Power station fan state prediction method based on clustering and dynamic division of time series
By combining k-means clustering and LSTM network, the wind turbine status is dynamically divided and weighted, which solves the problem of insufficient prediction accuracy in the existing technology and achieves more efficient wind turbine status prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NORTH CHINA ELECTRIC POWER UNIV
- Filing Date
- 2023-03-27
- Publication Date
- 2026-05-01
AI Technical Summary
In wind turbine condition prediction, conventional classification methods disrupt the temporal continuity of data, preventing LSTM networks from effectively utilizing time-series information and affecting prediction accuracy.
The k-means clustering algorithm is used to dynamically classify the wind turbine status, construct the training and test sets of the LSTM model, and optimize the weight coefficients through a search algorithm to achieve dynamic segmentation of the power plant wind turbine status prediction for time series.
It improves the accuracy and speed of wind turbine condition prediction, adapts to the characteristics of frequent changes in the operating conditions of power plant wind turbines, and achieves more efficient prediction results.
Smart Images
Figure CN116383686B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power plant wind turbine condition prediction technology, specifically a power plant wind turbine condition prediction method based on clustering and dynamic partitioning of time series. Background Technology
[0002] As a crucial auxiliary equipment in coal-fired power plants, wind turbines play a vital role in multiple processes, including flue gas recirculation, equipment sealing, and denitrification and desulfurization. Accurate state prediction of wind turbines is essential for ensuring the safe operation of the power plant and improving economic efficiency. Because wind turbines operate under frequently changing conditions, some monitoring parameters may fluctuate significantly, affecting prediction accuracy. Therefore, for time-varying sampled data such as wind turbine parameters, the sampled value at a given moment is related to the sampled value from previous periods, meaning the data is continuously changing. When using models for prediction, it is necessary to fully extract the temporal information from the input sequence.
[0003] Long Short-Term Memory (LSTM) networks are neural networks capable of accepting inputs from multiple time points. They can take data from time t to time tN as input, and during computation, the information obtained based on the input data at time tN is selectively passed down through the hidden layer neurons of the LSTM network to the computation process at time t-N+1. LSTM networks can be used to fully extract information from the input time series.
[0004] To improve the prediction accuracy of the model, it is necessary to classify the input data, because data of the same type have similar characteristics, which is beneficial for obtaining patterns during modeling. The k-means clustering algorithm is an iterative clustering analysis algorithm. Based on a given number of clusters k, it uses a certain distance function to divide the sample data into k categories, repeating the operation until a certain termination condition is met.
[0005] To improve the prediction accuracy of industrial data such as wind turbines, it is desirable to adopt a classification-first, prediction-later approach. However, conventional classification methods, such as k-means clustering or CNNs, classify data based on magnitude. This classification can disrupt the temporal continuity of the original data, making it unsuitable for prediction networks like LSTM that require continuous input time. Therefore, a method for predicting the state of power plant wind turbines based on clustering and dynamic time series partitioning is urgently needed. This method can improve the prediction accuracy of sub-models through classification, while leveraging networks like LSTM to extract temporal characteristics and obtain optimal values through weighted summaries. This is of great significance for improving the prediction accuracy of wind turbines. Summary of the Invention
[0006] The purpose of this invention is to provide a method for predicting the state of power plant wind turbines based on clustering and dynamic partitioning of time series, characterized by the following steps:
[0007] S1: Collect operating data of the power plant wind turbines over a span of one week, with a sampling period of 1 minute. Preprocess the collected dataset to form a time series composed of data from individual sampling points.
[0008] S2: Select the change in power station turbine current ΔI=I t -I t-1 Together with the power station fan current I, it serves as the benchmark parameter for classifying the power station fan status;
[0009] S3: Use k-means clustering technology to classify the state of power plant wind turbines, and set the number of state classifications k=2;
[0010] S4: Implement classification modeling. Label the original data according to the classification results of S3. Each class of data will correspond to a state prediction model. Select data from each class of samples and construct the training set and test set of each state prediction model by dynamically dividing the time series. Take the power station wind turbine current as the target variable for prediction. Select the input variable for each state prediction model and use correlation analysis or principal component analysis to train and test the corresponding state prediction model.
[0011] S5, during online prediction, simultaneously inputs data into various state prediction models, sets initial values for weight coefficients based on the data classification in the input sequence, and uses the deviation between the final calculated value and the actual observed value as an indicator. Based on the search algorithm, the weight coefficients are optimized to achieve accurate prediction of the power plant wind turbine state.
[0012] The dynamic partitioning of the time series in S4 includes the following steps:
[0013] S41: Dynamically divide the time series in S1 into multiple time group sequences of length N;
[0014] S42: Calculate the label for each time group sequence, wherein the label is in the format of "a / N,b / N";
[0015] Where: a is the number of sampling points in the first category in the time group; b is the number of sampling points in the second category in the time group;
[0016] S43: Based on the labels of the time group sequences in S42, select the sampling points with a larger proportion within the group for classification, and add the time group sequences to the training and test sets of each state prediction model.
[0017] The step in S41, which dynamically divides the time sequence into multiple time groups of length N, is as follows:
[0018] Select the data from the first to the Nth sampling points in the time series as the first time group;
[0019] Select the data from the 2nd to the (N+1)th sampling point in the time series as the second time group;
[0020] This process continues until the last sample point in the time series is selected.
[0021] The input variables selected for each state prediction model in S4 are: power plant turbine current, motor power, front bearing temperature, motor front bearing temperature, bearing horizontal vibration, and bearing vertical vibration.
[0022] The state prediction model in S4 is an LSTM model.
[0023] The steps in S5 for optimizing the weight coefficients based on the search algorithm are as follows:
[0024] S51: Calculate the initial weight w based on the label of each time group sequence obtained in S42. i That is: w1 = a / N, w2 = b / N;
[0025] S52: Set the bias b for the initial weights i Calculate the actual weight x i That is: x1 = w1 + b1, x2 = w2 + b2;
[0026] S53: The parameters to be searched are determined to be: b1, b2;
[0027] S54: Input the training set data into various state prediction models simultaneously to obtain the predicted output values o1, o2;
[0028] S55: Calculate the weighted output value as o = x1o1 + x2o2;
[0029] S56: Calculate the mean square error e between the weighted output value and the true value;
[0030] S57: Take minimizing e as the objective function, b1 and b2 as search variables, set the search range to (0, 0.1), use the genetic algorithm to find the optimal values for b1 and b2, and complete the optimization of the weight coefficients.
[0031] The beneficial effects of this invention are as follows:
[0032] (1) Improve the prediction accuracy of wind turbine status. Existing wind turbine status classification methods only perform static classification based on data and do not fully utilize the information in the input time series. This method classifies the induced draft fan status based on the k-means clustering algorithm, and then constructs the input set and label set of the status prediction model along the time series for training, which is a dynamic classification method. During online prediction, the input sequence is dynamically classified and predicted according to this method, which is suitable for the characteristics of frequent changes in the operating conditions of power plant wind turbines and can effectively improve the prediction accuracy.
[0033] (2) Improve the overall prediction speed. Use a search algorithm to optimize the weight coefficients of the output values of each prediction model. Since the search problem can be simplified to a simple single-objective linear optimization, it can effectively improve the overall prediction speed. Attached Figure Description
[0034] Figure 1 This invention discloses a flowchart of a power plant wind turbine status prediction method based on clustering and dynamic partitioning of time series.
[0035] Figure 2 This is a schematic diagram showing the results of classifying the state of power plant wind turbines using k-means clustering technology in an embodiment of the present invention;
[0036] Figure 3 This is a comparison chart of the prediction effects of four prediction methods in the embodiments of the present invention. Detailed Implementation
[0037] This invention provides a method for predicting the state of wind turbines in power plants based on clustering and dynamic partitioning of time series. The invention will be further described in detail below with reference to the accompanying drawings. Figure 1 The embodiment of the present invention disclosed herein provides a method for predicting the state of wind turbines in power plants based on clustering and dynamic partitioning of time series, specifically including the following steps:
[0038] S1: Collect operating data of the power plant wind turbines over a span of one week, with a sampling period of 1 minute. Preprocess the collected dataset to form a time series composed of data from individual sampling points.
[0039] The collected dataset is preprocessed, with the first 80% of samples used as the overall training sample set and the last 20% used as the overall validation sample set. The preprocessing includes, but is not limited to, data normalization.
[0040] S2: Analyze the wind turbine data. Based on engineering experience, and considering that the power plant wind turbine current I is an important parameter characterizing the operating status of the induced draft fan, select the power plant wind turbine current change ΔI = I. t -I t-1 Together with the power station fan current I, it serves as the benchmark parameter for classifying the power station fan status;
[0041] Choosing the power plant wind turbine current change ΔI and the power plant wind turbine current I is not only simple to calculate, but also reflects the most basic distribution and change characteristics of the data. The power plant wind turbine current data value reflects the size distribution of the data, and the power plant wind turbine current change reflects the change characteristics of the data.
[0042] S3: Use k-means clustering technology to classify the state of power plant wind turbines, and set the number of state classifications k=2;
[0043] The k-means method is used to classify the state of the induced draft fan. In this embodiment, the simplest classification is selected, and the number of classifications is set to 2. The classification results are as follows. Figure 2 As shown, k-means classification performs a simple classification based on the data size. The data distribution range is approximately (28, 40). According to the classification results, the two cluster centers are (32.69, 0.00181) and (29.56, -0.00155).
[0044] In this embodiment, the k-means method is simple in principle and easy to implement. Too many categories do not significantly improve the prediction effect. Therefore, in this embodiment, the data is divided into large data class K1 and small data class K2 based only on the data size. The label vector of K1 class data is set to (1,0) and the label vector of K2 class is set to (0,1). Using a simple clustering classification method can effectively improve the efficiency of model prediction.
[0045] S4: Implement classification modeling. Based on the classification results of S3, label the original data, i.e., large data class K1 or small data class K2. Each data class will correspond to a state prediction model, i.e., D1 and D2. Select data from each class of samples, construct training and test sets for each state prediction model by dynamically dividing the time series. Use the power plant wind turbine current as the target variable for prediction, select input variables for each state prediction model, and use correlation analysis or principal component analysis methods to train and test the corresponding state prediction model.
[0046] In this embodiment, both state prediction models D1 and D2 in S4 are LSTM models.
[0047] Considering the time-dependent characteristics of wind turbine parameters in power plants, it is crucial to fully leverage the temporal information within the input sequence when using models for prediction. LSTM networks, unlike ordinary neural networks which can only accept data from time t to calculate the output, are neural networks capable of accepting inputs from multiple time points. LSTM networks can accept data from time t to time tN. During computation, the information obtained based on the input data at time tN is selectively passed down through the hidden layer neurons of the LSTM network to the computation process at time t-N+1. This means that the computation based on the data at time t-N+1 utilizes not only the data from time t-N+1 but also the information from the previous time point, tN, and so on. Ultimately, the output based on time t can contain data from the previous N time points, thus fully utilizing the LSTM network to extract the temporal information from the input time series.
[0048] Based on the classification of individual sampled data in the time series in S3, in order to adapt to the input requirements of the LSTM network and to realize the input information mining using the LSTM network, it is necessary to further transform the time series composed of individual data into a time group data sequence, that is, to form a time group from time t to time t-N+1, and input the time group into the LSTM for calculation.
[0049] The dynamic partitioning of the time series in S4 includes the following steps:
[0050] S41: Dynamically divide the time series in S1 into multiple time group sequences of length N;
[0051] The step in S41, which dynamically divides the time sequence into multiple time groups of length N, is as follows:
[0052] Select the data from the first to the Nth sampling points in the time series as the first time group;
[0053] Select the data from the 2nd to the (N+1)th sampling point in the time series as the second time group;
[0054] This process continues until the last sample point in the time series is selected.
[0055] In this embodiment, the above grouping method can obtain a large number of time group data sequences with less collected data, that is, it can obtain a richer training set and test set for the state prediction model, which helps to improve the prediction effect of the state prediction model.
[0056] S42: Calculate the label for each time group sequence, wherein the label is in the format of "a / N,b / N";
[0057] Where: a is the number of sampling points in the first category in the time group; b is the number of sampling points in the second category in the time group;
[0058] In this embodiment, the time groups are divided according to the proportion of the two types of data within each time group. Specifically, a time group contains N data points, with 'a' data points of type K1 and 'b' data points of type K2. The label for this time group is then "a / N, b / N". Following this label setting method ensures that the label values are within the range of 0-1, and that the sum of the two sub-values within the label is 1, thus satisfying the normalization requirement.
[0059] S43: Based on the labels of the time group sequences in S42, select the sampling points with a larger proportion within the group for classification, and add the time group sequences to the training and test sets of each state prediction model.
[0060] In this embodiment, to achieve the goal of classification modeling, the time-group data is classified to train different LSTM models. Data of the same type have relatively consistent features, which is beneficial for obtaining patterns during modeling. For example, if a set of data is distributed between 0 and 10, the data distribution is relatively scattered. If it is further subdivided into two groups, 0-5 and 5-10, for feature extraction, the data distribution within each group becomes more concentrated, which can effectively improve the model's prediction accuracy.
[0061] In this embodiment, after obtaining the labels for all time groups in sequence, the sampling points with a larger proportion within each group are selected for classification, and the time group is assigned to the corresponding category. That is, the time groups are divided according to the classification results of the feature vector current. After all the divisions are completed, the data is divided into a training set (accounting for 80% of the total data) and a validation set (20%), and the corresponding data is used to complete the training and validation of two LSTM models (D1, D2).
[0062] The input variables selected for each state prediction model in S4 are: power plant turbine current, motor power, front bearing temperature, motor front bearing temperature, bearing horizontal vibration, and bearing vertical vibration.
[0063] S5, during online prediction, simultaneously inputs data into various state prediction models, sets initial values for weight coefficients based on the data classification in the input sequence, and uses the deviation between the final calculated value and the actual observed value as an indicator. Based on the search algorithm, the weight coefficients are optimized to achieve accurate prediction of the power plant wind turbine state.
[0064] The steps in S5 for optimizing the weight coefficients based on the search algorithm are as follows:
[0065] S51: Calculate the initial weight w based on the label of each time group sequence obtained in S42. i That is: w1 = a / N, w2 = b / N;
[0066] S52: Set the bias b for the initial weights i Calculate the actual weight x i That is: x1 = w1 + b1, x2 = w2 + b2;
[0067] S53: The parameters to be searched are determined to be: b1, b2;
[0068] S54: Input the training set data into various state prediction models simultaneously to obtain the predicted output values o1, o2;
[0069] S55: Calculate the weighted output value as o = x1o1 + x2o2;
[0070] S56: Calculate the mean square error e between the weighted output value and the true value;
[0071] S57: Take minimizing e as the objective function, b1 and b2 as search variables, set the search range to (0, 0.1), use the genetic algorithm to find the optimal values for b1 and b2, and complete the optimization of the weight coefficients.
[0072] In this implementation, the results obtained by optimizing the weight coefficients are: b1 = 0, b2 = 0.014.
[0073] This concludes the disclosure of a power plant wind turbine status prediction method based on clustering and dynamic time series partitioning in this embodiment. By using a dynamic time series partitioning prediction method, classification techniques are employed to improve the prediction accuracy of sub-models. LSTM networks are utilized to mine temporal characteristics, and finally, a weighted average is used to obtain a better value, achieving accurate power plant wind turbine status prediction.
[0074] The effectiveness of this embodiment can be further demonstrated by the following testing methods.
[0075] During operation, the current value is first extracted from the sampled data as a feature variable. The Euclidean distance between (i, Δi) and the two cluster centers is calculated. After normalization, the classification label of the sampled point is obtained, and the data of each sampled point is divided into K1 and K2 classes.
[0076] Following the method of dynamically dividing the time series in S4, we further construct each time group, divide the time group, and obtain the label values of the time group, which are the initial weights w1 and w2. After superimposing the biases b1 and b2, we obtain the actual weights x1 and x2.
[0077] The sampled data is simultaneously input into two trained LSTM models to obtain the corresponding predicted outputs o1 and o2, and the weighted output value o is calculated according to the formula in S55.
[0078] Build and train a standard LSTM prediction model D, and input the same set of parameters into D, D1, and D2 respectively, as follows: Figure 3 As shown in the table, the comparison reveals that the weighted calculated value based on the classification prediction model has a better predictive performance than the single predicted value. The comparison results of the prediction mean square error of the four models are shown in Table 1.
[0079] Table 1: Comparison of Prediction Root Mean Squares of the Four Models
[0080] MSE D 0.4916 D1 0.1839 D2 0.3585 Weighted forecast 0.1770
[0081] In summary, since the initial weights w1 and w2 are dynamically determined based on the range and variation characteristics of the feature variables of each input data set, dynamic partitioning and prediction are achieved. This overcomes the problems of conventional classification methods, which classify data based on magnitude, disrupting the temporal continuity of the original data and preventing the use of prediction networks that require continuous input time. By using a dynamic time series partitioning prediction method, employing classification techniques to improve the prediction accuracy of sub-models, utilizing LSTM networks to mine temporal characteristics, and finally obtaining optimal values through weighted summation, accurate prediction of power plant wind turbine status is achieved.
Claims
1. A method for predicting the state of power plant wind turbines based on clustering and dynamic partitioning of time series, characterized in that, Includes the following steps: S1: Collect operating data of the power plant wind turbines over a span of one week, with a sampling period of 1 minute. Preprocess the collected dataset to form a time series composed of data from individual sampling points. S2: Select the change in power station turbine current and power station wind turbine current Together they serve as the benchmark parameters for classifying the status of power plant wind turbines; S3: Use k-means clustering to classify the states of power plant wind turbines, and set the number of state categories. ; S4: Implement classification modeling. Label the original data according to the classification results of S3. Each class of data will correspond to a state prediction model. Select data from each class of samples and construct the training set and test set of each state prediction model by dynamically dividing the time series. Take the power station wind turbine current as the target variable for prediction. Select the input variable for each state prediction model and use correlation analysis or principal component analysis to train and test the corresponding state prediction model. S5, during online prediction, simultaneously inputs data into various state prediction models, sets initial values for weight coefficients based on the data classification in the input sequence, and uses the deviation between the final calculated value and the actual observed value as an indicator. Based on the search algorithm, the weight coefficients are optimized to achieve accurate prediction of the power plant's wind turbine state. The input variables selected for each state prediction model in S4 are: power plant turbine current, motor power, front bearing temperature, motor front bearing temperature, bearing horizontal vibration, and bearing vertical vibration. The state prediction model in S4 is an LSTM model; The dynamic partitioning of the time series in S4 includes the following steps: S41: Dynamically divide the time series in S1 into multiple time group sequences of length N; S42: Calculate the label for each time group sequence, wherein the label is in the format of "a / N, b / N"; Where: a is the number of sampling points in the first category in the time group; b is the number of sampling points in the second category in the time group; S43: Based on the labels of the time group sequences in S42, select the sampling points with a larger proportion within the group for classification, and add the time group sequences to the training and test sets of each state prediction model; The steps in S5 for optimizing the weight coefficients based on the search algorithm are as follows: S51: Calculate the initial weight w based on the label of each time group sequence obtained in S42. i That is: w1 = a / N, w2 = b / N; S52: Set the bias b for the initial weights i Calculate the actual weight x i That is: x1=w1+b1, x2=w2+b2; S53: The parameters to be searched are determined to be: b1, b2; S54: Input the training set data into various state prediction models simultaneously to obtain the predicted output values o1, o2; S55: Calculate the weighted output value as o = x1o1 + x2o2; S56: Calculate the mean square error e between the weighted output value and the true value; S57: Take minimizing e as the objective function, b1 and b2 as search variables, set the search range to (0, 0.1), use the genetic algorithm to find the optimal values for b1 and b2, and complete the optimization of the weight coefficients.
2. The power plant wind turbine status prediction method based on clustering and dynamic time series partitioning according to claim 1, characterized in that, The step in S41, which dynamically divides the time sequence into multiple time groups of length N, is as follows: Select the data from the first to the Nth sampling points in the time series as the first time group; Select the data from the 2nd to the (N+1)th sampling point in the time series as the second time group; This process continues until the last sample point in the time series is selected.