖는다. 또한, 전력

는

내지 정격 전력

의 범위를 갖는다. 예를 들어, 도 6(d)를 참조하면, WF은 각각 일정 개수의 터빈으로 구성된 13개의 하위 집합(M=13)으로 분리된다. 대규모 WF에 대한 실시간 제어를 보장하기 위해서는 통신 및 데이터 부담을 최소화할 필요가 있으며, 이 때 각 클러스터 하위 집합은 전체 숫자의 일부에 불과하므로 독립적으로 신속하게 제어될 수 있다. 한편, 중앙 집중형 알고리즘(centralized algorithm)은 경우, 하위 집합은 하나(M=1)이며, 이웃 터빈은 전체 터빈 K={25}으로 볼 수 있다. 즉, M과 K의 값을 다르게 설정하면 상이한 풍속, 풍향 및 제어 방식에 대한 최대 출력 전력을 설명하는 데 사용할 수 있다.where f(x) is the total power of the entire WF, i is a subset, j is a turbine, M is the number of discrete subsets of a WF, and K is the number of turbines in a WF. yaw angle

Is

come out

has a range of, and the axial coefficient

Is

pay

has a range of Also, power

Is

rated power

has a range of For example, referring to FIG. 6(d), WF is divided into 13 subsets (M=13) each composed of a certain number of turbines. To ensure real-time control over large WFs, it is necessary to minimize the communication and data burden, where each cluster subset is only a fraction of the total number and can be independently and quickly controlled. On the other hand, in the case of a centralized algorithm, the subset is one (M = 1), and the neighboring turbines can be viewed as all turbines K = {25}. That is, different values of M and K can be used to account for the maximum output power for different wind speeds, wind directions and control schemes.

5. Simulation and Review

로 설정되었으며, 하나의 제어 사이클 내에서 일정한 값을 갖는 것으로 가정하였다. Wind speed and direction are not constant due to wind irregularities and intermittence. Accordingly, this can be simplified by calculating the average value for 10 minutes based on the probability of wind rose to obtain an approximately constant value. In this simulation, the average wind speed is 8 m/s, and the wind direction range is

was set to , and it was assumed to have a constant value within one control cycle.

The simulation results using the same 25 turbines (NREL-5MW Type III WT) are used to verify the performance of the centralized and decentralized optimization methods. The layout structure of the WF is as shown in FIG. 3, where D = 126 m, nominal power P = 5 MW, and the vertical distance between each turbine is 200 m and the horizontal distance is 500 m.

초기값은 0이고 범위는 [-30, 30]으로, 축 계수

초기 값은 1/3이고 범위는 [0, 1/3]으로 설정되었다. The test is conducted for an average of 10 minutes in free wind speed and direction. In addition, to extend the algorithm, three WFs with N = 4, N = 9, and N = 25 WT were set, respectively.

Initial value is 0, range is [-30, 30], axis coefficient

The initial value is 1/3 and the range is set to [0, 1/3].

Meanwhile, as shown in FIG. 4 (a unique wake direction graph for wind directions of 0, 25, 45, and 90 degrees), it can be confirmed that the wake effect is highly influenced by the wind direction and wind speed according to the layout.

5. 1. Processing of adaptive pruning wake directed graphs

와 가지치기된 웨이크 방향 그래프

를 나타낸다. 분리된 통신 토폴리지는 적합한 가지치기 웨이크 방향 그래프에 의해 도출될 수 있다.A clustering method (see 3. above) for dividing a large WF into a plurality of sub-sets using the branched wake-directed graph clustering method according to the present invention will be described. 5 and 6 are graphs of unique wake directions for wind directions of 90 and 45 degrees, respectively.

and a pruned wake-directed graph

indicates The isolated communication topology can be derived by suitable pruning wake direction graphs.

Figure 5 shows that the eigenwakedirection graph is equivalent to the pruned wakedirection graph. This is because when the wind direction is 90 degrees, all wake effects are concentrated on the downstream turbine without diffusion. In this case, the power yield of the WF is low due to the low efficiency due to the high wake interaction.

는 k마다 상이할 수 있다. 도 6(c)의 경우 임계값은

로, 도 6(b)의 고유 웨이크 방향 그래프에서 해당 임계값보다 낮은 엣지(

)는 제거되는 것을 확인할 수 있다. 도 6(d)의 경우 임계값은

로, 도 6(c)의 가지치기 고유 웨이크 방향 그래프에서 해당 임계값보다 낮은 엣지가 제거되어, 공유 터빈이 없는 13개의 하위 집합으로 분리되는 것을 확인할 수 있다. 도 5 및 6을 참조하면, 외부 관련 변수(external relevant variable)

가 웨이크 효과에 영향을 미치며 이에 따라 웨이크 토폴로지가 파라미터에 의존한다는 것을 확인할 수 있다. 주어진 풍향에 대해, 최적으로 분리된 하위 집합을 도출하기 위한 최적의 k를 결정하는 것이 중요하다. 하이퍼 파라미터 k의 튜닝 방법은 도 7을 참조하여 보다 상세하게 설명한다. On the other hand, when the wind direction is 45 degrees, as shown in FIG. 6 , the unique wake direction graph is different from the adaptive pruning wake direction graph. Also, a pruned wake directed graph

may be different for each k. In the case of FIG. 6(c), the threshold is

, an edge lower than the corresponding threshold value in the unique wake direction graph of FIG. 6 (b) (

) can be confirmed to be removed. In the case of FIG. 6(d), the threshold is

, it can be confirmed that edges lower than the corresponding threshold are removed from the pruning eigenwake direction graph of FIG. 6(c) and separated into 13 subsets without shared turbines. 5 and 6, an external relevant variable

affects the wake effect, and it can be confirmed that the wake topology depends on the parameters. For a given wind direction, it is important to determine the optimal k to derive optimally separated subsets. The tuning method of the hyper parameter k will be described in detail with reference to FIG. 7 .

)이다. K의 범위에 따른 하위 집합의 수 및 공유 터빈의 유무를 정리하면 아래 표 1과 같다. 7 is a reference diagram for explaining a pruning process of an adaptive wake direction graph (wind direction of 45 degrees,

)am. Table 1 below summarizes the number of subsets according to the range of K and the presence or absence of shared turbines.

[Table 1]

Referring to Table 1, the number of subsets according to the change in k value and the existence of shared turbines can be confirmed. Here, it is necessary to pay attention to the range of k corresponding to 2.5 to 7.3 without a shared turbine (subsets are all separated) for establishing a non-centralized control strategy. Next, a methodology for determining the optimal k value within the range should be presented. An adaptive threshold may be proposed by comparing output power and calculation time, and the comparison result may be shown in Table 2 below. Table 2 compares the results of pruning wake direction graphs according to the change in k value.

[Table 2]

Referring to Table 2, when k>5.7, the output power is smaller than a preset reference value, so an optimum value cannot be selected in the corresponding range. Next, when k = 5.7, the calculation time is the smallest, and k = 5.7 may be selected as an optimal value. The speed of the controller is an important factor in terms of real-time control. Under these conditions, the pruned wake directed graph is separated into 13 subsets, more specifically N1={T1,T8,T15}, N2={T2,T9}, N3={T3,T10}, N4 ={T4}, N5={T5}, N6={T6,T13,T20}, N7={T7,T14}, N8={T11,T18,T25}, N9={T12,T19}, N10={ T16,T23}, N11={T17,T24}, N12={T21}, N13={T22}.

Using the method described above, the wind direction range can be expanded by 15 degrees from 0 degrees to 90 degrees. Under different wind direction conditions, in order to separate the communication topology through pruning of the wake direction graph, the test was conducted by changing k in the range of k1 to k3, and k2 was derived through the adaptive pruning algorithm according to the present invention. is the optimal value.

[Table 3]

On the other hand, for real-time control purposes, it is desirable to determine the optimum value k2 with more focus on computational efficiency. Referring to Table 3, it can be confirmed that the optimum value k2 represents an output power higher than the predetermined reference value, and is determined as a k value representing the shortest control time among them.

5. 2. Comparison of centralized and non-centralized MC-BAS algorithms

The efficiency of the centralized MC-BAS algorithm can be confirmed in 4. above. However, centralized algorithms are not suitable for large-scale WFs with high communication and computational demands. Therefore, we propose a non-centralized MC-BAS communication framework that has several advantages over conventional PSO and GA algorithms. In the following, centralized and non-centralized algorithms are compared and described in terms of computation time and accuracy.

Based on the adaptive pruning wake-directed graph, the communication architecture can be established once the cluster subsets are completely separated. A decentralized MC-BAS algorithm is proposed based on a separate solution communication architecture. To verify the effectiveness of the new algorithm, the non-centralized MC-BAS algorithm is compared with the centralized MC-BAS algorithm and the greedy algorithm. The results are shown in FIG. 8 and Table 4.

The result of each output power is shown in FIG. 8, and the comparison of the total output power is shown in Table 4. When the centralized and non-centralized methods are implemented using the MC-BAS control method, the total generated power is increased compared to the greedy control method. For MC-BAS centralized and non-centralized control schemes, the wake effect is taken into account to optimize the overall power output. Here, most of the upstream turbines can reduce the output power while the downstream turbines increase the output power to increase the total power output. The calculated time differs between centralized and non-centralized methods. The control speed of the decentralized method is higher than that of the centralized method because there are fewer turbines to solve, as shown in Table 4.

[Table 4]

)은 그리디 중앙 집중형 알고리즘의 P_total 기준선(baseline)에서 증가한 전력의 비율이다. 표 4는 비중앙 집중형 및 중앙 집중형 MC-BAS 방법에 따르면 다양한 풍향 조건에서

이 증가됨을 보여주며, 또한 비중앙 집중형 MC-BAS 방법의 계산 시간(T_total)이 중앙 집중형 방법의 1/3 미만으로 감소됨을 보여준다.power increase rate (

) is the ratio of power increased from the P_total baseline of the greedy centralized algorithm. Table 4 shows the non-centralized and centralized MC-BAS methods under various wind direction conditions.

increases, and also shows that the computation time (T_total) of the non-centralized MC-BAS method is reduced to less than 1/3 of that of the centralized method.

The average total power by the centralized and non-centralized MC-BAS algorithms improves by 14.4% and 11.3676%, respectively, compared to the baseline. This indicates that the non-centralized MC-BAS has a power loss of about 3.0324% compared to the centralized MC-BAS method. Thus, the control strategy according to the present invention is useful for improvement of power output and improvement of calculation speed in view of the benefits of real-time control and large-scale WF.

5. 3. Comparison of MC-BAS algorithm with other algorithms

In general, the more iterations, the more accurate the calculation. Hereinafter, a description will be made from the viewpoint of minimizing the number of iterations to be performed to reach the optimal control action.

Referring to FIG. 9 , it can be seen that the calculation time significantly increases as the number of iterations increases by 20 times from 100 to 300 times. The number of iterations plays an important role in reducing the computation time, thereby reducing the communication burden. Accordingly, the exact iteration value is an important tuning value of the optimization algorithm. The calculation time of MC-BAS is shorter than that of PSO (Particle swarm optimization) and GA (genetic algorithm). MC-BAS calculates at about 1/9 of PSO and 1/4 of GA. Thus, the MC-BAS method according to the present invention is superior to other intelligent methods (GA and PSO) in terms of speed.

Figure 10 shows the amount of power generation according to the wind direction and repetition by four different control algorithms. The PSO control method can produce more power than other methods in most iterations. However, when the number of iterations exceeds 140, the total yield of the MC-BAS algorithm becomes the same as that of the PSO algorithm. On the other hand, the main disadvantage of the GA algorithm is the instability of the output power that changes with repetition, as shown in FIGS. 10(b) and (c). Therefore, the GA algorithm is not suitable for WF.

11 shows the amount of power generation and calculation time according to the change of wind direction by different control algorithms. 11(a) and (c), it can be seen that the total power generation of MC-BAS is similar to that of PSO and is superior to the greedy algorithm in most wind directions. Also, referring to FIGS. 10(b) and (d), it can be confirmed that the MC-BAS algorithm shortens the calculation time compared to the PSO algorithm.

As described above, the MC-BAS algorithm according to the present invention has an advantage of increasing output power and reducing calculation time compared to other algorithms. In addition, the algorithm of the adaptive pruning wake-directed graph according to the present invention can separate a large-scale WF into a plurality of sub-sets, and each sub-set can be controlled in real time through the same controller.

6. Real-time power optimization device according to the present invention

Hereinafter, with reference to FIG. 12, a real-time power optimization device (hereinafter, a power optimization device) that performs the real-time power optimization method for a WF according to the present invention described above will be described.

A power optimization device according to the present invention is a computing device that performs the above-described power optimization method, and may include a network topology determining unit 100 and a power adjustment optimization unit 200 . Here, components of the power optimization device may be implemented by being included in a single device (eg, central control unit), or implemented by being distributed to a plurality of devices (eg, central control unit and control unit included in each WT, etc.) It can be. On the other hand, this example is not intended to limit the scope of the present invention, and if the device in which the power optimization method described above is implemented, it is not limited to the type, name, number of implemented devices, etc., and is interpreted as a power optimization device according to the present invention. It should be.

Hereinafter, operations of the network topology determining unit 100 and the power adjustment optimization unit 200 will be described, and descriptions overlapping with those described above will be omitted.

The network topology determination unit 100 generates a wake direction graph according to a wake interaction between a plurality of WTs, and determines a network topology of the plurality of WTs based on the wake direction graph.

The power adjustment optimizer 200 applies a predefined non-centralized optimization solution to independently control the clustered subsets according to the network topology determined by the network topology determiner 100 . Here, the non-centralized optimization solution includes optimization values for each of the yaw angle and axial coefficient of the WT maximizing the output power of the WF, and may be derived by a predefined optimization algorithm. In one embodiment, the optimization algorithm may correspond to an MC-BAS algorithm in which a Monte Carlo law (MC) law is embedded in a beetle antennae search (BAS) algorithm.

That is, in performing the power optimization method described in 1. to 5. above, the network topology determining unit 100 performs the process of determining the network topology for a plurality of WTs, and the power adjustment optimization unit 200 performs a non-centralized process. It is a configuration that creates a centralized optimization strategy (solution) and performs the process of applying a decentralized optimization solution to each of the subsets.

In one embodiment, the network topology determiner 100 may collect parameters for a plurality of WTs that are associated with the output power of the wind farm. Thereafter, the network topology determining unit 100 includes a wake weight coefficient matrix that numerically represents the wake effect acting on each of the plurality of WTs, and a unique matrix representing the wake effect acting between the WTs as a direction line with the WTs as nodes. You can create a wake-directed graph.

Thereafter, the network topology determination unit 100 performs a pruning process on the unique wake direction graph using a predefined adaptive pruning algorithm to generate a pruned wake direction graph in which subsets are completely separated. can do.

In an embodiment, the network topology determiner 100 changes a predefined threshold by a predetermined value (changes the threshold by changing the value of hyper parameter k) and performs a pruning process on the wake direction graph. can Here, the network topology determining unit 100 may perform a pruning process until the subsets are completely separated without sharing turbines between the subsets. (See 5.1 above)

In one embodiment, the network topology determining unit 100 derives a plurality of threshold values capable of completely separating subsets, and for each wake direction graph pruned according to the derived threshold values, power generation information and Calculation speed information can be calculated. Here, the network topology determining unit 100 determines an optimal threshold value among the derived threshold values based on the calculated power generation result information and calculation speed result information, and prunes using the determined optimal threshold value. You can create a wake-directed graph. (See 5.1 above)

The adjustment optimization method according to the present invention described above can be implemented as computer readable codes on a computer readable recording medium. A computer-readable recording medium includes all types of recording devices in which data that can be read by a computer system is stored. Examples of computer-readable recording media include ROM, RAM, CD-ROM, magnetic tape, floppy disk, and optical data storage devices. In addition, the computer-readable recording medium is distributed in computer systems connected through a network, so that computer-readable codes can be stored and executed in a distributed manner.

The preferred embodiments of the present invention described above have been disclosed for illustrative purposes, and those skilled in the art with ordinary knowledge of the present invention will be able to make various modifications, changes, and additions within the spirit and scope of the present invention, and these modifications, changes, and The additions should be viewed as falling within the scope of the following claims.

100: network topology determination unit
200: power adjustment optimization unit

Claims (6)

In the real-time power optimization method of a wind farm (WF) composed of a plurality of wind turbines (WT),
collecting parameters for the plurality of WTs associated with the output power of the wind farm;
Generating a wake weight coefficient matrix numerically representing the wake effect acting on each other of the plurality of WTs and a unique wake direction graph representing the wake effect acting on each other with the WTs as nodes and direction lines. ;
performing a pruning process on the unique wake directed graph to generate a pruned wake directed graph in which subsets are completely separated; and
applying a decentralized optimization solution to independently control clustered subsets according to the network topology according to the pruned wake-directed graph;
Including,
The pruning process is
A plurality of threshold values capable of completely separating subsets are derived, power yield information and calculation speed information are calculated for each of the pruned wake direction graphs according to the derived threshold values, and the calculated power yield results Based on the information and calculation speed result information, after determining an optimal threshold among the derived thresholds, generating the pruned wake direction graph using the determined optimal threshold;
The above decentralized optimization solution is
Includes optimized values for each of the yaw angle and axial coefficient of the WT that maximizes the output power of the WF, and the Monte Carlo law (MC) law is embedded in the BAS (beetle antennae search) algorithm MC-BAS algorithm Characterized in that it is derived by an optimization algorithm corresponding to,
A real-time power optimization method for wind farms.
The method of claim 1, wherein generating the pruned wake-directed graph comprises:
Characterized in that the pruning process is performed until the subsets are completely separated without a shared turbine between the subsets.
A real-time power optimization method for wind farms.
delete The method of claim 1, wherein determining the optimal threshold value comprises:
Characterized in that, among the derived threshold values, a threshold value exceeding a preset reference power output value and indicating the highest calculation speed is determined as an optimal threshold value,
A real-time power optimization method for wind farms.
The method of claim 1, wherein the non-centralized optimization solution,
Includes optimized values for each of the yaw angle and axial coefficient of the WT that maximizes the output power of the WF, and the Monte Carlo law (MC) law is embedded in the BAS (beetle antennae search) algorithm MC-BAS algorithm Characterized in that it is derived by an optimization algorithm corresponding to,
A real-time power optimization method for wind farms.
An apparatus for real-time power optimization of a wind farm for optimizing a wind farm (WF) composed of a plurality of wind turbines (WT), comprising:
a network topology determination unit generating a wake direction graph according to a wake interaction between the plurality of WTs and determining a network topology of the plurality of WTs based on the wake direction graph; and
A power adjustment optimization unit for applying a decentralized optimization solution to independently control clustered subsets according to the determined network topology;
The network topology determining unit,
A wake weighting coefficient matrix that collects parameters for the plurality of WTs associated with the output power of the wind farm and numerically represents a wake effect acting on each other among the plurality of WTs, and the WT as a node , generate a unique wake direction graph representing the wake influence acting on each other as a direction line, and perform a pruning process on the unique wake direction graph, so that the subsets are completely separated. generating a directed graph and determining a network topology for the plurality of WTs based on the generated pruned wake directed graph;
The pruning process is
A plurality of threshold values capable of completely separating subsets are derived, power yield information and calculation speed information are calculated for each of the pruned wake direction graphs according to the derived threshold values, and the calculated power yield results Based on the information and calculation speed result information, after determining an optimal threshold among the derived thresholds, generating the pruned wake direction graph using the determined optimal threshold;
The above decentralized optimization solution is
Includes optimized values for each of the yaw angle and axial coefficient of the WT that maximizes the output power of the WF, and the Monte Carlo law (MC) law is embedded in the BAS (beetle antennae search) algorithm MC-BAS algorithm Characterized in that it is derived by an optimization algorithm corresponding to,
A real-time power optimizer for wind farms.

Images