Classification Method for Extreme Weather Wind Power Ramping Events Based on Spatiotemporal Characteristics

Abstract

The present invention discloses a classification method for extreme weather wind power ramp events based on spatio-temporal characteristics, and the steps thereof include: 1. Filtering the historical wind power sequence; 2. Detecting the ramp events in the historical wind power; 3. Generating classification labels for the ramp events; 4. Screening out the meteorological characteristics of extreme weather; 5. Generating spatio-temporal characteristic data for classifying the ramp events; 6. Training a ConvLSTM classifier to classify and identify the wind power ramp events in the wind farm. The present invention detects the wind power ramp events under extreme weather, screens the meteorological characteristics affecting the ramp events, combines the spatio-temporal correlation characteristics of multiple wind farms, and further realizes the classification of the wind power ramp events in the wind farm, so as to provide an important reference for grid dispatching personnel to formulate control and dispatching strategies for different ramp events.

Description

Technical Field

[0001] The present invention relates to the field of wind power ramp detection and classification, and specifically refers to a method for classifying extreme weather wind power ramp events based on spatio-temporal characteristics. Background Art

[0002] Wind power ramp events are closely related to meteorological factors, with strong randomness and uncertainty, and have an important impact on the safe operation of power systems. When extreme weather occurs, it is extremely easy to trigger wind farm ramp events: when extreme weather such as cyclones, thunderstorms, and gusts occurs, it will cause a sudden increase in the power of the wind farm and an up-ramp event will occur; when the wind speed is too high and exceeds the cut-out wind speed, the wind turbines will shut down one after another, which will cause a sudden drop in the power of the wind farm and a down-ramp will occur. When a wind power ramp event occurs, its output power suddenly increases or decreases within a short period of time, which will seriously disrupt the supply-demand balance and safe dispatching of the power system, and even cause large-scale power outages and power system collapses in extreme weather. Therefore, when formulating the system operation plan, grid dispatchers have to consider the impact of wind power ramps. However, the impacts of different wind power ramp events on the grid operation are not the same. In order to more economically and effectively respond to the impacts of different ramp events, different dispatching strategies often need to be adopted. Therefore, studying the classification of wind power ramp events is of great significance for the safe operation of power grids with a high proportion of wind power access.

[0003] Currently, the classification methods for wind power ramp events mainly include: (1) classifying using the wind power data of ramp events; (2) classifying using ramp characteristics such as the power amplitude change, duration, and ramp rate when the ramp event occurs, etc. However, directly classifying using power data has a large input feature dimension, and the power data is easily affected by the noise of the measurement device, and the classification effect needs to be further improved; when classifying using ramp characteristics, although it can effectively distinguish ramp events from non-ramp events, it cannot further distinguish ramp events under different meteorological conditions, and the effective information contained in the features is less. In addition, the above classification methods do not consider the impact of meteorological factors, nor do they consider the spatial distribution characteristics of wind farms. Based on the classification results of ramp events by the above classification methods, the differences between different types of ramp events cannot be reflected, which is not conducive to in-depth analysis of the causes of different ramp events, and thus effective control and dispatching strategies cannot be formulated, which is not conducive to the safe and economic operation of the power system. Summary of the Invention

[0004] In order to solve the deficiencies of the above-mentioned existing technologies, the present invention proposes a classification method for extreme weather wind power ramp events based on spatio-temporal characteristics, aiming to further distinguish wind power ramp events under different meteorological conditions and different spatial distributions, realize the classification of wind power ramp events under extreme weather, and thus provide an important reference for grid dispatching personnel to formulate control and dispatching strategies for different ramp events, and further be able to formulate targeted control and dispatching strategies to ensure the safe operation of the power system.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0006] The classification method for extreme weather wind power ramp events based on spatio-temporal characteristics of the present invention is characterized by being carried out according to the following steps:

[0007] Step 1. Filter the historical wind power sequence:

[0008] Obtain the wind power sequence {p0(t)|t = 0, 1, 2, …, N - 1} of wind farm W0, where: t represents the sampling time; N represents the total number of sampling times; p0(t) represents the wind power of wind farm W0 at sampling time t;

[0009] After performing first-order exponential smoothing filtering on the wind power sequence {p0(t)|t = 0, 1, 2, …, N - 1} using Equation (1), obtain the filtered wind power sequence at sampling time t

[0010]

[0011] In Equation (1): α represents the smoothing exponent, α ∈ [0, 1];

[0012] Step 2. Detect ramp events in the historical wind power:

[0013] Judge whether the filtered wind power in the wind power filtered sequence satisfies the wind power ramp event defined by Equation (2). If it satisfies, it means that the wind power in the corresponding time period has a wind power ramp event, and it is added to the ramp event sample set {P k}, where P k represents the k ramp event sample sets, and t k represents the occurrence time of the kth ramp event, k = 1, 2, …, K, and K is the total number of ramp events in the sample set {P k};

[0014]

[0015] In Equation (2): P ref is the ramp threshold, P ref= w·P cap , P cap is the rated capacity of the wind farm, w is the threshold coefficient; Δt is the sampling interval of wind power; M is the number of sampling points within the ramp time window; (M - 1)·Δt represents the duration of the ramp time window;

[0016] Step 3. Generate ramp event classification labels:

[0017] Traverse the ramp event sample set {P k}, and according to the length of the ramp time window, the initial and final wind powers within the ramp time window, and the maximum power within the ramp time window, divide all the ramp events in the ramp event sample set {P k} into F types of ramp events, and the classification labels are respectively denoted as 1, 2, …, F;

[0018] Step 4. Screen out extreme weather meteorological characteristics:

[0019] Obtain the sequence of H meteorological factors {q 0,i (t)|t = 0, 1, 2, …, N - 1} of the wind farm W0, q 0,i (t) represents the value of the i-th meteorological factor of the wind farm W0 at the sampling time t, i = 1, 2, …, H, and H represents the total number of meteorological factors;

[0020] Perform first-order exponential filtering and smoothing on each meteorological factor using Equation (3) to obtain the filtered meteorological factor sequence Calculate the filtered wind power sequence of the wind farm W0 using Equation (4) and the correlation between each filtered meteorological factor sequence , and screen out the meteorological characteristics with ρ i > ρ ref as the extreme weather meteorological characteristics, thus obtaining C extreme weather meteorological characteristics, where ρ ref represents the correlation threshold;

[0021]

[0022]

[0023] In Equations (3) and (4): ρ i represents the correlation between the filtered wind power sequence and the filtered sequence of the i-th meteorological factor ; represents the mean value of the filtered wind power sequence, represents the mean value of the filtered sequence of the i-th meteorological factor,

[0024] Step 5. Generate spatio-temporal feature data for ramp event classification:

[0025] Obtain the wind power data {p d (t)|d = 1, 2, …, D} and extreme weather meteorological factor data {q d (t)|d = 1, 2, …, D, c = 1, 2, …, C} of D wind farms {W d,c |d = 1, 2, …, D} within the same geographical range as wind farm W0, and use them together as the spatio-temporal features for ramp event classification of wind farm W0; where p d (t) represents the wind power of the d-th wind farm at sampling time t, and q d,c (t) represents the value of the c-th extreme weather meteorological factor of the d-th wind farm at sampling time t;

[0026] Step 6. Train a ConvLSTM classifier to classify and identify wind farm wind ramp events;

[0027] Step 6.1. Normalize the spatio-temporal features of the wind ramp events to obtain the normalized spatio-temporal features, denoted as {p′ d (t)|d = 1, 2, …, D} and {q′ d,c (t)|d = 1, 2, …, D, c = 1, 2, …, C} respectively, where p′ d (t) represents the normalized wind power value of the d-th wind farm at sampling time t; q′ d,c (t) represents the normalized value of the c-th extreme weather meteorological feature of the d-th wind farm at sampling time t;

[0028] Step 6.2. For the k-th ramp event in the ramp event sample set {P k}, use Equation (5) to obtain the spatio-temporal features at the first sampling time t k in the corresponding ramp time window

[0029]

[0030] Similarly, find the spatio-temporal features at other sampling times t k + 1, t k + 2, …, t k + M - 1 in the ramp time window corresponding to the k-th ramp event where M is the number of sampling points in the ramp time window, thus constructing the spatio-temporal features within the corresponding ramp time window

[0031] Step 6.3. Use Equation (6) to construct the loss function L of the ConvLSTM classifier:

[0032]

[0033] In formula (6): represents the weight of correct classification; represents the weight of the output classification being j;

[0034] Step 6.4. Input the spatio-temporal feature X k into the ConvLSTM classifier for training, and continuously optimize the classifier parameters during the training process until the loss function L reaches the minimum, then stop the training, so as to obtain the trained extreme weather wind power ramp event classification model for classifying and identifying the wind power ramp events in the wind farm.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0036] The present invention takes into account the influence of the spatial distribution characteristics of wind power on the ramp events in the wind farm, and considers the influence of meteorological conditions at different geographical locations on the classification of wind farm ramp events. By detecting the wind farm ramp events, considering the spatio-temporal correlation between the wind farm power, meteorological conditions and ramp events in different spaces, a wind farm ramp event classifier based on ConvLSTM is established to achieve more accurate classification of wind farm ramp events. The proposed method not only considers the influence of meteorological characteristics of adjacent wind farms on ramp events under extreme weather, but also can accurately classify the ramp events in the wind farm, providing a reference basis for the wind farm or dispatching department to suppress wind power fluctuations or formulate targeted control strategies. The specific effects are as follows:

[0037] 1. The present invention uses Step 1 to smooth the wind farm power data and filter out the noise in the data, so as to reduce the influence of noise data on the detection of ramp events. Combining with Step 2 can improve the detection rate of ramp events and avoid over-detection, missed detection or misdetection of ramp events;

[0038] 2. The present invention uses Step 4 to screen the meteorological characteristics with a large correlation with wind power ramp events, and combines with Step 5 to obtain the spatio-temporal characteristics affecting wind power ramp, considering the correlation between the wind farm power, meteorological characteristics and the occurrence of wind power ramp events under different spatial distributions, which can improve the accuracy of wind power ramp event classification;

[0039] 3. The present invention uses Step 6.1 to normalize the wind farm power and meteorological characteristics, which can improve the training efficiency and accuracy of the classifier; the present invention uses Step 5 and Step 6.2 to construct the spatio-temporal characteristics input to the classifier, which can not only ensure the temporal continuity of different characteristics, but also reflect the spatial correlation between various characteristics, and consider the spatio-temporal correlation between wind power, meteorological characteristics and wind power ramp events;

[0040] 4. The present invention uses step 6.4 to train a ConvLSTM classifier, which can obtain the categories of different wind power ramp events based on spatio-temporal features, providing more information for wind farm or grid dispatchers to formulate more secure and effective control and dispatch strategies. Description of the Drawings

[0041] Figure 1 is a flowchart of the method of the present invention. Detailed Embodiments

[0042] In this embodiment, as Figure 1 shown, a method for classifying extreme weather wind power ramp events based on spatio-temporal features is carried out according to the following steps:

[0043] Step 1. Filter the historical wind power sequence:

[0044] Obtain the wind power sequence {p0(t)|t = 0, 1, 2, …, N - 1} of wind farm W0, where: t represents the sampling time; N represents the total number of sampling times; p0(t) represents the wind power of wind farm W0 at sampling time t;

[0045] After performing first-order exponential smoothing filtering on the wind power sequence {p0(t)|t = 0, 1, 2, …, N - 1} using Equation (1), the filtered wind power sequence at sampling time t is obtained

[0046]

[0047] In Equation (1): α represents the smoothing exponent, α ∈ [0, 1];

[0048] Using first-order exponential smoothing to smooth the wind power can reduce the influence of noise data on the detection of wind power ramp events, more accurately describe the variation law of wind power data, improve the detection rate of wind power ramp events, and avoid multiple detections, missed detections or misdetections caused by noise data.

[0049] Step 2. Detect ramp events in the historical wind power:

[0050] Judge whether the filtered wind power in the wind power filtered sequence satisfies the wind power ramp event defined by Equation (2). If it satisfies, it means that the wind power in the corresponding period has a wind power ramp event, and it is added to the ramp event sample set {P k}, where: t k represents the occurrence time of the kth ramp event, k = 1, 2, …, K, and K is the total number of ramp events in the sample set {P k};

[0051]

[0052] In formula (2): P ref is the ramp threshold, P ref = w·P cap , P cap is the rated capacity of the wind farm, w is the threshold coefficient; Δt is the sampling interval of wind power; M is the number of sampling points within the ramp time window; (M - 1)·Δt represents the duration of the ramp time window;

[0053] Step 3. Generate ramp event classification labels:

[0054] Traverse the ramp event sample set {P k}}, and based on the length of the ramp time window, the initial and final wind powers within the ramp time window, and the maximum power within the ramp time window, divide all the ramp events in the ramp event sample set {P k}} into F types of ramp events, and the classification labels are denoted as 1, 2, …, F respectively;

[0055] Step 4. Screen out extreme weather meteorological characteristics: Obtain the H meteorological factor sequences {q 0,i (t)|t = 0, 1, 2, …, N - 1} of the wind farm W0, q 0,i (t) represents the value of the i-th meteorological factor of the wind farm W0 at the sampling moment t, i = 1, 2, …, H, and H represents the total number of meteorological factors;

[0056] Perform first-order exponential filtering and smoothing on each meteorological factor using formula (3) to obtain the filtered meteorological factor sequence Calculate the filtered wind power sequence of the wind farm W0 using formula (4) and the correlation between the filtered sequences of each meteorological factor , and screen out the meteorological characteristics with ρ i > ρ ref as the extreme weather meteorological characteristics, so as to obtain C extreme weather meteorological characteristics, where ρ ref represents the correlation threshold;

[0057]

[0058]

[0059] In formulas (3) and (4): ρ i represents the correlation between the filtered wind power sequence and the filtered sequence of the i-th meteorological factor ; represents the mean value of the filtered wind power sequence, represents the mean of the filtered sequence of the i-th meteorological factor,

[0060] The Pearson correlation coefficient is used to quantify the correlation between meteorological factors and wind power ramping events. Meteorological features with a strong correlation with wind power ramping events are screened out according to the size of the correlation, which is conducive to reducing the complexity of the classification model, reducing the adverse effects of unnecessary features on the classification effect, and improving the classification accuracy.

[0061] Step 5. Generate spatiotemporal feature data for hill climbing event classification:

[0062] Get D wind farms {W d |d=1,2,…,D}wind power data {p d (t)|d=1,2,…,D} and extreme weather meteorological factor data {q d,c (t)|d=1,2,…,D,c=1,2,…,C}, and together they are used as the spatiotemporal features for the classification of the wind farm W0 ramp event. d (t) represents the wind power of the d-th wind farm at sampling time t, q d,c (t) represents the value of the cth extreme weather meteorological factor of the dth wind farm at sampling time t;

[0063] Step 6. Train the ConvLSTM classifier to classify and identify wind power ramp events in wind farms;

[0064] Step 6.1. Normalize the spatiotemporal characteristics of wind power ramp events to obtain normalized spatiotemporal characteristics, which are denoted as {p′ d (t)|d=1,2,…,D} and {q′ d,c (t)|d=1,2,…,D,c=1,2,…,C}, where: p′ d (t) represents the normalized wind power value of the d-th wind farm at sampling time t; q′ d,c (t) represents the normalized value of the cth extreme weather meteorological characteristic of the dth wind farm at sampling time t;

[0065] Normalizing different features separately can avoid mutual interference between features due to different orders of magnitude, and is also conducive to improving the training efficiency and classification accuracy of the classifier. In addition, the power between adjacent wind farms has a strong spatial correlation. In order to improve the classification accuracy of ramp events, the power characteristics and meteorological characteristics of adjacent wind farms are considered as the influencing factors of wind power ramp events, which is conducive to further distinguishing ramp events with different spatiotemporal characteristics and different meteorological conditions.

[0066] Step 6.2. For the hill climbing event sample set {Pk For the k-th ramp event in}, the first sampling moment t in the corresponding ramp time window is obtained using Equation (5). k spatiotemporal characteristics

[0067]

[0068] Similarly, the spatiotemporal characteristics of other sampling moments t+1, t+2, …, t+M-1 within the ramp time window corresponding to the k-th ramp event can be obtained. k +1, t k +2, …, t k +M - 1 can be obtained. where M is the number of sampling points within the ramp time window; thus, the spatiotemporal characteristics within the corresponding ramp time window are constructed.

[0069] Using Equation (5) to construct the input spatiotemporal characteristics of the classifier can ensure the temporal continuity of different features and is also conducive to exploring the spatial correlation of different features, thereby improving the accuracy of the ConvLSTM classifier.

[0070] Step 6.3. Construct the loss function L of the ConvLSTM classifier using Equation (6):

[0071]

[0072] In Equation (6): represents the weight of the correct classification; represents the weight of the output classification as j;

[0073] Step 6.4. Input the spatiotemporal feature X k into the ConvLSTM classifier for training, and the output is the classification label corresponding to the wind power ramp event. During the training process, continuously optimize the classifier parameters until the loss function L reaches the minimum, and then stop the training to obtain the trained classification model for extreme weather wind power ramp events, which is used to classify and identify the wind power ramp events in the wind farm.

[0074] Optimizing the parameters of the ConvLSTM classifier according to the loss function shown in Equation (6) is beneficial to improving the accuracy of correct classification and reducing the misclassification rate. In addition, the output result of the classifier is the category label corresponding to different types of ramp events, which can provide more information for the wind farm or grid dispatcher and serve as a basis for formulating control and dispatching strategies.

Claims

1. A classification method for extreme weather wind power ramp events based on spatio-temporal characteristics, characterized in that Proceed as follows: Step 1. Filter the historical wind power sequence: Obtain the wind power sequence {p0(t)|t = 0, 1, 2, …, N - 1} of wind farm W0, where: t represents the sampling time; N represents the total number of sampling times; p0(t) represents the wind power of wind farm W0 at sampling time t; After performing first-order exponential smoothing filtering on the wind power sequence {p0(t)|t = 0, 1, 2, …, N - 1} using Equation (1), the wind power filtering sequence at sampling time t is obtained In formula (1): α represents the smoothing exponent, α ∈ [0, 1]; Step 2. Detect the ramp events in the historical wind power; Judge the filtered wind power in the wind power filtering sequence to see if it satisfies the wind power ramp event defined by Equation (2). If it does, it means that the wind power has a wind power ramp event during the corresponding period, and it is added to the ramp event sample set {P k}, where P k represents the k ramp event sample sets, and t k represents the occurrence time of the k-th ramp event, k = 1, 2, …, K, and K is the total number of ramp events in the sample set {P k}; In formula (2): P ref is the ramp threshold, P ref = w·P cap , P cap is the rated capacity of the wind farm, w is the threshold coefficient; Δt is the sampling interval of wind power; M is the number of sampling points within the ramp time window; (M - 1)·Δt represents the duration of the ramp time window; Step 3. Generate ramp event classification labels; Traverse the ramp event sample set {P k}, and according to the length of the ramp time window, the initial and final wind power within the ramp time window, and the maximum power within the ramp time window, divide all the ramp events in the ramp event sample set {P k} into F types of ramp events, and the classification labels are denoted as 1, 2, …, F respectively; Step 4. Screen out the meteorological characteristics of extreme weather; Obtain the sequences of H meteorological factors {q 0,i (t)|t = 0, 1, 2, …, N - 1} for wind farm W0, where q 0,i (t) represents the value of the i-th meteorological factor of wind farm W0 at sampling time t, i = 1, 2, …, H, and H represents the total number of meteorological factors; Perform first-order exponential filtering and smoothing on each meteorological factor using Equation (3) to obtain the filtered meteorological factor sequence Calculate the filtered wind power sequence of wind farm W0 using Equation (4) And the correlation with each filtered meteorological factor sequence And select the meteorological characteristics with ρ i > ρ ref As extreme weather meteorological characteristics, thus obtaining C extreme weather meteorological characteristics, where ρ ref Represents the correlation threshold; In formulas (3) and (4): ρ i represents the correlation between the wind power filtering sequence and the filtering sequence of the i-th meteorological factor ; represents the mean value of the wind power filtering sequence, represents the mean value of the filtering sequence of the i-th meteorological factor, Step 5. Generate spatio-temporal feature data for ramp event classification; Obtain the wind power data {p d (t)|d = 1, 2, …, D} of D wind farms {W d |d = 1, 2, …, D} that are within the same geographical range as wind farm W0, and the extreme weather meteorological factor data {q d,c (t)|d = 1, 2, …, D, c = 1, 2, …, C}, and together use them as the spatio-temporal features for classifying the ramping events of wind farm W0; where p d (t) represents the wind power of the d-th wind farm at sampling time t, and q d,c (t) represents the value of the c-th extreme weather meteorological factor of the d-th wind farm at sampling time t; Step 6. Train the ConvLSTM classifier to classify and identify the wind farm wind power ramp events; Step 6.

1. Normalize the spatio-temporal characteristics of the wind power ramp event to obtain the normalized spatio-temporal characteristics, denoted as {p′ d (t)|d = 1, 2, …, D} and {q′ d,c (t)|d = 1, 2, …, D, c = 1, 2, …, C}, where p′ d (t) represents the normalized wind power value of the d-th wind farm at the sampling time t; q′ d,c (t) represents the normalized value of the c-th extreme weather meteorological characteristic of the d-th wind farm at the sampling time t; Step 6.

2. For the k-th ramp event in the ramp event sample set {P k}, the spatio-temporal features at the first sampling moment t k in the corresponding ramp time window are obtained using Equation (5) Similarly, find the other sampling times t in the climbing time window corresponding to the kth climbing event k +1,t k +2,…,t k +M-1's spatiotemporal characteristics Among them, M is the number of sampling points in the climbing time window, so as to construct the spatiotemporal features in the corresponding climbing time window Step 6.

3. Construct the loss function L of the ConvLSTM classifier using formula (6); In formula (6): represents the weight of correct classification; represents the weight of the output classification being j; Step 6.

4. Input the spatio-temporal feature X k into the ConvLSTM classifier for training, and continuously optimize the classifier parameters during the training process until the loss function L reaches the minimum, then stop the training, so as to obtain the trained extreme weather wind power ramp event classification model for classifying and identifying wind power ramp events in the wind farm.

Images