A method for predicting short-term wind power ramping
The CNN-LSTM-AM model and BA-SDT algorithm identify wind power power climbing events, which solves the problem of low recognition accuracy in the prior art, and improves the accuracy of wind power power prediction and grid stability.
Patent Information
- Application Number
- CN202311459264.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-11-03
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2043-11-03
AI Technical Summary
The existing wind power climbing event recognition methods are complex and have low recognition accuracy, which affects the accuracy of wind power prediction.
The CNN-LSTM-AM model is used to combine the BA-SDT algorithm to preprocess data, train the model, identify the climbing event and evaluate its intensity, and regularly verify the model accuracy and make adjustments to improve the accuracy of wind power prediction.
More accurate wind power prediction is achieved, wind farm operation and management efficiency is improved, wind energy fluctuations affect the instability of the power system, and grid stability and renewable energy predictability.
Smart Images

Figure CN117578404B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of wind power prediction, and in particular to a method for predicting short-term wind power ramping. Background Art
[0002] As a renewable energy source, wind power has been vigorously developed around the world. With the increase in the penetration rate of wind power in the power grid, its randomness and volatility will seriously affect the safety and stable operation of the power system. Therefore, it is very important to carry out wind power forecasting. For the power grid side, accurate power forecasting results can guide the power grid to reserve rotating reserve capacity in advance, timely load regulation, and ensure the safe and stable operation of the power grid. Wind power forecasting helps wind power companies to reasonably arrange maintenance plans and improve economic benefits and profitability. In addition, as a grid-connected power source, wind power needs to participate in the day-ahead power market bidding. Therefore, the dependence and demand for wind power forecasting are increasing.
[0003] Accurately identifying the characteristics of ramp events will be beneficial to wind power forecasting. However, existing wind power ramp event identification methods are either too complex, resulting in low algorithm execution efficiency, or have low recognition accuracy, which affects the subsequent power forecast accuracy. Therefore, exploring more accurate and effective wind power forecasting methods is of great practical significance. Summary of the invention
[0004] The present invention provides a method for predicting short-term wind power ramping, which solves the problem in the prior art that the wind power ramping event identification method is complex and the identification accuracy is low.
[0005] In order to solve the above-mentioned invention object, the technical solution provided by the present invention is as follows:
[0006] A method for predicting short-term wind power ramping includes the following steps:
[0007] S1. Obtain sample prediction data and preprocess the sample prediction data;
[0008] S2. According to the preprocessed sample prediction data, the initial CNN-LSTM-AM power prediction model is trained to obtain a trained CNN-LSTM-AM power prediction model;
[0009] S3, determining the power prediction sequence data based on the preprocessed sample prediction data and the trained CNN-LSTM-AM power prediction model;
[0010] S4. Based on the constructed BA-SDT ramp identification model, the power prediction sequence data is identified to obtain the predicted power of the wind turbine, and a ramp event is determined based on the predicted power of the wind turbine to evaluate the intensity index of the ramp event;
[0011] S5. Obtain the measured power of wind power, determine the measured power sequence data of the wind turbine according to the measured power of wind power, and verify the accuracy of the CNN-LSTM-AM power prediction model based on the measured power sequence data of the wind turbine and the power prediction sequence data to obtain an accuracy index;
[0012] S6. If the accuracy index is lower than the set accuracy threshold, the wind power measured power is preprocessed, and the preprocessed wind power measured power is integrated with the wind turbine prediction data and input into the CNN-LSTM-AM power prediction model for training.
[0013] A wind power short-term power ramp prediction device, comprising:
[0014] A preprocessing module, used to obtain sample prediction data and preprocess the sample prediction data;
[0015] A training module is used to train the initial CNN-LSTM-AM power prediction model according to the preprocessed sample prediction data to obtain a trained CNN-LSTM-AM power prediction model;
[0016] A determination module is used to determine the power prediction sequence data based on the preprocessed sample prediction data and the trained CNN-LSTM-AM power prediction model;
[0017] A determination module is used to identify the power prediction sequence data based on the constructed BA-SDT ramp identification model, obtain the predicted power of the wind turbine, determine the ramp event based on the predicted power of the wind turbine, and evaluate the intensity index of the ramp event;
[0018] A verification module is used to obtain the measured power of wind power, determine the measured power sequence data of the wind turbine according to the measured power of wind power, and verify the accuracy of the CNN-LSTM-AM power prediction model based on the measured power sequence data of the wind turbine and the power prediction sequence data to obtain an accuracy index;
[0019] The retraining module is used to preprocess the wind power measured power if the accuracy index is lower than the set accuracy threshold, integrate the preprocessed wind power measured power with the wind turbine prediction data, and then input them into the CNN-LSTM-AM power prediction model for training.
[0020] Compared with the prior art, the above technical solution has at least the following beneficial effects:
[0021] (1) This wind power short-term power ramp prediction method can more effectively plan and manage the operation of wind farms, maximize the use of wind energy resources, and improve power generation efficiency through accurate wind power prediction, identification and evaluation of ramp events, and maintenance of model accuracy. By identifying and evaluating the intensity of ramp events, wind power fluctuations can be predicted and corresponding measures can be taken to reduce the instability of the power system and improve the stability of the power grid.
[0022] (2) The wind power short-term power ramp prediction method uses automated methods such as the BA algorithm and the SDT algorithm. The model can automatically adjust parameters and tolerance coefficients to provide intelligent support for decision-making and improve efficiency. Accurate power forecasting and ramp event identification can improve the predictability of renewable energy and make the power system easier to manage and plan.
[0023] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0025] Figure 1 A schematic diagram of the BA-SDT algorithm for predicting short-term power ramping of wind power provided by the present invention;
[0026] Figure 2 A flow chart of a method for predicting short-term wind power ramping provided by the present invention;
[0027] Figure 3 A flow chart of a wind power short-term power ramp prediction method for ramp determination provided by the present invention;
[0028] Figure 4 A device block diagram of a wind power short-term power ramp prediction device provided by the present invention. DETAILED DESCRIPTION
[0029] In order to make the purpose, technical solution and advantages of the embodiment of the present invention clearer, the technical solution of the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings of the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all of the embodiments. Based on the described embodiment of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0030] Unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, words such as "one", "one" or "the" do not indicate a quantitative limitation, but indicate the existence of at least one. Words such as "include" or "comprise" mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Words such as "connect" or "connected" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect.
[0031] It should be noted that the terms "up", "down", "left", "right", "front", "back", etc. used in the present invention are only used to indicate relative position relationships. When the absolute position of the object being described changes, the relative position relationship may also change accordingly.
[0032] The present invention aims to realize similarity retrieval and real-time push of large-scale data, and can be used for various problems that require recommendations based on vector data, and provides a method for predicting short-term wind power ramping. The key point of the overall concept is to build an accurate wind power prediction model by comprehensively utilizing historical power, weather forecast, theoretical power and startup capacity data, supplemented by the identification and evaluation of ramping events, so as to improve the operating efficiency of wind farms and reduce the instability of wind energy fluctuations on the power system. At the same time, the model is regularly verified and improved to improve the predictability and manageability of renewable energy and achieve more sustainable and reliable wind power generation.
[0033] The CNN-LSTM-AM model uses the power time series as the input of the network. Referring to the representation method of word vectors in natural language, the power value at a certain moment is input as a single vector, and then input into the model in chronological order using a sliding window to obtain a feature map, so that the CNN's ability to extract features can be utilized, thereby filtering the effective content in the time series.
[0034] In actual use, the climbing features extracted from the power signal after CNN recognition are used as the input of LSTM. The three gates and memory units unique to LSTM can further screen the climbing features extracted by CNN, so that important features are retained through memory units, and unimportant features are forgotten through the forget gate. Finally, the output of LSTM is weighted through the attention mechanism to further highlight important features.
[0035] Due to the strong randomness of power, multi-step prediction is difficult and the accuracy is generally low. In real-time power prediction, the result of the next sampling point is unknown, and it is impossible to determine whether a long-term power ramp occurs, so it is more accurate to predict adjacent time points. The characteristic signal of ramping is easily submerged in a large amount of random and chaotic data. It is difficult for ordinary models to establish a mapping relationship from data input to ramp recognition output, and the accuracy of the model is bound to be low. A power ramp prediction method based on convolutional neural networks and long short-term memory neural networks is proposed. In order to solve the problem of difficult extraction and learning of ramp features, it is proposed to use an attention mechanism to strengthen the learning of ramp features. In view of the limitations of the artificial setting of the attention mechanism weights, it is proposed to use a fully connected layer to construct the weights of the attention mechanism to ensure the effective learning of the ramp features. Finally, the prediction ability of the model is verified.
[0036] The BA-SDT algorithm is based on the standard SDT algorithm. Through the optimization of the BA algorithm, the parameters that affect the compressibility of the SDT algorithm, namely the tolerance coefficient ΔE, are obtained. The principle diagram of the BA-SDT algorithm is shown in the figure. Figure 1 As shown, the parameters are first initialized, then the optimal bat is selected according to the evaluation function, and then the parameters of each bat are updated through loop iterations, and finally the optimal tolerance coefficient ΔE of the SDT algorithm is searched and output.
[0037] The revolving door algorithm (SDT) is often used to compress data in real-time databases. It is a relatively fast linear trend fitting compression algorithm. Its essence is to extract trends from massive data, retain the development trend of the original data, compress massive data into less data, and retain the position of the corresponding compression point in the original sequence. More specifically, the SDT algorithm is to filter the data to be stored by constructing a parallelogram. If the recorded data point is within the constructed parallelogram, the data is discarded and only the data outside the parallelogram is stored. A new compression starts from this point, and then the parallelogram is continued to be constructed. The parallelogram continues to "expand", and the length of the side in the vertical direction is always maintained at 2 times the tolerance coefficient ΔE.
[0038] like Figure 2 As shown, an embodiment of the present invention provides a method for predicting short-term power ramping of wind power, comprising the following steps: S1, obtaining sample prediction data, and preprocessing the sample prediction data;
[0039] S2. According to the preprocessed sample prediction data, the initial CNN-LSTM-AM power prediction model is trained to obtain a trained CNN-LSTM-AM power prediction model;
[0040] S3, determining the power prediction sequence data based on the preprocessed sample prediction data and the trained CNN-LSTM-AM power prediction model;
[0041] S4. Based on the constructed BA-SDT ramp identification model, the power prediction sequence data is identified to obtain the predicted power of the wind turbine, the ramp event is determined based on the predicted power of the wind turbine, and the intensity index of the ramp event is evaluated;
[0042] S5. Obtain the measured power of wind power, determine the measured power sequence data of the wind turbine according to the measured power of wind power, verify the accuracy of the CNN-LSTM-AM power prediction model based on the measured power sequence data of the wind turbine and the power prediction sequence data, and obtain the accuracy index;
[0043] S6. If the accuracy index is lower than the set accuracy threshold, the wind power measured power is preprocessed, and the preprocessed wind power measured power is integrated with the wind turbine prediction data and input into the CNN-LSTM-AM power prediction model for training.
[0044] In this implementation plan, historical actual wind power data, numerical weather forecast data, theoretical power data of wind towers and wind turbines, and wind turbine startup capacity data are collected, and the collected data are preprocessed to ensure data consistency and quality. Based on the preprocessed data, a power prediction model CNN-LSTM-AM, which is a combination of convolutional neural network and long short-term memory network, is established. The model is trained using historical data so that future wind power can be accurately predicted.
[0045] A BA-SDT (Ramp Identification) model is constructed to identify ramp events in power forecast sequences, quantify and evaluate the intensity of ramp events, and thus better understand the future fluctuations of wind farms.
[0046] The accuracy of the CNN-LSTM-AM power prediction model is verified regularly. If the accuracy is below the set threshold, the model is updated and retrained. Measured data can be used to improve the accuracy of the model, especially when facing new environmental conditions or equipment changes.
[0047] Specifically, preprocessing the prediction data includes: determining whether there are abnormal values in the same type of wind towers, wind towers of wind turbines, theoretical power data of wind turbines, and operating capacity data of wind turbines, and if so, deleting the abnormal values and filling in normal values;
[0048] The preprocessing also includes: aligning the historical actual wind power, numerical weather forecast data, the theoretical power of the wind tower and the wind turbine and the operating capacity of the wind turbine according to the set timestamp, and dividing the aligned prediction data into a training set and a validation set according to the set ratio.
[0049] In this implementation scheme, outliers in the theoretical power data of wind towers and wind turbines of the same type and the operating capacity data of wind turbines are judged and deleted to ensure the data quality of the input model. The historical actual power of wind power, numerical weather forecast data, theoretical power data and operating capacity data are aligned according to the set timestamps, which helps to ensure that all data are consistent in time for model training and prediction. The aligned data are divided into training sets and validation sets according to the set ratio, which helps to evaluate the performance and accuracy of the model.
[0050] In the data preprocessing stage, the data quality and consistency of the input model are ensured by cleaning and aligning the data. At the same time, the data is divided into training sets and validation sets to ensure that the established model can make accurate predictions and evaluate fluctuation events on new data, which helps to improve the reliability and practicality of the model and provide more reliable results for the operation and prediction of wind farms.
[0051] Specifically, the process of judging whether there are abnormal values in the theoretical power data of wind towers and wind turbines and the startup capacity data of wind turbines of the same type is as follows: number the wind towers and wind turbines respectively; obtain the pre-stored power data of the wind towers and wind turbines and the pre-stored startup capacity data of the wind turbines under each number; judge whether the difference between the theoretical power data of the wind towers and wind turbines and the pre-stored power data of the wind towers and wind turbines under each number is zero, if not zero, determine the theoretical power data of the wind towers and wind turbines corresponding to the number as abnormal values, delete the abnormal values, and fill in the corresponding pre-stored power data of the wind towers and wind turbines; judge whether the difference between the startup capacity data of the wind turbines under each number and the pre-stored startup capacity data of the wind turbines is zero, if not zero, determine the pre-stored startup capacity data of the wind turbines corresponding to the number as abnormal values, delete the abnormal values, and fill in the corresponding pre-stored startup capacity data of the wind turbines.
[0052] In this implementation scheme, by comparing the difference between theoretical power data and actual pre-stored power data, as well as the difference between startup capacity data and pre-stored startup capacity data, it is possible to identify situations where outliers exist. When outliers are identified, not only are the outliers deleted, but the corresponding pre-stored power data or startup capacity pre-stored data are also supplemented to ensure the integrity of the data set. Cleaning and correcting data can improve the accuracy of the wind power prediction model.
[0053] Specifically, the process of obtaining power prediction sequence data based on the trained CNN-LSTM-AM power prediction model is as follows:
[0054] y i =f(W*x i +b) (1)
[0055] Among them, y iis the output of the convolutional layer, f is the activation function, W is the convolution kernel, b is the bias, and x i is the data in the prediction data;
[0056] The pooling operation of the pooling layer is expressed by the following formula (2):
[0057] z i =pooling(y i ) (2)
[0058] Among them, z i is the output of the convolutional layer;
[0059] The local feature sequence Z containing the time series output by CNN is Z = [z1, z2, z3, ..., z i , ..., z n ] is input to LSTM, hidden units are obtained based on the determined number in LSTM, long-term dependencies in the local feature sequence containing the time series are captured, and the input gate, forget gate and output gate are used for processing and output. At the same time, the dropout layer is used to prevent overfitting, and the deep sequence H = [h1, h2, h3, ..., h i , ..., h n ];
[0060] The deep sequence containing the time series is input into AM, and the attention mechanism is introduced to weight the deep sequence containing the time series, and different weights are assigned to different features to generate the final power prediction sequence data A = [a1, a2, a3, ..., a i , …, a n ], each time step in the power prediction sequence data contains a prediction of future power, where:
[0061]
[0062] Among them, e i is the feature h in the deep sequence containing the time series i The normalized weight of .
[0063] In this embodiment, features can be extracted from multiple data sources, long-term dependencies can be learned, important features can be weighted, and a power prediction sequence can be generated. This power prediction sequence can be used to predict power changes in wind power ramp events for proper operation and management.
[0064] Specifically, Figure 3 As shown, the process of performing climbing judgment is as follows: based on the predicted power of the wind turbine in the climbing identification time window of the set width, determine whether a climbing event occurs within the climbing identification time window of the set width; if so, determine whether the climbing event is an up-climbing event or a down-climbing event.
[0065] The process of determining whether a ramp event occurs within the ramp identification time window of the set width is as follows: obtaining the predicted power of the wind turbine within the ramp identification time window, screening the maximum power value and the minimum power value; and calculating the power difference by the difference between the maximum power value and the minimum power value based on the following formula (3):
[0066] ΔP=|P max(t,t+Δt) -P min(t,t+Δt) | (3)
[0067] Where ΔP is the power difference, P max(t,t+Δt) To identify the maximum power within the time window, P min(t,t+Δt) To identify the minimum power value within the time window;
[0068] It is determined whether the power difference index is greater than a set threshold index. If so, it is determined that a ramp event has occurred; otherwise, it is determined that no ramp event has occurred.
[0069] The process of determining whether the climbing event is an up-climbing event or a down-climbing event is as follows: obtaining the power at the end time and the power at the start time of the climbing identification time window; calculating the ratio of the difference between the power at the end time and the power at the start time to the climbing duration, and the calculation formula is as follows (4):
[0070]
[0071] Among them, Γ is the ratio, Δt is the climbing duration, P t is the power at the beginning of climbing, P t+Δt The power at the end of the climbing; determine whether the ratio is greater than 0, if so, it is determined as an uphill climbing event, otherwise it is determined as a downhill climbing event.
[0072] In this implementation scheme, the process can be divided into two stages: climbing determination and climbing direction identification. First, a climbing identification time window of a certain width is selected to extract the maximum and minimum power values within the time period. Then, the extracted maximum and minimum power values are subtracted and compared with the set power threshold. If the difference is less than the power threshold, no climbing event has occurred. If the difference is greater than the power threshold, it is determined that a climbing event has occurred, and the climbing event direction identification stage is entered. Finally, the ratio of the difference between the power at the end moment and the power at the start moment of the selected time period and the climbing duration is compared with 0. If the ratio is greater than 0, an uphill climbing event has occurred. If the ratio is less than 0, a downhill climbing event has occurred. Among them, the time window width for climbing identification is set to 4 hours, and the power threshold is set to 20% of the rated installed capacity.
[0073] Specifically, the process of evaluating the intensity index of a climbing event is as follows: obtaining the difference between the maximum power and the minimum power of historical climbing events, and calculating the average power difference of historical climbing events; based on the difference between the maximum power and the minimum power of this climbing event and the average power difference of historical climbing events, calculating the intensity index, the calculation formula is:
[0074]
[0075] Where Λ is the intensity index, i = 1, 2, 3, ..., n is the number of historical climbing events, is the maximum power of the i-th historical climbing event, is the minimum power value of the i-th historical climbing event, and e is a natural constant.
[0076] In this implementation, by calculating the average power difference of historical ramping events and comparing it with the power difference of the current ramping event, the intensity of the current ramping event can be evaluated, and the historical ramping event data is used as a benchmark to evaluate the current ramping event, which helps to compare the current event with past situations and better understand the performance of the wind farm.
[0077] Specifically, the process of obtaining the accuracy index is as follows: draw the measured power series data based on the measured power of wind power, set several sampling points within the time window where the power ramp occurs, and sample the measured power at each sampling point; align the power prediction series data with the measured power series data through timestamps, and sample the predicted power; calculate the accuracy index based on the measured power and predicted power after sampling.
[0078] The calculation formula of the accuracy index is:
[0079]
[0080] Where Ψ is the accuracy index, j = 1, 2, 3, ..., n is the number of sampling points, is the predicted power of the jth sampling point, For The measured power at the corresponding sampling point, Θ r is the tolerance index of the measured power and predicted power at each sampling point.
[0081] In this implementation scheme, the accuracy of the prediction model is evaluated by comparing the difference between the measured power and the predicted power and calculating the accuracy index. A number of sampling points are set within the time window where the power ramp occurs, and the measured power and the corresponding predicted power are sampled respectively. For each sampling point, the difference between the measured power and the predicted power is calculated, and their root variance is found. The root variance is divided by the number of sampling points and the tolerance index, which represents the average error between the power predicted by the model and the actually observed power. A smaller value indicates a more accurate prediction.
[0082] By setting the accuracy index, the degree of difference between the power predicted by the model and the actual observed power can be evaluated. The accuracy of the model can be evaluated based on the degree of difference, so that the various parameters of the prediction model can be adjusted and the model can be adjusted purposefully. In addition, the degree of difference can be used to compensate the prediction model. The predicted power can be compensated or adjusted according to the degree of difference, so that the predicted power can be closer to the actual measured power, thereby improving the accuracy of the prediction.
[0083] Specifically, the process of constructing the BA-SDT climbing identification model is as follows: the historical data of the known number of climbing events within a set time is input into the BA-SDT climbing identification model; the evaluation function is set, and the evaluation function is the difference between the number of climbing events identified by the SDT algorithm and the actual number of climbing events within a period of time. The tolerance coefficient is continuously optimized through the BA algorithm and the calculation is iterated multiple times until the fitness of the bat population is optimal when the difference is 0; the tolerance coefficient when the difference is 0 is output, and the BA-SDT climbing identification model is constructed.
[0084] In this implementation scheme, the tolerance coefficient is continuously optimized through the BA algorithm so that the model can adapt to different data and environmental conditions, which helps to improve the robustness of the model without manually adjusting parameters. By comparing the actual number of climbing events with the number of events identified by the SDT algorithm, the model can accurately identify climbing events and achieve the optimal match through iteration, thereby constructing a BA-SDT climbing identification model, which has higher adaptability and reliability in actual data and can better solve the problem of identifying climbing events in wind farms.
[0085] The embodiment of the present invention provides a wind power short-term power ramp prediction device, which is used to implement the above-mentioned wind power short-term power ramp prediction method, such as Figure 4 As shown, the device comprises:
[0086] A preprocessing module 410 is used to obtain sample prediction data and preprocess the sample prediction data;
[0087] A training module 420 is used to train the initial CNN-LSTM-AM power prediction model according to the preprocessed sample prediction data to obtain a trained CNN-LSTM-AM power prediction model;
[0088] A determination module 430 is used to determine power prediction sequence data based on the preprocessed sample prediction data and the trained CNN-LSTM-AM power prediction model;
[0089] The determination module 440 is used to identify the power prediction sequence data based on the constructed BA-SDT ramp identification model, obtain the predicted power of the wind turbine, perform ramp event determination based on the predicted power of the wind turbine, and evaluate the intensity index of the ramp event;
[0090] Verification module 450, used to obtain the measured power of wind power, determine the measured power sequence data of the wind turbine according to the measured power of wind power, and verify the accuracy of the CNN-LSTM-AM power prediction model based on the measured power sequence data of the wind turbine and the power prediction sequence data to obtain an accuracy index;
[0091] The retraining module 460 is used to pre-process the wind power measured power if the accuracy index is lower than the set accuracy threshold, integrate the pre-processed wind power measured power with the wind turbine prediction data, and input them into the CNN-LSTM-AM power prediction model for training.
[0092] Optionally, the forecast data includes historical actual wind power, numerical weather forecast data, theoretical power data of wind towers and wind turbines, and operating capacity data of wind turbines;
[0093] The preprocessing module is used to:
[0094] S11, determining whether there are abnormal values in the wind towers, wind towers of the wind turbines, theoretical power data of the wind turbines, and operating capacity data of the wind turbines of the same type, and if so, deleting the abnormal values and filling in normal values;
[0095] S12. Align the historical actual wind power, numerical weather forecast data, the theoretical power of the wind tower and the wind turbine, and the operating capacity of the wind turbine according to the set timestamp, and divide the aligned prediction data into a training set and a validation set according to the set ratio.
[0096] Optionally, the preprocessing module is used to:
[0097] S111. Number the wind towers and wind turbines respectively;
[0098] S112, obtaining pre-stored power data of the wind tower and the wind turbine under each number and pre-stored startup capacity data of the wind turbine;
[0099] S113, determining whether the difference between the theoretical power data of the wind tower and the wind turbine under each number and the pre-stored power data of the wind tower and the wind turbine is zero; if not, determining the theoretical power data of the wind tower and the wind turbine corresponding to the number as an abnormal value, deleting the abnormal value, and filling in the corresponding pre-stored power data of the wind tower and the wind turbine;
[0100] S114. Determine whether the difference between the startup capacity data of the fan under each number and the pre-stored startup capacity data of the fan is zero. If it is not zero, determine the pre-stored startup capacity data of the fan corresponding to the number as an abnormal value, delete the abnormal value, and fill in the corresponding pre-stored startup capacity data of the fan.
[0101] Optionally, the determining module is used to:
[0102] The preprocessed sample prediction data is input into CNN, and based on the convolution layer and pooling layer of the CNN, the local feature sequence Z=[z1, z2, z3, ..., z i , ..., z n ], the convolution operation of the convolution layer is expressed by the following formula (1):
[0103] y i =f(W*x i +b) (1)
[0104] Among them, y i is the output of the convolutional layer, f is the activation function, W is the convolution kernel, b is the bias, and x i is the data in the prediction data;
[0105] The pooling operation of the pooling layer is expressed by the following formula (2):
[0106] z i =pooling(y i ) (2)
[0107] Among them, z i is the output of the convolutional layer;
[0108] The local feature sequence Z containing the time series output by CNN is Z = [z1, z2, z3, ..., z i , ..., z n ] is input to LSTM, hidden units are obtained based on the determined number in the LSTM, long-term dependencies in the local feature sequence containing the time series are captured, and the input gate, forget gate and output gate are used for processing and output, and the dropout layer is used to prevent overfitting, and the deep hierarchical sequence H = [h1, h2, h3, ..., h i , ..., h n ];
[0109] The deep sequence containing the time series is input into AM, and the attention mechanism is introduced to weight the deep sequence containing the time series, and different weights are assigned to different features to generate the final power prediction sequence data A = [a1, a2, a3, ..., a i , …, an ], each time step in the power prediction sequence data contains a prediction of future power, where:
[0110]
[0111] Among them, e i is the feature h in the deep sequence containing the time series i The normalized weight of .
[0112] Optionally, the determination module is used to:
[0113] S41, based on the predicted power of the wind turbine in the ramp identification time window of the set width, determining whether a ramp event occurs within the ramp identification time window of the set width;
[0114] S42: If yes, determine whether the climbing event is an uphill climbing event or a downhill climbing event.
[0115] Optionally, the determination module is used to:
[0116] Obtain the predicted power of the wind turbine within the ramp identification time window, and filter the maximum and minimum power values;
[0117] The power difference is calculated by the difference between the maximum power and the minimum power based on the following formula (3):
[0118] ΔP=|P max(t,t+Δt) -P min(t,t+Δt) | (3)
[0119] Where ΔP is the power difference, P max(t,t+Δt) To identify the maximum power within the time window, P min(t,t+Δt) To identify the minimum power value within the time window;
[0120] Determine whether the power difference is greater than a set threshold indicator, and if so, determine that a climbing event occurs;
[0121] Otherwise, it is determined that no hill climbing event has occurred.
[0122] Optionally, the determination module is used to:
[0123] Obtain the power at the end time and the power at the start time of the hill climbing identification time window;
[0124] The ratio of the difference between the power at the end time and the power at the start time to the climbing duration is calculated using the following formula (4):
[0125]
[0126] Among them, Γ is the ratio, Δt is the climbing duration, P tis the power at the beginning of climbing, P t+Δt Power at the end of climbing;
[0127] It is determined whether the ratio is greater than 0, if so, it is determined as an uphill climbing event, otherwise it is determined as a downhill climbing event.
[0128] Optionally, the determination module is used to:
[0129] Obtain the difference between the maximum power and the minimum power of historical climbing events, and calculate the average power difference of historical climbing events;
[0130] Based on the difference between the maximum power and the minimum power of this climbing event and the average power difference of the historical climbing events, the intensity index is calculated, and the calculation formula is:
[0131]
[0132] Where Λ is the intensity index, i = 1, 2, 3, ..., n is the number of historical climbing events, is the maximum power of the i-th historical climbing event, is the minimum power value of the i-th historical climbing event, and e is a natural constant.
[0133] Optionally, the verification module is used to:
[0134] Based on the measured power of wind power, the measured power sequence data is drawn, and several sampling points are set within the time window where the power ramp occurs, and the measured power at each sampling point is sampled;
[0135] Align the power prediction sequence data with the measured power sequence data through timestamps, and sample the predicted power;
[0136] The accuracy index is calculated based on the measured power and predicted power after sampling. The calculation formula of the accuracy index is:
[0137]
[0138] Where Ψ is the accuracy index, j = 1, 2, 3, ..., m is the number of sampling points, is the predicted power of the jth sampling point, For The measured power at the corresponding sampling point, Θ r is the tolerance index of the measured power and predicted power at each sampling point.
[0139] Optionally, the device further comprises a building module;
[0140] The building blocks are used to:
[0141] The historical data of the number of ramp events known to have occurred within a set time period is input into the BA-SDT ramp identification model;
[0142] An evaluation function is set, wherein the function of the evaluation function is: within a period of time, the number of climbing events identified by the SDT algorithm is subtracted from the actual number of climbing events, and the tolerance coefficient is continuously optimized by the BA algorithm and iterated multiple times until the fitness of the bat population is optimal when the difference is 0;
[0143] The tolerance coefficient when the difference is 0 is output, and the construction of the BA-SDT hill climbing recognition model is completed.
[0144] There are a few points to note:
[0145] (1) The drawings of the embodiments of the present invention only relate to the structures related to the embodiments of the present invention, and other structures may refer to the general design.
[0146] (2) For the sake of clarity, in the drawings used to describe the embodiments of the present invention, the thickness of the layers or regions is exaggerated or reduced, that is, these drawings are not drawn according to the actual scale. It is understood that when an element such as a layer, film, region or substrate is referred to as being "on" or "under" another element, the element may be "directly" "on" or "under" the other element or there may be intermediate elements.
[0147] (3) In the absence of conflict, the embodiments of the present invention and the features therein may be combined with each other to obtain new embodiments.
[0148] The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. The protection scope of the present invention shall be based on the protection scope of the claims.
Claims
1. A method for predicting short-term wind power ramping. It is characterized in that The following steps are involved: S1. Obtain sample prediction data and preprocess the sample prediction data; S2. According to the preprocessed sample prediction data, the initial CNN-LSTM-AM power prediction model is trained to obtain a trained CNN-LSTM-AM power prediction model; S3, determining the power prediction sequence data based on the preprocessed sample prediction data and the trained CNN-LSTM-AM power prediction model; S4. Based on the constructed BA-SDT ramp identification model, the power prediction sequence data is identified to obtain the predicted power of the wind turbine, and a ramp event is determined based on the predicted power of the wind turbine to evaluate the intensity index of the ramp event; S5. Obtain the measured power of wind power, determine the measured power sequence data of the wind turbine according to the measured power of wind power, and verify the accuracy of the CNN-LSTM-AM power prediction model based on the measured power sequence data of the wind turbine and the power prediction sequence data to obtain an accuracy index; S6. If the accuracy index is lower than the set accuracy threshold, preprocess the wind power measured power, integrate the preprocessed wind power measured power with the wind turbine prediction data, and input the preprocessed wind power measured power into the CNN-LSTM-AM power prediction model for training; The step S4 of determining a ramp event based on the predicted power of the wind turbine includes: S41, based on the predicted power of the wind turbine in the ramp identification time window of the set width, determining whether a ramp event occurs within the ramp identification time window of the set width; S42: If yes, determining whether the climbing event is an uphill climbing event or a downhill climbing event; Among them, the predicted power of the wind turbine based on the ramp identification time window of the set width in S41 determines whether a ramp event occurs within the ramp identification time window of the set width, including: Obtain the predicted power of the wind turbine within the ramp identification time window, and filter the maximum and minimum power values; The power difference is calculated by the difference between the maximum power and the minimum power based on the following formula (3): ΔP=|P max(t,t+Δt) -P min(t;t+Δt) | (3) Where ΔP is the power difference, P max(t,t+Δt) To identify the maximum power within the time window, P min(t,t+Δt) To identify the minimum power value within the time window; Determine whether the power difference is greater than a set threshold indicator, and if so, determine that a climbing event occurs; If not, it is determined that no climbing event occurred; The intensity index for evaluating the climbing event includes: Obtain the difference between the maximum power and the minimum power of historical climbing events, and calculate the average power difference of historical climbing events; Based on the difference between the maximum power and the minimum power of this climbing event and the average power difference of the historical climbing events, the intensity index is calculated, and the calculation formula is: Where Λ is the intensity index, i = 1, 2, 3, ..., n is the number of historical climbing events, is the maximum power of the i-th historical climbing event, is the minimum power value of the i-th historical climbing event, and e is a natural constant; The method of determining the measured power sequence data of the wind turbine according to the measured power of wind power, and verifying the accuracy of the CNN-LSTM-AM power prediction model based on the measured power sequence data of the wind turbine, and obtaining the accuracy index includes: Based on the measured power of wind power, the measured power sequence data is drawn, and several sampling points are set within the time window where the power ramp occurs, and the measured power at each sampling point is sampled; Align the power prediction sequence data with the measured power sequence data through timestamps, and sample the predicted power; The accuracy index is calculated based on the measured power and predicted power after sampling. The calculation formula of the accuracy index is: Where ψ is the accuracy index, j = 1, 2, 3, ..., m is the number of sampling points, is the predicted power of the jth sampling point, For The measured power at the corresponding sampling point, Θ r is the tolerance index of the measured power and predicted power at each sampling point.
2. The method for predicting short-term wind power ramping according to claim 1, It is characterized in that The forecast data includes historical actual wind power, numerical weather forecast data, theoretical power data of wind towers and wind turbines, and operating capacity data of wind turbines; The step S1 of preprocessing the sample prediction data includes: S11, determining whether there are abnormal values in the wind towers, wind towers of the wind turbines, theoretical power data of the wind turbines, and operating capacity data of the wind turbines of the same type, and if so, deleting the abnormal values and filling in normal values; S12. Align the historical actual wind power, numerical weather forecast data, the theoretical power of the wind tower and the wind turbine, and the operating capacity of the wind turbine according to the set timestamp, and divide the aligned prediction data into a training set and a validation set according to the set ratio.
3. The method for predicting short-term wind power ramping according to claim 2, It is characterized in that The step S11 of judging whether there are abnormal values in the same type of wind towers, wind towers of wind turbines, theoretical power data of wind turbines, and operating capacity data of wind turbines includes: S111. Number the wind towers and wind turbines respectively; S112, obtaining pre-stored power data of the wind tower and the wind turbine under each number and pre-stored startup capacity data of the wind turbine; S113, determining whether the difference between the theoretical power data of the wind tower and the wind turbine under each number and the pre-stored power data of the wind tower and the wind turbine is zero; if not, determining the theoretical power data of the wind tower and the wind turbine corresponding to the number as an abnormal value, deleting the abnormal value, and filling in the corresponding pre-stored power data of the wind tower and the wind turbine; S114. Determine whether the difference between the startup capacity data of the fan under each number and the pre-stored startup capacity data of the fan is zero. If it is not zero, determine the pre-stored startup capacity data of the fan corresponding to the number as an abnormal value, delete the abnormal value, and fill in the corresponding pre-stored startup capacity data of the fan.
4. The method for predicting short-term wind power ramping according to claim 3, It is characterized in that The S3 obtains power prediction sequence data based on the preprocessed sample prediction data and the trained CNN-LSTM-AM power prediction model, including: The preprocessed sample prediction data is input into CNN, and based on the convolution layer and pooling layer of CNN, the local feature sequence Z=[z 1 , z 2 , z 3 , ..., z i , ..., z n ], the convolution operation of the convolution layer is expressed by the following formula (1): y i =f(W*x i +b) (1) Among them, y i is the output of the convolutional layer, f is the activation function, W is the convolution kernel, b is the bias, and x i is the data in the prediction data; The pooling operation of the pooling layer is expressed by the following formula (2): z i =pooling(y i ) (2) Among them, z i is the output of the convolutional layer; The local feature sequence Z containing the time series output by CNN is Z = [z 1 , z 2 , z 3 , ..., z i , ..., z n ] is input to LSTM, hidden units are obtained based on the determined number in the LSTM, long-term dependencies in the local feature sequence containing the time series are captured, and the input gate, forget gate and output gate are used for processing and output, and the dropout layer is used to prevent overfitting, and the deep sequence H=[h 1 ,h 2 ,h 3 , ..., h i , ..., h n ]; The deep sequence containing the time series is input into AM, and the attention mechanism is introduced to weight the deep sequence containing the time series, and different weights are assigned to different features to generate the final power prediction sequence data A = [a 1 , a 2 , a 3 , ..., a i , ..., a n ], each time step in the power prediction sequence data contains a prediction of future power, where: Among them, e i is the feature h in the deep sequence containing the time series i The normalized weight of .
5. A method for predicting short-term wind power ramping according to claim 1, It is characterized in that The step S42 of determining whether the climbing event is an uphill climbing event or a downhill climbing event includes: Obtain the power at the end time and the power at the start time of the hill climbing identification time window; The ratio of the difference between the power at the end time and the power at the start time to the climbing duration is calculated using the following formula (4): Among them, Γ is the ratio, Δt is the climbing duration, P t is the power at the beginning of climbing, P t+Δt Power at the end of climbing; It is determined whether the ratio is greater than 0, if so, it is determined as an uphill climbing event, otherwise it is determined as a downhill climbing event.
6. A method for predicting short-term wind power ramping according to claim 1, It is characterized in that The process of building the BA-SDT hill climbing identification model is as follows: The historical data of the number of ramp events known to have occurred within a set time period is input into the BA-SDT ramp identification model; An evaluation function is set, wherein the function of the evaluation function is: within a period of time, the number of climbing events identified by the SDT algorithm is subtracted from the actual number of climbing events, and the tolerance coefficient is continuously optimized by the BA algorithm and iterated multiple times until the fitness of the bat population is optimal when the difference is 0; The tolerance coefficient when the difference is 0 is output, and the construction of the BA-SDT hill climbing recognition model is completed.
Citation Information
Cited By
Offshore wind power climbing event prediction method based on adaptive graph convolutional network
CN120256817A
