Offshore wind plant active power distribution method considering steady-state fan response error

Abstract

The invention discloses an offshore wind plant active power distribution method considering steady-state fan response errors, which relates to the technical field of power distribution, and comprises the following steps: selecting characteristic quantities strongly correlated with the fan response errors by utilizing a Pearson's correlation coefficient method, and carrying out fuzzification processing on the characteristic data to obtain an active power distribution result; concluding a fuzzy rule between the response error and the characteristic quantity of the fan, establishing a category mapping relation between real-time target power and the response error, and solving an optimal fan action number when dispatching is issued; introducing a time convolutional network to carry out deep learning on the historical feature data, and establishing an error function between the response error and the feature quantity of each fan; and establishing an active power distribution model based on a real-time layering result and an error function of each fan by taking the minimum response error of the whole wind power plant to the scheduling instruction as a target, and solving the model by using an adaptive particle swarm algorithm. The method has better effects in the aspects of response precision and reliability.

Description

Technical Field

[0001] The present invention relates to the technical field of power distribution, and in particular to an offshore wind farm active power distribution method taking into account steady-state wind turbine response errors. Background Art

[0002] Offshore wind power is an important strategic support for promoting the transformation of my country's energy structure and ensuring energy security. By the end of 2022, my country's total installed capacity of offshore wind power will be about 31.5GW; it is expected that by 2030, my country's total installed capacity of offshore wind power will reach 58.8GW, becoming the country with the largest cumulative installed capacity of offshore wind power. On the one hand, large-scale offshore wind power is mostly connected to the eastern power grid. Due to the large-scale feed-in of ultra-high voltage transmission in the west, the local synchronous power supply is greatly reduced and the system backup capacity is insufficient. Improving the response accuracy of offshore wind farms to dispatching instructions is the basis for exerting their active support for the power grid. On the other hand, the operating environment of offshore wind turbines is complex and changeable, and wind power output has strong intermittent and volatile characteristics, which brings difficulties to the formulation of reasonable power generation plans. In order to achieve friendly grid connection of offshore wind power, it is necessary to obey the dispatching requirements of the power grid dispatching center. Therefore, when the real-time prediction accuracy of offshore wind power is difficult to be greatly improved, improving the response accuracy of offshore wind farms to power grid dispatching instructions and optimizing the active power distribution problem when offshore wind farms participate in dispatching response can reduce the economic losses caused by dispatching response errors.

[0003] When receiving the dispatching instructions from the control center, the offshore wind turbine adjusts its power output by controlling the speed and pitch angle of the unit to achieve the dispatching target power value. However, if the dispatching instructions change frequently, the pitch angle and generator torque of the unit will frequently move, causing the wind turbine to bear complex alternating loads, thereby reducing the life of the wind turbine. Therefore, when performing minute-level active power distribution and dispatching, the response capability of each unit should be fully considered, so as to select a part of the wind turbines for power adjustment to meet the needs of the dispatching instructions.

[0004] At present, there are few studies on the optimization of active power distribution schemes for offshore wind farms, and most of them are studies on onshore wind farms. Traditional wind farm active power distribution mostly adopts average distribution or distributes active power to wind farms according to the operating status of the units and the predicted wind speed. Due to the limitation of prediction ability and only considering the single random factor of wind speed, the response accuracy of wind turbines needs to be improved. The current research does not consider the difference in wind speed in the area where each unit of the wind farm is located, and the classification method cannot fully reflect the differences in the power generation capacity and operating status of the units. In general, previous studies have ignored the randomness of multiple factors such as the environment and the impact of differences in the operating status of the units on the power response accuracy, resulting in large errors in the output power of wind farms. Summary of the invention

[0005] In view of the above-mentioned problems, the present invention is proposed.

[0006] Therefore, the technical problem solved by the present invention is that the existing offshore wind farm active power optimization allocation method has insufficient response accuracy, ignores the randomness of multiple factors such as the environment, and how to consider the difference in wind speed in the area where each unit of the wind farm is located.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions: an offshore wind farm active power distribution method taking into account the steady-state wind turbine response error, including using the Pearson correlation coefficient method to select characteristic quantities that are strongly correlated with the wind turbine response error, and fuzzifying the characteristic data, summarizing the fuzzy rules between the wind turbine response error and the characteristic quantity, establishing a category mapping relationship between the real-time target power and the response error, and solving the optimal number of wind turbine actions when the dispatch is issued. A time convolution network is introduced to perform deep learning on historical characteristic data, and an error function between the response error of each wind turbine and the characteristic quantity is established. Based on the real-time stratification results and the error functions of each wind turbine, an active power distribution model is established with the goal of minimizing the response error of the entire wind farm to the dispatch instruction, and the model is solved using an adaptive particle swarm algorithm.

[0008] As a preferred scheme of the offshore wind farm active power allocation method taking into account the steady-state wind turbine response error described in the present invention, the processing of selecting feature quantities strongly correlated with the wind turbine response error using the Pearson correlation coefficient method includes using the Pearson correlation coefficient |ρ|>0.5 to perform feature dimension reduction, and obtaining the target power value, power output value, wind speed, air density, wind direction, and temperature as the optimal feature subset for wind turbine response error evaluation.

[0009] As a preferred solution of the offshore wind farm active power distribution method taking into account the steady-state wind turbine response error described in the present invention, wherein: the solution to the optimal number of wind turbine actions when the scheduling is issued includes giving the center point of each membership function through the k-means++ clustering algorithm, using the triangular membership function to fuzzy the data set of wind turbine attributes, each attribute obtains a set of class labels, and each class label value is represented by the cluster center value. Based on the fuzzified data set, the fuzzy ID3 algorithm is used to select the attribute with the smallest average fuzzy classification entropy as the segmentation attribute, and a power response error fuzzy decision tree is established. After the fuzzy decision tree is constructed, the entire fuzzy decision tree model is traversed, each category of the target power is traversed, and the category mapping relationship between the target power and the response error is obtained by using fuzzy rules. All mapping results are sorted from small to large according to the class label value of the response error, and the top N wind turbines with the smallest sum of response errors and the sum of target powers that meet the full-field scheduling instruction value are selected for scheduling.

[0010] As a preferred solution of the offshore wind farm active power allocation method taking into account the steady-state wind turbine response error described in the present invention, wherein: the introduction of the time convolution network to perform deep learning on the historical feature data includes dividing the screened feature data into a training feature set and a verification feature set. In the training feature set, X i TS is the multi-dimensional input feature of the model, and verifies the feature set Y i TS Another input of the model is the final steady-state response error of the wind turbine after receiving the power regulation command.

[0011] The validation feature set only inputs feature X i TE To evaluate the accuracy of the model, the training feature set and the validation feature set are expressed as:

[0012]

[0013]

[0014] Among them, i represents the number of each wind turbine, t is the length of the data sequence in the training feature set, and h is the length of the data sequence in the verification feature set. and From top to bottom, they are the wind turbine target power value, power output value, wind speed, air density, wind direction, and temperature of the same dimension under the data sequence length. For The actual power error response value under the corresponding dimension.

[0015] As a preferred solution of the offshore wind farm active power allocation method taking into account the steady-state wind turbine response error of the present invention, the error function between the response error of each wind turbine and the characteristic quantity is established, including performing a mapping relationship between the input and output characteristics of the TCN network offline training, and quantitatively calculating the steady-state response error of the wind turbine. The power response error is expressed as:

[0016]

[0017] in, is the steady-state response error of the fan numbered i, f i (x) is the functional mapping relationship between the six input features of wind turbine i trained by the TCN network and the steady-state response error, V, ρ, D, and T are the target power value, power output value, wind speed, air density, wind direction, and temperature of wind turbine i, respectively.

[0018] As a preferred solution of the offshore wind farm active power allocation method taking into account the steady-state wind turbine response error of the present invention, wherein: the establishment of the active power allocation model with the goal of minimizing the response error of the entire wind farm to the dispatching instruction includes establishing an objective function of reducing the steady-state response error of the entire station when the wind farm receives the dispatching instruction, which is expressed as:

[0019]

[0020] Establish constraint conditions, active power balance constraints between wind farm output value and dispatch command value:

[0021]

[0022]

[0023] Among them, P T It is the total dispatching instruction value issued by the wind farm to the wind turbine. is the power regulation of fan i, is the target power value and power output value of wind turbine i at this moment.

[0024] Operating output constraints of the fan:

[0025]

[0026] in, is the rated power of fan i. is the maximum power that wind turbine i can generate at the current moment.

[0027] As a preferred solution of the offshore wind farm active power distribution method taking into account the steady-state wind turbine response error described in the present invention, the method of solving the model using the adaptive particle swarm algorithm includes generating an error sorting sequence of all wind turbines in the field by evenly distributing the power adjustment amount and based on the experimental results of the wind turbine error function, selecting the first k wind turbines in the sorting sequence to participate in power regulation, and in the process of hierarchical sorting, only inputting parameters into the error function to calculate the dispatching instruction adjustment amount ΔP of the whole field. T , expressed as:

[0028]

[0029] Known target power category range for each layer Determine the interval range of the dispatch instruction PT and calculate the number of wind turbine layers m and number k that need to participate in the dispatch response.

[0030] Calculate the active power experimental adjustment ΔP of a single fan test , expressed as:

[0031] ΔP test =ΔP T / k

[0032] For all the fans in the field, the active output of the fans at the current moment is and ΔP test The difference is The error function of each fan is input, and the fan error sorting sequence is generated based on the error size output by the error function of each fan. The first k fans with smaller response errors in the error sorting are selected for power adjustment, and the PSO algorithm is used to find the optimal solution for the adjustment amount of k fans.

[0033] Another object of the present invention is to provide an offshore wind farm active power distribution system that takes into account the steady-state wind turbine response error, which can reduce the impact of the difference in wind speed in the area where each unit is located by calculating the response error, thereby solving the problem that the current offshore wind farm active power distribution includes considering the difference in wind speed in the area where each unit of the wind farm is located.

[0034] As a preferred solution of the offshore wind farm active power distribution system taking into account the steady-state wind turbine response error described in the present invention, it includes: a data processing module, an error calculation module, and a distribution module. The data processing module is used to select characteristic quantities that are strongly correlated with the wind turbine response error using the Pearson correlation coefficient method, and fuzzify the characteristic data, summarize the fuzzy rules between the response error of the wind turbine and the characteristic quantity, establish the category mapping relationship between the real-time target power and the response error, and solve the optimal number of wind turbine actions when the dispatch is issued. The error calculation module is used to introduce a time convolution network to perform deep learning on historical characteristic data, and establish an error function between the response error of each wind turbine and the characteristic quantity. The distribution module is used to establish an active power distribution model based on the real-time stratification results and the error functions of each wind turbine, with the goal of minimizing the response error of the entire wind farm to the dispatch instruction, and solve the model using an adaptive particle swarm algorithm.

[0035] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of an offshore wind farm active power distribution method taking into account steady-state wind turbine response errors.

[0036] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of an offshore wind farm active power distribution method taking into account steady-state wind turbine response errors.

[0037] Beneficial effects of the present invention: The active power distribution method for offshore wind farms taking into account the response error of steady-state wind turbines provided by the present invention establishes a model of response error under time-varying conditions based on historical data, adopts a fuzzy decision tree method to roughly evaluate the range of response error and combines the precise response error model mined by the TCN network, and proposes an active power distribution method taking into account the response error, thereby improving the response accuracy of the wind farm when the dispatch is issued. The influence of the difference in wind speed in the area where each unit is located is reduced by calculating the response error, and the calculation result is more reliable. The present invention achieves better results in terms of response accuracy and reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0039] Figure 1 An overall flow chart of an offshore wind farm active power distribution method taking into account steady-state wind turbine response errors provided in the first embodiment of the present invention.

[0040] Figure 2 A flowchart of constructing fuzzy rules for an offshore wind farm active power allocation method taking into account steady-state wind turbine response errors is provided for the first embodiment of the present invention.

[0041] Figure 3 A flow chart of the relationship between target power and power response error classification obtained based on fuzzy rules in an offshore wind farm active power allocation method taking into account steady-state wind turbine response errors provided in the first embodiment of the present invention.

[0042] Figure 4 A sorting diagram of mapping results based on power response error class label values ​​of an offshore wind farm active power allocation method taking into account steady-state wind turbine response errors provided by the first embodiment of the present invention.

[0043] Figure 5 A wind turbine stratification strategy diagram based on a mapping result between target power and power response error categories of an offshore wind farm active power allocation method taking into account steady-state wind turbine response errors provided in a first embodiment of the present invention.

[0044] Figure 6 An expanded causal convolution structure diagram of an offshore wind farm active power allocation method taking into account steady-state wind turbine response errors provided in the first embodiment of the present invention.

[0045] Figure 7A residual module and deep TCN structure diagram of an offshore wind farm active power allocation method taking into account steady-state wind turbine response errors provided in the first embodiment of the present invention.

[0046] Figure 8 A process diagram of solving active power allocation based on wind turbine response error of an offshore wind farm active power allocation method taking into account steady-state wind turbine response error provided by the first embodiment of the present invention.

[0047] Fig. 9 A k-means++ algorithm is used to obtain a clustering number map of input attributes for an offshore wind farm active power allocation method taking into account steady-state wind turbine response errors provided in a second embodiment of the present invention.

[0048] Fig.10 A No. 22 wind turbine classification fuzzy decision tree for an offshore wind farm active power allocation method taking into account steady-state wind turbine response errors provided in a second embodiment of the present invention.

[0049] Fig.11 A fuzzy rule number mined for each wind turbine in an offshore wind farm according to a method for allocating active power of an offshore wind farm taking into account steady-state wind turbine response errors provided by a second embodiment of the present invention.

[0050] Fig.12 A comparison chart of the accuracy and recall rate of each wind turbine based on a fuzzy decision tree and a decision tree for an offshore wind farm active power allocation method taking into account steady-state wind turbine response errors provided in a second embodiment of the present invention.

[0051] Fig.13 A comparison diagram of evaluation indicators of a validation set model of each wind turbine in an offshore wind farm active power allocation method taking into account steady-state wind turbine response errors provided in a second embodiment of the present invention.

[0052] Fig.14 A comparison diagram of mean values ​​of evaluation indexes of offshore wind farm error function verification results of an offshore wind farm active power allocation method taking into account steady-state wind turbine response errors provided by a second embodiment of the present invention.

[0053] Fig.15 A power response error curve diagram of an offshore wind farm according to different allocation schemes of an offshore wind farm active power allocation method taking into account steady-state wind turbine response errors provided by a second embodiment of the present invention. DETAILED DESCRIPTION

[0054] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0055] Example 1, reference Figure 1-Figure 8 , which is an embodiment of the present invention, provides an offshore wind farm active power allocation method taking into account steady-state wind turbine response errors, comprising:

[0056] S1: The Pearson correlation coefficient method is used to select characteristic quantities that are strongly correlated with the wind turbine response error, and the characteristic data is fuzzy processed. The fuzzy rules between the wind turbine response error and the characteristic quantity are summarized, and the category mapping relationship between the real-time target power and the response error is established to solve the optimal number of wind turbine actions when the dispatch is issued.

[0057] Furthermore, since the relationship between the fan response error and the characteristic quantity is often unclear, and both the fan response error and the characteristic quantity are continuous quantities, a large number of continuous quantities have information redundancy, resulting in a large amount of real-time stratification calculation and unsatisfactory stratification effect. The fuzzy decision tree combines the advantages of fuzzy theory and decision tree, has high interpretability, is easy to handle fuzzy and uncertainty problems, and reduces noise interference while retaining the influence of characteristic quantities on the results.

[0058] Fuzzy decision tree is used to mine classification rules with 6 characteristic quantities as input attributes and response error as target attribute, and the classification rules are used to evaluate the power response capability of wind turbines in real time. When the characteristic quantities other than target power and the dispatching instructions of the whole station are known, the category mapping relationship between target power and response error is established according to the mined fuzzy classification rules, and the power response capability of wind turbines is roughly evaluated according to the mapping relationship, so as to stratify the wind turbines.

[0059] It should be noted that the use of the Pearson correlation coefficient method to select feature quantities that are strongly correlated with the wind turbine response error includes using the Pearson correlation coefficient |ρ|>0.5 for feature dimensionality reduction, and obtaining the target power value, power output value, wind speed, air density, wind direction, and temperature as the optimal feature subset for wind turbine response error evaluation.

[0060] S2: Introduce a temporal convolutional network to perform deep learning on historical feature data and establish an error function between the response error of each wind turbine and the feature quantity.

[0061] Furthermore, solving the optimal number of wind turbine actions when scheduling is issued includes giving the center point of each membership function through the k-means++ clustering algorithm, using the triangular membership function to fuzzify the data set of wind turbine attributes, and each attribute obtains a set of class labels. Each class label value is represented by the cluster center value. Based on the fuzzified data set, the fuzzy ID3 algorithm is used to select the attribute with the smallest average fuzzy classification entropy as the segmentation attribute, and a power response error fuzzy decision tree is established. After the fuzzy decision tree is constructed, the entire fuzzy decision tree model is traversed, each category of the target power is traversed, and the category mapping relationship between the target power and the response error is obtained by using fuzzy rules. All mapping results are sorted from small to large according to the class label value of the response error, and the top N wind turbines with the smallest sum of response errors and the sum of target powers that meet the overall scheduling instruction value are selected for scheduling.

[0062] It should be noted that the K-means++ algorithm that avoids local optimization is first used to cluster the data set, and the K-means++ algorithm is optimized using the silhouette coefficient index to obtain the optimal number of clusters mmax and mmax cluster centers of the K-means++ algorithm. The segmentation points of the triangular membership function of each attribute are determined according to the cluster center. After the data is fuzzified by the triangular membership function, each attribute can be divided into mmax category labels, and each class label value is represented by mmax cluster center values. For each attribute cluster, the value of the cluster center can be regarded as the representative discrete value of each cluster. By mapping the data point of each attribute to the discrete value corresponding to the cluster center to which it belongs, the continuous data can be effectively mapped to mmax class labels. The class label divides each attribute domain into mmax partitions, and each generated partition has two tangent points: the lower tangent point and the upper tangent point.

[0063] It should also be noted that fuzzy decision tree induction has two main components: the fuzzy decision tree construction process and the process of assigning class labels to new examples. The fuzzy ID3 algorithm is used to generate a fuzzy decision tree by selecting the root node and child nodes and stopping the growth of the decision tree. 1 ,x 2 ,…,x n}, there are l attributes that can be selected as non-leaf node B, and the attributes can be expressed as A (1) ,A (2) ,...,A (l) Each fuzzy input attribute A (s) (1≤s≤l) has k s fuzzy semantic values, respectively Decision attribute A (l+1) Has m fuzzy semantic values, which are For each attribute Its relative to the category on non-leaf node B The relative frequency is defined as:

[0064]

[0065] Where M represents the cardinality measure of the fuzzy set, ∩ represents the intersection operation of the fuzzy set, represents the degree of membership of x to the attribute value, represents the degree of membership of x to the category, μ B (x) represents the membership of x to the non-leaf node B.

[0066] For each attribute value Its fuzzy classification entropy on non-leaf node B is:

[0067]

[0068] For each attribute A (s) (1≤s≤l), the average fuzzy classification entropy on non-leaf node B is:

[0069]

[0070] Where ω i Indicates the sth attribute value A (s) The i-th attribute value of The weight is calculated as follows:

[0071]

[0072] Step 1: Select the root node. Calculate the average fuzzy classification entropy of the six input attributes and select the attribute with the smallest average fuzzy classification entropy as the root node.

[0073] Step 2: Expand the root node. Expand the selected root node, generate its child nodes, and use the fuzzy partitioning rule of the attribute to create the child nodes.

[0074] Step 3: Calculate the truth of internal nodes. Traverse each internal node and calculate its truth. The truth refers to the proportion of the main category of samples on the node.

[0075] Step 4: Compare the truth with the threshold β. Compare the truth of the internal node with the threshold β. If the truth is greater than β, mark the internal node as a leaf node. If the internal node uses all attributes, mark it as a leaf node. Otherwise, select the unused attribute with the smallest average fuzzy classification entropy as the extended attribute and continue to generate child nodes.

[0076] The true level β is a threshold value, which plays a vital role in the generation of decision tree leaves. It directly controls the generation of fuzzy decision tree leaves. The membership calculation formula of a sample of class C at node A. Too high or too low β values ​​will affect the effect of the fuzzy decision tree. In the present invention, the value of β is 0.7. Repeat steps 2-4 until the entire decision tree is constructed.

[0077]

[0078] The decision tree is converted into a set of if-then rules. Each branch path from the root node to the leaf node can be converted into a fuzzy rule. The if-then rule is a category mapping rule between the characteristic quantity and the power response error. The fuzzy decision rule of each wind turbine provides a real-time stratification basis for the wind turbines in the wind farm.

[0079] Furthermore, the fuzzy decision tree of each wind turbine can be converted into a set of fuzzy classification rules. The mined fuzzy rules are used to roughly evaluate the power response capability of the wind turbine during real-time power distribution. Before the wind farm distributes active power, the five attributes of each wind turbine's input attributes except the target power are known quantities. Based on the fuzzy rules of each wind turbine, the category mapping relationship between the target power and power response error of all wind turbines can be obtained, such as Figure 3 As shown. There are i max There are 10 wind turbines and SCADA collects real-time data of each wind turbine once a minute. According to the upper and lower tangent points of the input attribute class label partition of each wind turbine, class labels are assigned to the real-time data with known attributes. For the i-th wind turbine, its target power and wind turbine power response error are divided into p imax Class and m imax When the target power category is p, according to the matching degree between the class label value of the 6 real-time conditional attributes and the rule, it is judged that when the target power is p, the wind turbine power response error is m. Therefore, the interval of the response error can be obtained, so as to judge the performance of the wind turbine power regulation task. imax class, get p imax The target power and power response error category mapping results. max The same operation is performed on all wind turbines to obtain the current The category mapping results provide a reference for active power allocation in wind farms.

[0080] It should be noted that in order to preferentially select wind turbines to participate in power regulation, the wind turbines in the wind farm are layered. The wind turbines in the entire farm are divided into different levels, and based on the size of the dispatching instructions, the number of wind turbines in the corresponding level is preferentially selected to participate in power regulation. Figure 3The schematic diagram of the wind turbine stratification strategy based on the target power and power response error mapping results is as follows:

[0081] Step 1: According to the above category mapping relationship flow chart, find the p of the i-th wind turbine at the current moment imax The category mapping result of target power and power response error (p imax →m imax ), wind farm i max The typhoon machine received Category mapping results.

[0082] Step 2: When performing cluster analysis, it is ensured that there are only m categories of power response errors of all wind turbines in the wind farm. max Class. Figure 4 As shown, The class label values ​​of the power response errors are sorted from small to large, and the error categories after sorting can form at most m max The target power categories mapped with the response errors are also sorted in this order.

[0083] Step 3: Retrieve the upper and lower tangent points of the class label partition to obtain the error category set {m 1 ,m 2 ,…,m max} The corresponding category interval upper and lower tangent point set {(m 1min ,m 1max ),(m 2min ,m 2max ),…(m maxmin ,m maxmax )}, and the upper and lower tangent points (p imin ,p imax ).

[0084] Step 4: Figure 5 As shown, the number k of fans in the error class label m1 is obtained m1 , record the number r1 of target power categories in m1 class labels that are mapped one-to-one with the response error categories. In the error set m1, select k from the r1 mapping results m1 The maximum target power category of the wind turbines is k. m1 The upper and lower tangent points of the target power category interval are added to obtain a new set At this time, the total dispatching instruction value issued by the wind farm to the wind turbine is in the set When , we get the k of the first layer that the wind farm needs to dispatch m1 A fan.

[0085] Step 5: If the instruction is greater than When considering the next error set m2 . Get the error set m 1 and m 2 The number of fans k m2 and number, record m 1 and m 2 The number of target power categories in the set that are mapped one-to-one with the response error categories r 2 . In r 2 Select k from the mapping results m2 The maximum target power category of the wind turbines is k. m2 The upper and lower tangent points of the target power category interval are added to obtain a new set At this time, the total dispatching instruction value issued by the wind farm to the wind turbine is in the set When , we can get the k of the first and second layers that the wind farm needs to dispatch m2 This process is repeated until a set of total dispatching instruction values ​​that satisfy the current wind farm station’s requirements for wind turbines is found, thereby dispatching the number of wind turbines k in the set.

[0086] Furthermore, TCN is based on a one-dimensional fully convolutional network and integrates the structures of causal convolution, dilated convolution and residual connection. Causal convolution means that different layers have causal relationships and can maintain the same length and width as the input layer, without missing historical information and using future information, which meets the requirements of solving the power response error problem of the controller. TCN can effectively understand the features at different scales in the time series through convolution operations. The convolution kernel extracts local features by sliding on the input sequence. Since the scheduling data volume of offshore wind farms is large and the time span is long, the following is adopted: Figure 6 By expanding the causal convolution structure as shown in the figure, TCN can capture global patterns in a longer range by increasing the filter size k and the expansion coefficient d, thereby accepting a wider range of input information and increasing the receptive field. However, as the depth of TCN increases, it will bring about problems such as gradient explosion and gradient disappearance. Therefore, by Figure 7 The residual module shown on the left builds a deep TCN structure. The residual module corrects the error of the network output to avoid the degradation of the convolution layer. By stacking the residual modules, TCN can achieve the construction of a deeper network. The residual module of TCN is mainly composed of an expanded causal convolution layer, weight normalization, activation function, and random inactivation. The 1×1 convolution ensures that the input and output data dimensions of the residual module are the same.

[0087] It should be noted that the introduction of a temporal convolutional network to perform deep learning on historical feature data includes dividing the selected feature data into a training feature set and a verification feature set. Multi-dimensional input features for the model, validation feature set Another input of the model is the final steady-state response error of the wind turbine after receiving the power regulation command.

[0088] Verify that the feature set only has input features To evaluate the accuracy of the model, the training feature set and the validation feature set are expressed as:

[0089]

[0090]

[0091] Among them, i represents the number of each wind turbine, t is the length of the data sequence in the training feature set, and h is the length of the data sequence in the verification feature set. and From top to bottom, they are the wind turbine target power value, power output value, wind speed, air density, wind direction, and temperature of the same dimension under the data sequence length. For The actual power error response value under the corresponding dimension.

[0092] Establishing the error function between the response error and characteristic quantity of each wind turbine includes the mapping relationship between the input and output characteristics of the TCN network offline training, and quantitatively calculating the steady-state response error of the wind turbine. The power response error is expressed as:

[0093]

[0094] in, is the steady-state response error of the fan numbered i, f i (x) is the functional mapping relationship between the six input features of wind turbine i trained by the TCN network and the steady-state response error, V, ρ, D, and T are the target power value, power output value, wind speed, air density, wind direction, and temperature of wind turbine i, respectively.

[0095] It should also be noted that when the six characteristic inputs of the wind turbine are known, the response error of the wind turbine can be qualitatively solved through the error function of TCN network training, but the target power is the power command issued by the control center, which is a substitute quantity. Therefore, the wind farm control center can use the PSO algorithm to find the optimal target power value of each wind turbine based on the power response error function of each wind turbine, with the goal of minimizing the sum of the power response errors of the wind turbines in the entire station, so as to improve the accuracy of the overall dispatch response of the power station.

[0096] S3: Based on the real-time stratification results and the error functions of each wind turbine, an active power distribution model is established with the goal of minimizing the response error of the entire wind farm to the dispatch command, and the model is solved using an adaptive particle swarm algorithm.

[0097] Furthermore, the solution process of wind farm active power allocation based on wind turbine response error is as follows: Figure 8 As shown in the figure, first, the offline training module uses the TCN network to extract the characteristics of massive data sequences and construct the wind turbine error function. Secondly, the offline mining module uses the fuzzy rules mined by the fuzzy decision tree. Finally, the online application module uses the error function and fuzzy rules of the wind turbine as the basis for active power allocation to improve the accuracy of the overall dispatch response of the wind farm.

[0098] It should be noted that, with the goal of minimizing the response error of the entire wind farm to the dispatching instruction, establishing the active power allocation model includes establishing an objective function to reduce the steady-state response error of the entire station when the wind farm receives the dispatching instruction, which is expressed as:

[0099]

[0100] Establish constraint conditions, active power balance constraints between wind farm output value and dispatch command value:

[0101]

[0102] Among them, P T It is the total dispatching instruction value issued by the wind farm to the wind turbine. is the power regulation of fan i, is the target power value and power output value of wind turbine i at this moment.

[0103] Operating output constraints of the fan:

[0104]

[0105] in, is the rated power of fan i. is the maximum power that wind turbine i can generate at the current moment.

[0106] It should also be noted that the use of the adaptive particle swarm algorithm to solve the model includes generating an error sorting sequence of all fans in the field by evenly distributing the power adjustment amount and based on the experimental results of the fan error function, selecting the first k fans in the sorting sequence to participate in power adjustment, and in the process of hierarchical sorting, only inputting parameters into the error function to calculate the dispatch instruction adjustment amount ΔP for the whole field. T , expressed as:

[0107]

[0108] Known target power category range for each layer Determine the interval range of the dispatch instruction PT and calculate the number of wind turbine layers m and number k that need to participate in the dispatch response.

[0109] Calculate the active power experimental adjustment ΔP of a single fan test, expressed as:

[0110] ΔP test =ΔP T / k

[0111] For all the fans in the field, the active output of the fans at the current moment is and ΔP test The difference is The error function of each fan is input, and the fan error sorting sequence is generated based on the error size output by the error function of each fan. The first k fans with smaller response errors in the error sorting are selected for power adjustment, and the PSO algorithm is used to find the optimal solution for the adjustment amount of k fans.

[0112] Example 2, reference Figure 9-14 , which is an embodiment of the present invention, provides an offshore wind farm active power distribution method taking into account steady-state wind turbine response errors. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0113] First, the example data comes from an offshore wind farm with an installed capacity of 100.8MW, which consists of 28 wind turbines with a rated capacity of 3.6MW. The SCADA data of the 28 units from June 2018 to June 2019 are used for training and mining. The offline training of fuzzy decision trees and TCN networks and the online solution of the power allocation value of each wind turbine in the wind farm using the adaptive PSO algorithm are all completed on the Anaconda3 simulation software platform. The program is compiled in Python3.8 and uses extension libraries such as Fuzzytree, Scikit-learn, Keras and TensorFlow.

[0114] SCADA data may contain a large amount of abnormal data and duplicate data. Too much duplicate data may complicate the calculation and may lead to unnecessary output. Before data mining and analysis, SCADA data needs to be preprocessed and the data is converted into a similar scale using standardized methods to improve the convergence speed and accuracy of model training and delete abnormal data that deviates from the distribution of most samples. When offshore wind turbines do not need to participate in the dispatch response, they are often in the MPPT operation state. Therefore, it is necessary to exclude historical data in the MPPT state. In addition, exclude data in the off-grid and shutdown states to avoid increasing training noise. After optimization and screening, a total of 18,396 valid data moments were obtained.

[0115] The SCADA data of 28 wind turbines in a wind farm were used for research. The fuzzy rules of the 28 wind turbines were mined according to the historical data to provide a basis for the real-time stratification of wind turbines. The response errors of all wind turbines were clustered using the K-means++ algorithm. When the contour coefficient was the largest, the errors were divided into 4 categories. Therefore, the upper and lower tangent points of the response error E1-E4 categories were (0, 22) (22, 43) (43, 56) (89, 150) respectively. Each piece of historical data contains 7 parameters. The input attributes are target power value (kW), power output value (kW), wind speed (m / s), air density (kg / m3), wind direction (°), and temperature (℃). The target attribute is the response error category (E1-E4). The historical data of the wind turbine input attributes were clustered, partitioned, and fuzzified. The partitions of the input attributes were represented by class labels and membership degrees. Based on the classification ability of the fuzzy decision tree, the category relationship between the input attributes and the target attributes, i.e., fuzzy rules, was established to qualitatively evaluate the response error of the wind turbine. Each wind turbine mines fuzzy rules according to the steps of data clustering, data fuzzification, and building a fuzzy decision tree. The following specifically shows the process and results of fuzzy rule mining for wind turbine No. 22. Due to the qualitative evaluation of the response error of the wind turbine, 60 data are randomly selected from the valid data every month, for a total of 720 samples, including 80 errors in category 1, 268 errors in category 2, 304 errors in category 3, and 68 errors in category 4. The input attribute related parameters of the samples are shown in Table 1, and the 5-fold cross validation method is used.

[0116] Table 1 Brief statistical analysis of sample data parameters

[0117] property Mean Standard Deviation Power output (kW) 1665.3 634.3 wind direction(.) 102.2 16.7 Target power value (kW) 1334.1 585.4 Wind speed (m / s) 5.9 2.51 Temperature(℃) 15.15 0.58 Air density (kg / m3) 1.3 0.2

[0118] Fig. 9 The relationship between the number of clusters and the average silhouette coefficient obtained by clustering sample data using the K-means++ algorithm is presented. The average silhouette coefficient is in the range of [-1,1], and the larger the value, the better the clustering. It can be seen that when the value of the number of clusters K is 2 and the silhouette coefficient is the largest, the clustering effect is the best. Therefore, in the present invention, the number of clusters is set to 2. Therefore, each fuzzy input attribute has two fuzzy categories, as shown in Table 2.

[0119] Table 2 Fuzzy input attribute center point

[0120]

[0121]

[0122] The continuous data of the 22nd fan are fuzzy processed. According to the triangular membership function equation and the cluster center in Table 2, the membership function corresponding to each fuzzy input attribute is calculated, and the continuous fan data is converted into a membership matrix. Fig.10As shown, according to the method in 2.3, the fuzzy decision tree of wind turbine No. 22 is established, and each path from the root node to the leaf node of the fuzzy decision tree is converted into the corresponding fuzzy classification rules, as shown in Table 3. Fig.10 It contains 1 root node, 12 internal nodes and 14 leaf nodes. Each leaf node corresponds to a classification rule in Table 4. In this way, the classification rules of the fuzzy decision tree can be used to roughly evaluate the range of the fan response error.

[0123] Table 4 Fuzzy classification rules

[0124]

[0125]

[0126] Fuzzy rules are mined for each wind turbine. The number of fuzzy rules for each wind turbine is as follows: Fig.11 As shown in Figure 2, to show the accuracy of fuzzy decision tree classification, the prediction results of each wind turbine fuzzy decision tree and decision tree classifier are compared, as shown in Figure 2. Fig.12 As shown in Figure 3, the accuracy and recall of the fuzzy decision tree have stronger classification capabilities than the traditional decision tree. Therefore, the proposed fuzzy decision tree model is helpful for class prediction of response errors under real-time variable conditions, with sufficient accuracy, and can give reliable prediction results.

[0127] The effective data set of 18,396 moments obtained after optimization and screening is divided into the first 98% of the data as the training set and the last 2% of the data as the validation set for training the fan response error function. In order to verify the accuracy of the fan error function output regulation error extracted by the TCN network, the accuracy of the training model of the long short-term memory network and the gated recurrent unit network is compared with that of the TCN network, and the root mean squared error (RMSE) and mean absolute percentage error (MAPE) are selected as evaluation indicators of the offline training results of the neural network. The accuracy analysis of the output results of the three neural network validation sets of 28 fans is shown in the figure. Fig.13 As shown in Figure 2, it can be seen that TCN can better summarize historical data, and the RMSE and MAPE of each wind turbine are the smallest. For the entire wind farm, Fig.14 As shown in the figure, the RMSE and MAPE of the TCN network validation set are 0.06kW and 0.75% respectively, which has stronger time series feature extraction capabilities than traditional neural networks LSTM and GRU. Therefore, the wind turbine response error function extracted by TCN training can accurately calculate the steady-state power response error of the wind turbine, which can be used as a reliable basis for the active power allocation model.

[0128] In order to further verify the optimization effect, the commonly used average allocation, rated capacity proportional allocation and the optimized allocation scheme of the present invention are used to simulate and analyze the dispatch response. During the period from 00:00 to 02:30, the wind speed of offshore wind power is relatively stable, and the offshore wind farm is in a dispatch response state. It is necessary to continuously adjust the power according to the command value issued by the dispatch center every minute. Fig.15 The error curves of the wind farm response experiment when the above three power allocation schemes are applied are recorded. The purple curve in the figure is the error curve of the scheme of the present invention. The scheme of the present invention can effectively reduce the power response error of the wind farm. The average steady-state response error of the scheme of the present invention to the dispatching instruction is 26.41kW, which is reduced by 60.05% and 44.68% respectively compared with the average allocation and rated capacity proportional allocation schemes.

[0129] Embodiment 3, an embodiment of the present invention provides an offshore wind farm active power distribution system taking into account steady-state wind turbine response errors, including a data processing module, an error calculation module, and a distribution module.

[0130] Among them, the data processing module is used to select characteristic quantities that are strongly correlated with the wind turbine response error using the Pearson correlation coefficient method, and to fuzzify the characteristic data, summarize the fuzzy rules between the wind turbine response error and the characteristic quantity, establish the category mapping relationship between the real-time target power and the response error, and solve the optimal number of wind turbine actions when the dispatch is issued; the error calculation module is used to introduce a time convolution network to perform deep learning on historical characteristic data, and establish an error function between the response error of each wind turbine and the characteristic quantity; the allocation module is used to establish an active power allocation model based on the real-time stratification results and the error function of each wind turbine, with the goal of minimizing the response error of the entire wind farm to the dispatch instruction, and solve the model using an adaptive particle swarm algorithm.

[0131] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0132] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0133] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0134] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc. It should be noted that the above embodiments are only used to illustrate the technical solution of the present invention and are not limited. Although the present invention is described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

[0135] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for allocating active power of an offshore wind farm taking into account steady-state wind turbine response errors, characterized in that: include: The Pearson correlation coefficient method is used to select characteristic quantities that are strongly correlated with the response error of the wind turbine, and the characteristic data is fuzzy processed to summarize the fuzzy rules between the response error of the wind turbine and the characteristic quantity, establish the category mapping relationship between the real-time target power and the response error, and solve the optimal number of wind turbine actions when the dispatch is issued; A temporal convolutional network is introduced to conduct deep learning on historical feature data, and an error function between the response error of each wind turbine and the feature quantity is established; Based on the real-time stratification results and the error function of each wind turbine, an active power distribution model is established with the goal of minimizing the response error of the entire wind farm to the dispatch command, and the model is solved using an adaptive particle swarm algorithm.

2. The offshore wind farm active power allocation method taking into account steady-state wind turbine response errors according to claim 1, characterized in that: The method of using the Pearson correlation coefficient to select feature quantities that are strongly correlated with the fan response error includes using the Pearson correlation coefficient |ρ|>0.5 to perform feature dimension reduction, and obtaining the six feature quantities of target power value, power output value, wind speed, air density, wind direction, and temperature as the optimal feature subset for fan response error evaluation.

3. The offshore wind farm active power allocation method taking into account steady-state wind turbine response errors as claimed in claim 2, characterized in that: The method for solving the optimal number of wind turbine actions when scheduling is issued includes giving the center point of each membership function through the k-means++ clustering algorithm, using the triangular membership function to fuzzify the data set of wind turbine attributes, each attribute obtains a set of class labels, and each class label value is represented by the cluster center value. Based on the fuzzified data set, the fuzzy ID3 algorithm is used to select the attribute with the smallest average fuzzy classification entropy as the segmentation attribute, and a power response error fuzzy decision tree is established. After the fuzzy decision tree is constructed, the entire fuzzy decision tree model is traversed, each category of the target power is traversed, and the category mapping relationship between the target power and the response error is obtained by using fuzzy rules. All mapping results are sorted from small to large according to the class label value of the response error, and the top N wind turbines with the smallest sum of response errors and the sum of target powers that meet the full-field scheduling instruction value are selected for scheduling.

4. The method for allocating active power of an offshore wind farm taking into account steady-state wind turbine response errors as claimed in claim 3, characterized in that: The introduction of the temporal convolutional network to perform deep learning on historical feature data includes dividing the selected feature data into a training feature set and a verification feature set. Multi-dimensional input features for the model, validation feature set is another input of the model, i.e., the final steady-state response error of the wind turbine after receiving the power regulation command; Verify that the feature set only has input features To evaluate the accuracy of the model, the training feature set and the validation feature set are expressed as: Where i represents the number of each wind turbine, t is the length of the data sequence in the training feature set, and h is the length of the data sequence in the verification feature set. and From top to bottom, they are the wind turbine target power value, power output value, wind speed, air density, wind direction, and temperature of the same dimension under the data sequence length. For The actual power error response value under the corresponding dimension.

5. The method for allocating active power of an offshore wind farm taking into account steady-state wind turbine response errors as claimed in claim 4, characterized in that: The error function between the response error of each wind turbine and the characteristic quantity is established by performing a mapping relationship between input and output characteristics of the TCN network offline training, and quantitatively calculating the steady-state response error of the wind turbine. The power response error is expressed as: in, is the steady-state response error of the fan numbered i, f i (x) is the functional mapping relationship between the six input features of wind turbine i trained by the TCN network and the steady-state response error, V, ρ, D, and T are the target power value, power output value, wind speed, air density, wind direction, and temperature of wind turbine i, respectively.

6. The method for allocating active power of an offshore wind farm taking into account steady-state wind turbine response errors according to claim 5, characterized in that: The objective of establishing the active power allocation model with the minimum response error of the entire wind farm to the dispatching instruction is to establish an objective function for reducing the steady-state response error of the entire station when the wind farm receives the dispatching instruction, which is expressed as: Establish constraint conditions, active power balance constraints between wind farm output value and dispatch command value: Among them, P T It is the total dispatching instruction value issued by the wind farm to the wind turbine. is the power regulation of fan i, is the target power value and power output value of wind turbine i at this moment; Operating output constraints of the fan: in, is the rated power of fan i; is the maximum power that wind turbine i can generate at the current moment.

7. The method for allocating active power of an offshore wind farm taking into account steady-state wind turbine response errors according to claim 6, characterized in that: The method of solving the model using the adaptive particle swarm algorithm includes generating an error sorting sequence of all fans in the field by evenly distributing the power adjustment amount and based on the experimental results of the fan error function, selecting the first k fans in the sorting sequence to participate in power adjustment, and only inputting the parameters into the error function during the hierarchical sorting process to calculate the dispatch instruction adjustment amount ΔP of the whole field. T , expressed as: Known target power category range for each layer Determine the interval range of the dispatch instruction PT, and calculate the number of wind turbine layers m and number k that need to participate in the dispatch response; Calculate the active power experimental adjustment ΔP of a single fan test , expressed as: ΔP test =ΔP T / k For all the fans in the field, the active output of the fans at the current moment is and ΔP test The difference is The error function of each fan is input, and the fan error sorting sequence is generated based on the error size output by the error function of each fan. The first k fans with smaller response errors in the error sorting are selected for power adjustment, and the PSO algorithm is used to find the optimal solution for the adjustment amount of k fans.

8. A system adopting the offshore wind farm active power allocation method taking into account steady-state wind turbine response errors as claimed in any one of claims 1 to 7, characterized in that: It includes data processing module, error calculation module and distribution module; The data processing module is used to select characteristic quantities that are strongly correlated with the response error of the fan by using the Pearson correlation coefficient method, and to perform fuzzy processing on the characteristic data, summarize the fuzzy rules between the response error of the fan and the characteristic quantity, establish a category mapping relationship between the real-time target power and the response error, and solve the optimal number of fan actions when the dispatch is issued; The error calculation module is used to introduce a time convolution network to perform deep learning on historical feature data and establish an error function between the response error of each wind turbine and the feature quantity; The allocation module is used to establish an active power allocation model based on real-time stratification results and the error functions of each wind turbine, with the goal of minimizing the response error of the entire wind farm to the dispatching instruction, and solve the model using an adaptive particle swarm algorithm.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the offshore wind farm active power allocation method taking into account steady-state wind turbine response errors according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the offshore wind farm active power allocation method taking into account steady-state wind turbine response errors according to any one of claims 1 to 7 are implemented.

Images