Intelligent prediction method of wind turbine power based on deep learning

Abstract

The present invention relates to the technical field of power prediction, and proposes a method for intelligent prediction of wind turbine power based on deep learning, including data input, data fusion, data normalization processing, calculation of correlation coefficients and weights, screening out samples with strong correlation as training samples of a power prediction model, and training to obtain a power prediction model with relatively accurate calculations. The power prediction model can accurately and effectively output predicted power values. Through the prediction results of the wind power prediction system, the power grid dispatching department can reasonably arrange the power generation plan, reduce the system's rotating reserve capacity, and improve the economy of the power grid operation. At the same time, by predicting the fluctuation of wind power in advance, reasonably arranging the operation mode and response measures, the safety and reliability of the power grid can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power prediction, and in particular to a method for intelligently predicting wind turbine power based on deep learning. Background Art

[0002] With the rapid development of new energy in China, a large number of wind farms and power stations have been gradually built. With the gradual deepening of daily management work, how wind farms and power stations can "reduce power loss and increase power generation" and thereby carry out the "double improvement" work has become an important management goal of new energy companies.

[0003] Wind power prediction is an important technical means to ensure safe, stable and efficient wind power generation. By collecting numerical weather forecast data, real-time weather station data, real-time output power data, inverter unit status and other data, the wind farm power prediction is completed and can be uploaded to the dispatching side power prediction system according to the requirements of the power grid. Through the prediction results of the wind power prediction system, the power grid dispatching department can reasonably arrange the power generation plan, reduce the system's rotating reserve capacity, and improve the economy of power grid operation. At the same time, by predicting the fluctuation of wind power in advance, reasonably arrange the operation mode and response measures, and improve the safety and reliability of the power grid.

[0004] Power forecasting has important guiding significance for wind power generation, but the main method at present is to fit the wind turbine power curve by collecting static data. The accuracy cannot be guaranteed, and it is impossible to achieve the effect of "from passive repair to active operation and maintenance". It is impossible to timely grasp and judge the current status of the wind turbine, and it is impossible to effectively evaluate the wind turbine performance and provide effective support for fault prediction. Summary of the invention

[0005] The present invention proposes an intelligent prediction method for wind turbine power based on deep learning, uses mature artificial intelligence technology, conducts value mining and intelligent analysis based on a large amount of historically accumulated data, and proposes a wind turbine prediction model, which solves the problem in the prior art that the accuracy of wind turbine power calculation cannot be guaranteed and the wind turbine power cannot be well predicted.

[0006] The technical solution of the present invention is as follows:

[0007] A wind turbine power intelligent prediction method based on deep learning, including the training steps of the power prediction model,

[0008] Step 100: input multiple groups of feature samples and target samples at the same time, the feature samples include power data of wind turbines, and the target samples include meteorological data of wind turbines and wind measurement data of the station;

[0009] Step 200: associate and fuse the feature samples and the target samples with the time period as the association condition, and output a wide data table;

[0010] Step 300: Input the data wide table into the normalization model and output the normalized data wide table. Each sample value in the normalized data wide table Among them, x max is the maximum sample value at all times, x min is the minimum sample value of all moments, and x is the original sample value;

[0011] Step 400: Calculate the correlation coefficient and weight of the target sample and the feature sample according to the normalized data wide table, and remove the target sample whose correlation coefficient is less than the set value of the correlation coefficient and the target sample whose weight is less than the set value of the weight from the data wide table;

[0012] Step 500: construct a wind turbine power prediction model based on a bidirectional LSTM network, input the remaining target samples and feature samples in the data wide table into the wind turbine power prediction model for training, input the feature samples at time t-1, and output the target samples corresponding to time t;

[0013] Step 600: Calculate the root mean square error of the wind turbine power prediction model and the coefficient of determination When the root mean square error is less than the root mean square error setting value and the coefficient of determination value is greater than the coefficient of determination setting value, the training is completed and the power prediction model is obtained.

[0014] The working principle and beneficial effects of the present invention are:

[0015] The present invention converts raw data into model training data through data processing, and obtains better training data features through feature selection, so that the machine learning model can approach the upper limit of the optimal effect. The power prediction model trained based on the selected sample features can accurately and effectively output the predicted power value. Through the prediction results of the wind power prediction system, the power grid dispatching department can reasonably arrange the power generation plan, reduce the system's rotating reserve capacity, and improve the economy of power grid operation. At the same time, by predicting the fluctuation of wind power in advance, the operation mode and response measures can be reasonably arranged to improve the safety and reliability of the power grid.

[0016] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic diagram of the process of the present invention;

[0018] Figure 2 This is the application effect diagram of the traditional LSTM model in the present invention;

[0019] Figure 3 This is a diagram showing the application effect of the bidirectional LSTM model in the present invention. DETAILED DESCRIPTION

[0020] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments 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.

[0021] like Figure 1 As shown, the present invention is a wind turbine power intelligent prediction method based on deep learning, including

[0022] 1. Data access

[0023] Data access mainly imports various types of data from external data sources into the data analysis domain storage system for calculation and analysis. Real-time data collection access uses message queue components + stream computing components; external data collection access uses file import data platform, external interface collection and other methods. The main data accessed include wind turbine archive data, operation data, station wind measurement data, external meteorological data, power data, etc. Because the data that affect power prediction are mainly station wind measurement data, meteorological data and power data, the power data of wind turbines is used as feature samples, and the meteorological data of wind turbines and station wind measurement data are used as target samples.

[0024] 2. Data preprocessing

[0025] Data preprocessing plays a decisive role in modeling and prediction. The quality of data directly determines the prediction and generalization capabilities of the model. It involves many factors, including accuracy, completeness, consistency, timeliness, credibility and interpretability. The present invention cleans the acquired data, integrates it into a data wide table, and then reduces the data wide table to only leave wind measurement data, meteorological data and power data, and finally normalizes the data in the data wide table.

[0026] 2.1 Data Cleaning

[0027] Data cleaning mainly "cleans" data by filling missing values, smoothing noisy data, smoothing or deleting outliers, and resolving data inconsistencies. If data cleaning is not up to standard, the mining results based on these data will be inaccurate, that is, the output results are unreliable. Outlier processing statistics detect numerical attributes, calculate the mean and standard deviation of field values, and consider the confidence interval of each field to identify abnormal fields and records. Data mining methods are introduced into data cleaning, such as clustering methods for detecting abnormal records, model methods for discovering abnormal records that do not conform to existing patterns, and association rule methods for discovering abnormal data in the data set that do not conform to rules with high confidence and support.

[0028] 2.2 Data Integration

[0029] In the process of data processing, different data exist in different business systems, involving the association analysis of multiple tables. Data integration combines data from multiple data sources and stores them in a consistent data warehouse. Inconsistent naming of attributes or dimensions during the data integration process will lead to redundancy in the data set. Use correlation to detect redundancy: numerical variables can calculate the correlation coefficient matrix, nominal variables can calculate the chi-square test, and different data sources should be standardized and deduplicated when they are merged. Data integration uses the pattern integration method, which mainly uses a custom global pattern to provide a virtual view of all data sources, and the data is still retained at the data source. The pattern integration method has become a common method for data integration due to its flexibility and typicality.

[0030] 2.3 Data Specification

[0031] Data reduction technology is used to obtain a reduced representation of a data set, which can scale the data set while still closely maintaining the integrity of the original data. In this way, mining on the reduced data set will be more efficient and produce the same analysis results. Generally, there are the following strategies: Dimensionality reduction. The data used for data analysis will contain hundreds of attributes, most of which are irrelevant to the mining task and are redundant. Dimensionality reduction reduces the amount of data by deleting irrelevant attributes and ensures minimal information loss. Dimension transformation. Dimension transformation reduces existing data to a smaller dimension, tries to ensure the integrity of data information, and will greatly improve the efficiency of modeling.

[0032] 2.4 Data Transformation

[0033] Data transformation includes normalizing, discretizing, and sparse processing of data to achieve the purpose of mining. Generally, the following strategies are included: Normalization: The dimensions of different features in the data may be inconsistent, and the difference between the values ​​may be large. Failure to process will affect the results of data analysis. Therefore, the data needs to be scaled according to a certain proportion so that it falls in a specific area for comprehensive analysis. In particular, distance-based mining methods, clustering, KNN, and SVM must be normalized. Maximum-minimum normalization maps data to the [0,1] interval to ensure the uniformity of dimensions. The calculation formula is as follows: Among them, x max The maximum sample value at all times, x min is the minimum sample value of all moments, and x is the original sample value.

[0034] 3. Feature Selection

[0035] Traditional feature selection only relies on correlation coefficients to carry out feature engineering, but the effect is not ideal. Therefore, the correlation coefficient is combined with the ensemble learning algorithm. First, the GBDT algorithm is used to calculate the weight of the feature factor, and the features with a larger weight are preliminarily screened. At the same time, the Pearson correlation coefficient is used to analyze the correlation between the input feature and the target field.

[0036] 3.1 Calculation of correlation coefficient

[0037] The Pearson correlation coefficient is also called the Pearson product-moment correlation coefficient, which is a linear correlation coefficient. The sample correlation coefficient is obtained by estimating the standard scores of the X and Y sample points. Where n is the number of samples, X i is the feature sample value, is the feature sample mean, Y i is the target sample value, is the target sample mean, δ X is the standard deviation of the characteristic sample value, δ Y is the standard deviation of the target sample value;

[0038] The Pearson correlation coefficient reflects the strength of the linear correlation between two variables. The larger the absolute value of r, the stronger the correlation.

[0039] When r>0, it indicates that the two variables are positively correlated, that is, the larger the value of one variable, the larger the value of the other variable.

[0040] When r<0, it indicates that the two variables are negatively correlated, that is, the larger the value of one variable, the smaller the value of the other variable.

[0041] When r = 0, it indicates that the two variables are not linearly related, but may be related in other ways, such as a curve.

[0042] When r = 1 and -1, it indicates that the two variables X and Y can be described by a straight line equation, that is, the sample points all fall well on a straight line.

[0043] 3.2 Calculating weights

[0044] Using the negative gradient fitting idea of ​​GBDT, the original data distribution can be further optimized, the independence between feature vectors can be ensured, overfitting of calculation results can be avoided, and the calculation difficulties caused by "sparse" and "dense" two-dimensional array matrices can be avoided, thus improving the operation efficiency. In the process of GBDT outputting weights, the Pearson correlation coefficient values ​​of the data wide table and the neighborhood are calculated iteratively multiple times, and the second-order Taylor expansion is used for mapping.

[0045] The multi-feature GBDT corresponds to multivariate logistic regression and binary logistic regression. Assuming the number of categories is K, the corresponding log-likelihood loss function is: The probability expression of the kth class is: The information entropy is calculated as: Where p(x) is the probability value of the category to which sample x belongs. After obtaining the information entropy, the weight calculation idea of ​​multiple base classifiers of GBDT is used to incorporate the result into the weight calculation formula Where E i is the information entropy of each feature.

[0046] 4. Model construction

[0047] A wind turbine power prediction model is built based on a bidirectional LSTM network. The remaining target samples and feature samples in the data table are input into the wind turbine power prediction model for training. The feature samples at time t-1 are input and the target samples corresponding to time t are output. The root mean square error of the wind turbine power prediction model is calculated. and the coefficient of determination When the root mean square error is less than the set value and the coefficient of determination is greater than the set value, the training is completed and the power prediction model is obtained. The meteorological data of the current wind turbine and the wind measurement data of the station are input into the power prediction model, and the predicted power data of the wind turbine is output.

[0048] The time series data was generated using the selected features, and the long short-term memory neural network model (LSTM) was constructed using the complete data from 2019 to 2020. The model input was the wind speed, wind direction, power, etc. at time t-1, and the output was the power data corresponding to time t. The training was iterated 200 times, and 128 sample data were input for each training to obtain the LSTM-based intelligent power prediction model. During the model training process, the loss function continued to decrease with the number of training times, and the final loss was 0.0471. The root mean square error of the model was 0.03875, and the coefficient of determination was 0.8879. From the perspective of technical indicators, the model has a good effect and can accurately learn and memorize the historical power change law.

[0049] Taking into account the under-fitting ability of traditional LSTM for complex data distribution and business application scenarios, the performance of LSTM is improved and two-way feedback processing is performed to enable the data source to flow in two-way feedback within the LSTM, enhance the model's memory and learning capabilities, and further optimize the weight coefficients of the input gate, forget gate, and output gate in the LSTM, thereby greatly improving the performance of LSTM.

[0050] Among them, the Forward layer and the Backward layer are linked to the output layer, which contains 6 shared weights w1-w6. The Forward layer performs a forward calculation from time 1 to time t, and obtains and saves the output of the forward hidden layer at each time. The Backward layer performs a reverse calculation from time t to time 1, and obtains and saves the output of the backward hidden layer at each time. Finally, at each time, the final output is obtained by combining the output results of the Forward layer and the Backward layer at the corresponding time. The mathematical expression is as follows: h t =f(w1x t +w2h t-1 );h t ′=f(w3x t +w5h t ' -1 );o t =g(w4x t +w6h t ′)

[0051] In the calculation process, the particularity of power for power fluctuation is taken into account. Only the LSTM algorithm cannot organically integrate business and technology. Therefore, the business characteristics of power itself are taken into account, and the power difference is divided into the following formula: L(t) = f(t) + p(t) + x(t)

[0052] Where L(t) is the real-time power, f(t) is the trend term, which reflects the changing trend of L(t), p(t) is the periodic term, which reflects the periodic change of L(t), and x(t) is the random term, which reflects the stability of L(t).

[0053] Finally, the split power features are incorporated into the bidirectional LSTM for iterative calculation, and the impact of business factors on the model is taken into account, which greatly improves the performance of LSTM and increases the determination coefficient from 0.673 to 0.8879.

[0054] The application effect of the traditional LSTM model is as follows Figure 2 As shown, the application effect of the bidirectional LSTM model is as follows Figure 3 As shown in the figure, the bidirectional LSTM model can accurately fit the future power data, and the overall trend and power value are consistent with the real power change trend, and has good generalization ability.

[0055] 5. Model Adjustment

[0056] The predicted power data output by the power prediction model is verified using the theoretical power generation balance analysis method, and the model is optimized and adjusted based on the verification results.

[0057] Theoretical power generation is essentially the sum of the actual power generation of the unit and five power losses. SCADA collects wind speed and power data for power curve fitting, and calculates each power loss through the fitting curve. The theoretical power generation is obtained by adding the actual power generation. The five power losses are as follows:

[0058] Planned maintenance loss: refers to the amount of electricity lost due to shutdown of power generation equipment during the statistical period due to planned inspections, regular maintenance, seasonal inspections, condition maintenance and testing of equipment. It includes two parts: planned maintenance loss of power generation equipment and planned maintenance loss of power transmission and transformation equipment.

[0059] Electricity loss due to unplanned maintenance: refers to the electricity loss due to wind turbine shutdown caused by failure of wind turbines, box transformers, collection lines, and equipment in the substation in the wind farm during the statistical period, including two parts: electricity loss due to wind turbine failure and electricity loss due to failure of substation equipment.

[0060] Electricity loss due to power restriction: including power outage and power reduction. The electricity loss during the power restriction period is called electricity loss due to power restriction.

[0061] Lost electricity: refers to the loss of electricity caused by power generation equipment shutdown due to planned maintenance, unplanned maintenance and emergency repair (unplanned shutdown), failure or force majeure by the power grid company during the statistical period. It includes three parts: power loss due to planned maintenance of the power grid, power loss due to power grid failure, and power loss due to force majeure.

[0062] Performance loss electricity: includes two types: self-derating of the unit and unit performance not meeting the standard. Self-derating of the unit refers to the unit operating at reduced output due to the unit's own reasons, such as protection operation after the replacement of major components (bearings). Unit performance not meeting the standard means that the operating output does not reach the standard value or the unit cannot be cut in after the wind speed reaches the rated cut-in wind speed. The electricity loss during this period is called performance loss electricity.

[0063] Example 1

[0064] Taking a certain wind farm as an example, the power data, wind measurement data, meteorological data, etc. in the corresponding time period are screened.

[0065] The power data mainly includes the measured power value, one point every 15 minutes, a total of 70176 data. The sample data is shown in Table 1 below:

[0066] Table 1

[0067] Serial number time Actual power (MW) 1 2019-01-01 00:00:00 0.254383 2 2019-01-01 00:15:00 0.329703 3 2019-01-01 00:30:00 0.296306 4 2019-01-01 00:45:00 0.18759 5 2019-01-01 01:00:00 0.081005

[0068] The wind measurement data and meteorological data mainly include wind speed at a height of 10 meters, wind direction at a height of 10 meters, wind speed at a height of 30 meters, wind direction at a height of 30 meters, wind speed at a height of 50 meters, wind direction at a height of 50 meters, wind speed at a height of 70 meters, wind direction at a height of 70 meters, wind speed at the height of the wind turbine hub, wind direction at the height of the wind turbine hub, temperature, air pressure, relative humidity and other factors. The corresponding data are recorded every 15 minutes, with a total of 72,001 data. The sample data is shown in Table 2 below:

[0069] Table 2

[0070]

[0071]

[0072] The measured power value, wind measurement data and meteorological data are integrated into a wide data table, and the data in the wide data table are normalized. The normalized sample data is shown in Table 3:

[0073] Table 3

[0074]

[0075]

[0076] Using the Pearson correlation coefficient, we analyze the correlation between the input features and the target field, and the correlation relationship is shown in Table 4 below:

[0077] Table 4

[0078]

[0079]

[0080] From the correlation relationship, we can see that the correlations of temperature, air pressure, and relative humidity are 0.18, 0.015, and 0.021, respectively, which are relatively small. Therefore, these three factors are removed during model construction and iterative training. At the same time, the GBDT algorithm is used to calculate feature weights. The calculation results are shown in Table 5 below:

[0081] Table 5

[0082] factor Weight Wind speed at 10m height 0.021075 Wind direction at 10 meters height 0.028713 Wind speed at 30 meters 0.018114 Wind direction at 30 meters 0.015512 Wind speed at 50m 0.160435 Wind direction at 50 meters 0.015239 Wind speed at 70 meters 0.452199 Wind direction at 70 meters 0.011585 Wind speed at fan hub height 0.265231 Wind direction at fan hub height 0.011897

[0083] It can be seen from the weights that the weights of the wind direction at a height of 70 meters and at the height of the wind turbine hub are 0.011585 and 0.011897 respectively. The weights are relatively low. Therefore, these two factors are eliminated during the model construction and iterative training process.

[0084] The samples in the wide table of the screened data are used as training samples for the power prediction model for training.

[0085] The present invention is based on big data and cloud computing, uses deep neural networks as the main algorithm model, constructs an intelligent power prediction model based on bidirectional LSTM, and combines the business characteristics of wind turbines, uses theoretical power generation balance analysis, etc. to analyze the power structure, calculate the weights and correlation of indicator features, so that power prediction can be truly used to guide business. In addition, it technically improves its shortcomings of being prone to local extreme values, low convergence accuracy, and slow convergence speed in the later stage of the algorithm. The improved algorithm is combined with an extreme learning machine with a kernel function to improve the stability and accuracy of the extreme learning machine.

[0086] From the perspective of benefits, this achievement can strengthen the safety risk control of power grid operation equipment, control the operation and maintenance in the process to pre-control and post-evaluation, and provide auxiliary decision-making basis for ensuring the safe and economic operation of the power grid. Effectively realize business integration, realize the safety risk control of wind turbine equipment, realize the full business information management of power grid equipment, scientific and effective operation monitoring, improve the basis of power grid equipment, timely discover abnormal information of wind turbine equipment, achieve safe and timely operation and maintenance of equipment, increase the life of wind turbine equipment, ensure its reliable operation, improve investment efficiency, and strengthen the organic integration of various businesses and big data applications.

[0087] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A wind turbine power intelligent prediction method based on deep learning, characterized in that: Including the training steps of the power prediction model, Step 100: input multiple groups of feature samples and target samples at the same time, the feature samples include power data of wind turbines, and the target samples include meteorological data of wind turbines and wind measurement data of the station; Step 200: associate and fuse the feature samples and the target samples with the time period as the association condition, and output a wide data table; Step 300: Input the data wide table into the normalization model and output the normalized data wide table. Each sample value in the normalized data wide table ,in, is the maximum sample value at all times, is the minimum sample value at all times, is the original sample value; Step 400: According to the normalized data wide table, the correlation coefficient is combined with the ensemble learning algorithm. First, the weight of the feature factor is calculated by the GBDT algorithm. At the same time, the Pearson correlation coefficient is used to analyze the correlation between the input feature and the target field, that is, the correlation coefficient and weight of the target sample and the feature sample are calculated. The target samples with a correlation coefficient less than the set value of the correlation coefficient and the target samples with a weight less than the set value of the weight are removed from the data wide table. Step 500: construct a wind turbine power prediction model based on a bidirectional LSTM network, input the remaining target samples and feature samples in the data wide table into the wind turbine power prediction model for training, input the feature samples at time t-1, output the target samples corresponding to time t, and incorporate the split power features into the bidirectional LSTM for iterative calculation; split the power into the following formula: ,in is the real-time power, is a trend item, reflecting The changing trend of is a periodic term, the reaction The periodic changes of is a random item, the response stability; Step 600: Calculate the root mean square error of the wind turbine power prediction model and the coefficient of determination , when the root mean square error is less than the root mean square error setting value and the coefficient of determination value is greater than the coefficient of determination setting value, the training is completed and the power prediction model is obtained; The step 400 includes: Step 401: Calculate the correlation coefficient between the meteorological data, wind measurement data and power data in the data table , where n is the number of samples, is the feature sample value, is the feature sample mean, is the target sample value, is the target sample mean, is the standard deviation of the characteristic sample value, is the standard deviation of the target sample value; Step 402: Incorporate the correlation coefficient and characteristic factors into the second-order Taylor expansion and calculate the information entropy ,in is the probability value of the category to which the sample x belongs; Step 403: Using the GBDT multiple base classifier weight calculation idea, the information entropy is incorporated into the weight calculation formula , (i=1,2,…,k) where is the information entropy size of each feature.

2. The method for intelligent prediction of wind turbine power based on deep learning according to claim 1 is characterized in that: The method further includes step 700: using a power prediction model, inputting the meteorological data of the current wind turbine generator set and the wind measurement data of the station into the power prediction model, and outputting the predicted power data of the wind turbine generator set.

3. The method for intelligent prediction of wind turbine power based on deep learning according to claim 1 is characterized in that: The power data includes the measured power value recorded every 15 minutes, the station wind measurement data includes the wind speed at a height of 10 meters, the wind direction at a height of 10 meters, the wind speed at a height of 30 meters, the wind direction at a height of 30 meters, the wind speed at a height of 50 meters, the wind direction at a height of 50 meters, the wind speed at a height of 70 meters, the wind direction at a height of 70 meters, the wind speed at the height of the wind turbine hub, and the wind direction at the height of the wind turbine hub, recorded once every 15 minutes, and the meteorological data includes temperature, air pressure, and relative humidity.

4. The method for intelligent prediction of wind turbine power based on deep learning according to claim 1, characterized in that: Step 200 includes, Step 201: cleaning the feature samples and target samples, including filling missing values ​​with the mean of the sample data and removing outliers; Step 202: Dimension reduction is performed on the feature samples and the target samples, and data in the target samples that do not change over time are removed; Step 203: The feature samples and the target samples are associated and fused with each other based on the time period to obtain a wide data table.

5. The method for intelligent prediction of wind turbine power based on deep learning according to claim 1, characterized in that: The step 600 also includes tuning the parameters of the wind turbine power prediction model through a grid search method, combining the number of neurons in the LSTM layer, the number of neurons in the fully connected layer, the selection of the optimizer, and the selection of the activation function, and adding a cross-validation method during the search process to select parameters while performing cross-validation.

6. The method for intelligent prediction of wind turbine power based on deep learning according to claim 2 is characterized in that: The method further includes step 800: verifying the predicted power data of the wind turbine generator set by using a theoretical power generation balance analysis method, outputting the verification result, and adjusting the wind turbine generator set prediction model based on the verification result.

7. The method for intelligent prediction of wind turbine power based on deep learning according to claim 6, characterized in that: The step 800 includes, Step 801: input wind speed and power data of the wind turbine and perform power curve fitting on them, and calculate the power loss through the fitting curve; Step 802: Calculate theoretical power generation, where the theoretical power generation includes the sum of actual power generation and power loss of the wind turbine generator set; Step 803: Use the theoretical power generation balance analysis method to verify the predicted power data of the wind turbine, output the verification result, and adjust the wind turbine prediction model based on the verification result.

8. The method for intelligent prediction of wind turbine power based on deep learning according to claim 7 is characterized in that: The power loss includes: Planned maintenance loss: refers to the amount of electricity lost due to shutdown of power generation equipment due to planned inspections, regular maintenance, seasonal inspections, condition maintenance and testing of equipment during the statistical period, including planned maintenance loss of power generation equipment and planned maintenance loss of power transmission and transformation equipment; Unplanned maintenance loss: refers to the amount of electricity lost due to wind turbine shutdown caused by failure of wind turbines, transformers, collector lines, and equipment in booster stations during the statistical period, including loss of electricity due to wind turbine failure and loss of electricity due to failure of booster station equipment; Power loss due to power restriction: refers to the power loss during the power restriction period, including power outage and power reduction; Lost electricity: refers to the loss of electricity caused by the shutdown of power generation equipment due to planned maintenance, unplanned maintenance, emergency repair, failure or force majeure of the power grid company during the statistical period, including the loss of electricity due to planned maintenance of the power grid, the loss of electricity due to power grid failure and the loss of electricity due to force majeure; Performance loss electricity: includes self-derating of the unit and failure of the unit to meet performance standards. Self-derating of the unit refers to the unit operating at reduced output due to the unit's own reasons. Failure of the unit to meet performance standards refers to the operating output failing to reach the standard value or the unit being unable to cut in after the wind speed reaches the rated cut-in wind speed.

Images