Multi-source optimization neural network ultra-short-term prediction algorithm based on health assessment

What is AI technical title?
AI technical title is built by Patsnap AI team. It summarizes the technical point description of the patent document.
Through a multi-source optimization neural network ultra-short-term prediction algorithm based on health assessment, the problems of low prediction accuracy and slow speed of the existing wind turbine power prediction algorithm model are solved, and higher prediction accuracy and speed are achieved, which improves wind power equipment regulating accuracy, grid stability and power generation efficiency.

CN119476633APending Publication Date: 2025-02-18HUNAN INSTITUTE OF ENGINEERING

View PDF 0 Cites 0 Cited by

Patent Information

Application Number: CN202411701953.0
Authority / Receiving Office: CN · China
Patent Type: Applications(China)
Current Assignee / Owner
Filing Date: 2024-11-26
Publication Date: 2025-02-18

AI Technical Summary

Technical Problem

The prediction accuracy and slow speed of the existing wind turbine power prediction algorithm model lead to poor regulation accuracy of wind power equipment, resulting in reduced grid stability and low power generation efficiency.

Method used

Design a multi-source optimized neural network ultra-short-term prediction algorithm based on health assessment, including operation and maintenance data processing module, construction of health matrix module and optimization of BP neural network module, and optimize prediction power through feature screening, health assessment and neural network training.

Benefits of technology

It improves the accuracy and speed of wind turbine power prediction, enhances the adjustment accuracy of wind power equipment, and improves the stability and power generation efficiency of the power grid.

✦ Generated by Eureka AI based on patent content.

Smart Images

Figure CN119476633A_ABST

Patent Text Reader

Abstract

The invention relates to the technical field of health assessment prediction algorithms, in particular to a multi-source optimization neural network ultra-short-term prediction algorithm based on health assessment, which comprises an operation and maintenance data processing module, a health matrix construction module and a BP neural network optimization module. The output end of the operation and maintenance data processing module is provided with a health matrix construction module; the output end of the health matrix building module is provided with an optimized BP neural network module, and the operation and maintenance data such as the humidity, the yaw angle, the blade deflection angle, the wind speed and the temperature of the target wind driven generator are obtained by arranging the original data; the method comprises the steps of preprocessing original data, removing null values and abnormal values, and performing normalization processing to enable the original data to meet an input format of a neural network;

Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of health assessment prediction algorithms, and in particular relates to a multi-source optimized neural network ultra-short-term prediction algorithm based on health assessment. Background Art

[0002] As a renewable energy source, wind energy has the volatility and intermittency to bring challenges to the stable operation and power generation efficiency of the power grid. In order to meet these challenges, wind power prediction technology has emerged. This technology collects historical power, wind speed, topography, numerical weather forecast and other data of wind farms to establish a prediction model for wind farm output power. These models can be based on wind speed, power or numerical weather forecast data as input, and combined with the equipment status and operating conditions of the wind farm units to obtain the future output power of the wind farm.

[0003] However, the existing wind turbine power prediction algorithm model has low prediction accuracy, which is mainly due to the low accuracy of wind speed prediction and the limitations of the model in dealing with complex meteorological conditions and topography. On the other hand, the prediction speed is slow, which limits the application of the model in real-time scheduling and optimization of resource allocation.

[0004] Therefore, in order to address the problem that the above-mentioned existing wind turbine power prediction algorithm model has low prediction accuracy and slow speed, resulting in poor adjustment accuracy of wind power equipment and reduced grid stability and low power generation efficiency, a low-noise robot reducer sealed housing can be designed. Summary of the invention

[0005] In order to overcome the problem that the existing wind turbine power prediction algorithm model has low prediction accuracy and slow speed, resulting in poor adjustment accuracy of wind power equipment and reduced grid stability and low power generation efficiency.

[0006] The technical solution of the present invention is: a multi-source optimized neural network ultra-short-term prediction algorithm based on health assessment, including an operation and maintenance data processing module, a health matrix construction module, and an optimization BP neural network module; the output end of the operation and maintenance data processing module is provided with a health matrix construction module; the output end of the health matrix construction module is provided with an optimization BP neural network module; the operation and maintenance data processing module includes operation and maintenance data and historical data; the health matrix construction module includes matrix elements, optimization parameters, and matrix dimensions; the optimization BP neural network module includes a health matrix and approximate hyperparameters.

[0007] Preferably, the operation and maintenance data of the target wind turbine, such as humidity, yaw angle, blade deflection angle, wind speed and temperature, are obtained by collating the original data; the original data is preprocessed to remove null values and abnormal values, and normalized to meet the input format of the neural network; the wind turbine operation and maintenance data is feature screened using the Lasso algorithm and the Pearson correlation coefficient method to eliminate irrelevant data; based on the screened data, the FNN model is trained to obtain the functional relationship between the wind turbine power generation and the operation and maintenance characteristics; based on the trained FNN model, the correlation coefficient k of the wind power is obtained using the chain derivative rule; the correlation coefficient k and the historical power value are used together for health assessment; a health matrix is formed, the power weight is changed, the optimal hyperparameter is found, and the BP neural network is trained; the predicted power is optimized, and the test results are output for comparative analysis, and conclusions are drawn, which solves the problem that the existing wind turbine power prediction algorithm model has low prediction accuracy and slow speed, resulting in poor adjustment accuracy of wind power equipment and decreased grid stability and low power generation efficiency.

[0008] Preferably, the output end of the operation and maintenance data is provided with feature screening; the output end of the feature screening is provided with obtaining a fitting function; the output end of obtaining the fitting function is provided with a correlation coefficient K; the output end of the correlation coefficient K is provided with feature data; and the input end of the feature data is provided with historical data.

[0009] Preferably, the output end of the feature data is provided with matrix elements, optimization parameters, and matrix dimensions.

[0010] Preferably, the output end of the matrix element is provided with efficiency calculation; the output end of the efficiency calculation is provided with state distribution; and the output end of the optimization parameter is provided with parameter calculation.

[0011] Preferably, a health matrix is provided at a common output terminal of the state distribution and the matrix dimension; and an approximate hyperparameter is provided at an output terminal of the parameter calculation.

[0012] Preferably, the output end of the health matrix is provided with an improved power weight; and the output end of the approximate hyperparameter is provided with a BP neural network.

[0013] Preferably, the common output terminal of the improved power weight and the BP neural network is provided with optimized predicted power.

[0014] Beneficial effects of the present invention:

[0015] 1. Arrange the original data to obtain the target wind turbine's humidity, yaw angle, blade deflection angle, wind speed, temperature and other operation and maintenance data; pre-process the original data, remove null values, outliers, and normalize them to meet the input format of the neural network; use the Lasso algorithm and Pearson correlation coefficient method to perform feature screening on the wind turbine operation and maintenance data and eliminate irrelevant data; train the FNN model based on the screened data to obtain the functional relationship between the wind turbine power generation and operation and maintenance characteristics; based on the trained FNN model, use the chain derivative rule to obtain the correlation coefficient k of wind power; use the correlation coefficient k and historical power values to perform health assessment; form a health matrix, change the power weight, find the optimal hyperparameter, and train the BP neural network; optimize the predicted power, output the test results for comparative analysis, and draw conclusions;

[0016] 2. Feature screening is achieved through Lasso and Pearson coefficient methods, and FNN uses BP algorithm to adjust the weight of the model. According to the universal approximation theorem, FNN can fit any continuous nonlinear function, including the high-dimensional nonlinear function in this algorithm; the data objects are wind speed S, blade deflection angle θ and power P, which cannot change suddenly and are all continuous, and continuity must be differentiable, so the partial derivative of power P to wind speed S can be obtained through FNN; health assessment is a weight optimization algorithm. It converts the two-dimensional probability density function into a weighted health matrix based on the characteristics of the statistical probability distribution of the model output; data collection: through sensors and SCADA systems, it collects wind speed, temperature, pressure, speed, power and other fan operation data; data cleaning: cleaning the missing values, outliers, and transmission noise in the operation data to improve data reliability; feature screening: from the cleaned operation data, the feature data is screened out, and based on the dimension of the feature data, the dimension of the health matrix is determined; state division: using feature data, the feature weight is calculated, and the fan state is divided into normal, minor fault, serious fault and other states according to the weight value. Finally, the above data is used to derive a more calibrated conclusion, and a health matrix is obtained through health assessment. After calculating the improved power weight, the predicted power is output using the neural network, which reduces the prediction error and improves the network prediction accuracy and speed without increasing the data processing cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 What is shown is a schematic diagram of the algorithm flow of the multi-source optimized neural network ultra-short-term prediction algorithm based on health assessment of the present invention;

[0018] Figure 2 What is shown is the BP prediction effect diagram of the multi-source optimized neural network ultra-short-term prediction algorithm based on health assessment of the present invention without adding k value optimization;

[0019] Figure 3What is shown is the BP prediction effect diagram of adding k value optimization to the multi-source optimized neural network ultra-short-term prediction algorithm based on health assessment of the present invention;

[0020] Figure 4 What is shown is a schematic diagram comparing the results of different optimization degrees of the multi-source optimized neural network ultra-short-term prediction algorithm based on health assessment of the present invention;

[0021] Figure 5 What is shown is a schematic diagram of a neural network prediction error comparison embodiment of the multi-source optimized neural network ultra-short-term prediction algorithm based on health assessment of the present invention.

[0022] The markings in the attached figure are: 1. Operation and maintenance data processing module; 2. Health matrix construction module; 3. BP neural network optimization module; 101. Operation and maintenance data; 102. Feature screening; 103. Training FHN; 104. Obtaining fitting function; 105. Correlation coefficient K; 106. Historical data; 107. Feature data; 201. Matrix elements; 202. Optimization parameters; 203. Efficiency calculation; 204. Parameter calculation; 205. State distribution; 206. Matrix dimension; 301. Health matrix; 302. Approximate hyperparameters; 303. Improved power weight; 304. BP neural network; 305. Optimized predicted power. DETAILED DESCRIPTION

[0023] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0024] See also Figure 1-5 The present invention provides an embodiment: a multi-source optimized neural network ultra-short-term prediction algorithm based on health assessment includes an operation and maintenance data processing module 1, a health matrix construction module 2, and an optimization BP neural network module 3; the output end of the operation and maintenance data processing module 1 is provided with a health matrix construction module 2; the output end of the health matrix construction module 2 is provided with an optimization BP neural network module 3; the operation and maintenance data processing module 1 includes operation and maintenance data 101 and historical data 106; the health matrix construction module 2 includes matrix elements 201, optimization parameters 202, and matrix dimensions 206; the optimization BP neural network module 3 includes a health matrix 301 and an approximate hyperparameter 302.

[0025] See also Figure 1-5 In this embodiment, the output end of the operation and maintenance data 101 is provided with feature screening 102; the output end of the feature screening 102 is provided with obtaining a fitting function 104; the output end of obtaining the fitting function 104 is provided with a correlation coefficient K105; the output end of the correlation coefficient K105 is provided with feature data 107; the input end of the feature data 107 is provided with historical data 106.

[0026] See also Figure 1-5In this embodiment, the output end of the feature data 107 is provided with a matrix element 201, an optimization parameter 202, and a matrix dimension 206; the output end of the matrix element 201 is provided with an efficiency calculation 203; the output end of the efficiency calculation 203 is provided with a state distribution 205; the output end of the optimization parameter 202 is provided with a parameter calculation 204; the common output end of the state distribution 205 and the matrix dimension 206 is provided with a health matrix 301; the output end of the parameter calculation 204 is provided with an approximate hyperparameter 302; the output end of the health matrix 301 is provided with an improved power weight 303; the output end of the approximate hyperparameter 302 is provided with a BP neural network 304; the common output end of the improved power weight 303 and the BP neural network 304 is provided with an optimized predicted power 305.

[0027] When working, feature screening is achieved through Lasso and Pearson coefficient methods, and FNN uses BP algorithm to adjust the weight of the model. According to the general approximation theorem, FNN can fit any continuous nonlinear function, including the high-dimensional nonlinear function in this algorithm; the data objects are wind speed S, blade deflection angle θ and power P, which cannot change suddenly and are all continuous, and continuity must be differentiable, and the partial derivative of power P with respect to wind speed S can be obtained through FNN; health assessment is a weight optimization algorithm, which converts the two-dimensional probability density function into a weighted health matrix based on the characteristics of the statistical probability distribution of the model output; data collection: collect wind speed, temperature, pressure, speed, power and other fan operation data through sensors and SCADA system; data cleaning: for the operation data Missing values, outliers, and transmission noise are cleaned to improve data reliability; Feature screening: Feature data is screened from the cleaned operating data, and the dimension of the health matrix is determined based on the dimension of the feature data; State division: Feature data is used to calculate feature weights, and the wind turbine states such as normal, minor faults, and serious faults are divided according to the weight values. Finally, the above data is used to draw more calibrated conclusions, and the health matrix is obtained through health assessment. After calculating the improved power weights, the predicted power is output using the neural network, which reduces the prediction error without increasing the data processing cost, and improves the network prediction accuracy and speed;

[0028] The original data is sorted to obtain the operation and maintenance data of the target wind turbine, such as humidity, yaw angle, blade deflection angle, wind speed and temperature; the original data is preprocessed to remove null values and outliers, and normalized to meet the input format of the neural network; the wind turbine operation and maintenance data is feature screened using the Lasso algorithm and the Pearson correlation coefficient method to eliminate irrelevant data; based on the screened data, the FNN model is trained to obtain the functional relationship between the wind turbine power generation and the operation and maintenance characteristics; based on the trained FNN model, the chain derivative rule is used to obtain the correlation coefficient k of wind power; the correlation coefficient k and the historical power value are used together for health assessment; a health matrix is formed, the power weight is changed, the optimal hyperparameter is found, and the BP neural network is trained; the predicted power is optimized, and the test results are output for comparative analysis to draw conclusions.

[0029] Through the above steps, the multi-source optimization algorithm significantly reduces the amount of data processing, reduces the lag of prediction, and improves the quality of the BP network power weight. After optimization, the prediction accuracy of the BP network is significantly improved. At the same time, the matrix approximation regularization parameter is used as the starting value for optimization, which not only improves the optimization efficiency, but also effectively reduces the BP prediction time. According to the relationship between multiple variables and power during the actual operation of the wind turbine, the data characteristics and correlation are considered, the intermediate variable k is added, the health matrix is calculated, and the BP neural network ultra-short-term power prediction algorithm is optimized. Tests show that compared with existing algorithms, this algorithm has shorter prediction time, lower data cost, and better prediction effect.

[0030] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge scope of those skilled in the art without departing from the spirit of the present invention.

Claims

1. A multi-source optimized neural network ultra-short-term prediction algorithm based on health assessment, including an operation and maintenance data processing module (1), characterized in that: It also includes a health matrix construction module (2) and an optimization BP neural network module (3); the output end of the operation and maintenance data processing module (1) is provided with a health matrix construction module (2); the output end of the health matrix construction module (2) is provided with an optimization BP neural network module (3); the operation and maintenance data processing module (1) includes operation and maintenance data (101) and historical data (106); the health matrix construction module (2) includes matrix elements (201), optimization parameters (202) and matrix dimensions (206); The optimized BP neural network module (3) includes a health matrix (301) and an approximate hyperparameter (302).

2. The multi-source optimized neural network ultra-short-term prediction algorithm based on health assessment according to claim 1 is characterized by: The output end of the operation and maintenance data (101) is provided with a feature screening (102); the output end of the feature screening (102) is provided with an acquisition fitting function (104); the output end of the acquisition fitting function (104) is provided with a correlation coefficient K (105); the output end of the correlation coefficient K (105) is provided with feature data (107); the input end of the feature data (107) is provided with historical data (106).

3. The multi-source optimized neural network ultra-short-term prediction algorithm based on health assessment according to claim 2 is characterized by: The output end of the feature data (107) is provided with matrix elements (201), optimization parameters (202), and matrix dimensions (206).

4. The multi-source optimized neural network ultra-short-term prediction algorithm based on health assessment according to claim 3 is characterized by: The output end of the matrix element (201) is provided with an efficiency calculation (203); the output end of the efficiency calculation (203) is provided with a state distribution (205); and the output end of the optimization parameter (202) is provided with a parameter calculation (204).

5. The multi-source optimized neural network ultra-short-term prediction algorithm based on health assessment according to claim 4 is characterized by: The common output terminal of the state distribution (205) and the matrix dimension (206) is provided with a health matrix (301); the output terminal of the parameter calculation (204) is provided with an approximate hyperparameter (302).

6. The multi-source optimized neural network ultra-short-term prediction algorithm based on health assessment according to claim 1 is characterized by: The output end of the health matrix (301) is provided with an improved power weight (303); the output end of the approximate hyperparameter (302) is provided with a BP neural network (304).

7. The multi-source optimized neural network ultra-short-term prediction algorithm based on health assessment according to claim 6 is characterized by: The common output terminal of the improved power weight (303) and the BP neural network (304) is provided with an optimized predicted power (305).