Wind power prediction error decoupling evaluation method based on k-nearest neighbor search

Abstract

According to the wind power prediction error decoupling evaluation method based on # imgabs0 # nearest neighbor search, prediction errors caused by a power correction link are calculated by using planned and actual startup capacity information, then real meteorological data closest to # imgabs2 # data are searched from historical operation data based on # imgabs1 # nearest neighbor thought, and the prediction errors are calculated according to the real meteorological data and the # imgabs2 # nearest neighbor thought. And estimating a prediction error caused by a # imgabs4 # link through average approximation of # imgabs3 # order neighbor, and finally calculating a prediction error caused by a # imgabs5 # link based on a total prediction error. According to the method, a prediction wind power conversion model of a wind power plant does not need to be directly obtained, high-reliability quantitative evaluation is carried out on errors of different links in the whole wind power prediction process, then the error contribution rate of each link is obtained, and accurate positioning of weak links of a wind power prediction algorithm is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power control, and in particular to a method based on A decoupled evaluation method for wind power forecasting error based on nearest neighbor search. Background Art

[0002] In recent years, with the large-scale grid connection of wind power, the intermittent and volatile nature of its output has made the randomness and uncertainty in the daily operation of the power system more prominent, which has brought great challenges to the safe and economic operation of the power system. In order to effectively deal with the adverse effects of large-scale grid connection of wind power on the power system, accurate and reliable prediction of wind power is required.

[0003] At present, data-driven wind power forecasting methods are the mainstream output forecasting methods. Since data-driven methods start directly from observed data and use machine learning algorithms such as LSTM, RF, and SVM to mine and learn the mapping relationship between observed data and wind power forecast values, data-driven wind power forecasting methods cannot fully reflect the complex coupling relationship between various links in the wind power forecasting process. Therefore, in terms of improving forecast accuracy, existing research basically focuses on the evaluation of the final error results of power forecasting, and rarely achieves the full process time series evaluation of power forecasting. In fact, according to the business process, the wind power forecasting process can be divided into numerical weather prediction (NWP), meteorological-power conversion ( ), power correction. If the key links that cause prediction errors in wind power forecasting can be effectively identified, the accuracy of wind power prediction can be more effectively improved. Therefore, on the basis of a comprehensive analysis of the statistical characteristics of wind power errors, an efficient and reliable decoupling analysis of the errors generated in each link of wind power prediction is carried out, which can provide key technical support for the performance evaluation, improvement and optimization of power prediction algorithms. Summary of the invention

[0004] In order to effectively identify the key links that cause prediction errors in wind power prediction, the present invention provides a method based on A decoupled evaluation method for wind power forecasting error based on nearest neighbor search.

[0005] In order to solve the above technical problems, the present invention adopts the following technical methods: a The wind power prediction error decoupling evaluation method based on nearest neighbor search includes: Step S1, collecting wind farm weather, capacity, and power data; Step S2, the whole process of wind power prediction includes Links, Link, power correction link, calculate the total prediction error of the wind farm , quantitatively evaluate the prediction errors caused by each link; S201, predict the wind farm power The actual power corresponding to Make a difference and get the total prediction error of the wind farm ; S202, first calculate the equivalent power prediction value under the condition of accurate startup capacity , and then the wind farm predicted power Equivalent power prediction value The difference between ; S203, build a method for quickly finding the nearest neighbor Nearest Neighbor Model, using historical weather data right The nearest neighbor model is fitted to form historical meteorological data With historical power data The index relationship of Nearest neighbor model Dataset Perform a nearest neighbor search for each sample in the The average actual power under the nearest neighbor , get the average actual power data set composed of all samples , the equivalent power prediction value The average actual power data set The difference is Prediction error caused by ; S204, the total wind farm prediction error is obtained Prediction error caused by power correction and Prediction error caused by Do the wrong thing and get Prediction error caused by ; Step S3: Links, The prediction errors caused by the wind power correction link and the power correction link are normalized respectively to obtain the error contribution rate of the corresponding link. The error contribution rates of the three links are compared to determine the key links that cause prediction errors in the entire process of wind power prediction.

[0006] Further, step S203 includes: 1) Obtain the actual power of the wind farm At the same time Dataset , and by Historical meteorological data collected by the system With historical power data ; 2) Initialize a method for quickly finding the nearest neighbor The nearest neighbor model uses The dimensional tree structure is used as the feature space partitioning algorithm, and the Euclidean distance is used to calculate the distance between samples. right The nearest neighbor model is fitted. Build a spatial index for subsequent nearest neighbor search; 3) Remember The The sample at each moment is , using the fitting Nearest neighbor model Perform a nearest neighbor search, starting from Find and The closest Historical meteorological data, and record the historical meteorological data index. The corresponding actual power data set is obtained from ,calculate Inside The average value of actual power data is obtained The average actual power under the nearest neighbor ; 4) Repeat step 3) until the traversal is complete All samples, the average actual power data set of all samples is recorded as ,Will and The difference as Prediction error caused by .

[0007] Furthermore, in step S202, the planned startup capacity of the wind farm is recorded. , actual startup capacity , Indicates a moment, is the total number of samples. First, calculate the equivalent power prediction value under the condition of accurate startup capacity. , and then calculate the prediction error caused by the power correction link , as follows: (1) (2) Furthermore, step S3 includes: 1) Calculate separately The sum of the absolute values ​​of all samples in , as follows: (3) (4) (5) In the formula, Respectively Middle The sample value at each moment, the total number of samples is ; 2) Respectively After normalization, we get Error contribution rate of each link , Error contribution rate of each link , Error contribution rate of power correction link , as follows: (6) (7) (8) 3) Comparison Error contribution rate of each link , Error contribution rate of each link , Error contribution rate of power correction link The larger the value, the greater the impact of the corresponding link on the prediction error in the whole process of wind power prediction.

[0008] Preferably, the Dataset And historical weather data All of them are wind farm meteorological data, and the wind farm meteorological data include wind speed, wind direction, temperature, humidity and air pressure at different heights.

[0009] The present invention provides The wind power prediction error decoupling evaluation method based on the nearest neighbor search can realize the error decoupling evaluation of each link in the whole process of wind power prediction. Specifically, the present invention first uses the planned and actual startup capacity information to calculate the prediction error caused by the power correction link, and then based on The idea of ​​the nearest neighbor is to find the corresponding The data is closest to the real meteorological data. The average approximation of the nearest neighbors The prediction error caused by each link is finally calculated based on the total prediction error The present invention does not need to directly obtain the predicted wind power conversion model of the wind farm. It performs a highly reliable quantitative evaluation of the errors in different links in the entire process of wind power prediction, and then obtains the error contribution rate of each link, thereby accurately locating the weak links of the wind power prediction algorithm, and providing a more comprehensive prediction error analysis from the perspective of the business process of power prediction. The present invention can guide the existing prediction algorithm to improve and optimize the weak links of the prediction, thereby further improving the accuracy of wind power prediction, and help to improve the wind power dispatching platform to implement more comprehensive and reliable full-process real-time perception of the power prediction level of each wind farm station in each region. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 The present invention provides Flowchart of the wind power forecast error decoupling evaluation method based on nearest neighbor search. DETAILED DESCRIPTION

[0011] In order to facilitate the understanding of those skilled in the art, the present invention is further described below in conjunction with embodiments and drawings. The contents mentioned in the implementation modes are not intended to limit the present invention.

[0012] like Figure 1 As shown, a The wind power prediction error decoupling evaluation method based on nearest neighbor search includes: Step S1, collecting weather, capacity and power data of a wind farm in central China at different times.

[0013] Wind farm meteorological data includes Dataset And historical weather data , which are wind speed, wind direction, temperature, humidity and air pressure at different altitudes. Dataset The meteorological data in is the numerical weather forecast. , Indicates a moment, is the total number of samples, Provided by power forecasting vendors, specifically: first, download the global weather forecast field from authoritative meteorological agencies, then standardize the format of the global atmospheric forecast field data to make it capable of driving mesoscale The conditions for the model software to run, complete all the preparations before the model runs, and finally, the power forecast service provider runs the mesoscale numerical weather model software according to the forecast requirements of its detailed geographic coordinates, completes the downscaling calculation of the local target area, and finally obtains the atmospheric state of the geographical area where the wind farm is located at different times in the future. It is the real meteorological data collected by the SCADA (Supervisory Control And Data Acquisition) system.

[0014] Wind farm capacity data includes the planned start-up capacity of wind farms and actual startup capacity , Indicates a moment, is the total number of samples.

[0015] Wind farm power data includes actual power during the same period With predicted power , and historical meteorological data Corresponding historical power data .in, and All by The actual power generation of the wind farm collected by the system; For the general Dataset through The power obtained after the wind power conversion (it is worth mentioning here that The predictive wind power conversion model provided by the manufacturer is used for data processing). In addition, Indicates a moment, is the total number of samples.

[0016] In this embodiment, the aforementioned Preferred is 2880.

[0017] Step S2, the whole process of wind power prediction includes Links, Link, power correction link, calculate the total prediction error of the wind farm , and quantitatively evaluate the prediction errors caused by each link.

[0018] S201, predict the wind farm power The actual power corresponding to Make a difference and get the total prediction error of the wind farm , as follows: (9) S202, first calculate the equivalent power prediction value under the condition of accurate startup capacity , and then calculate the prediction error caused by the power correction link , as follows: (1) (2) S203, through Average calculation of nearest neighbors The prediction error caused by the link is as follows: 1) Obtain the actual power of the wind farm At the same time Dataset , and by Actual historical meteorological data collected by the system Compared with the actual historical power data .

[0019] 2) Initialize a method for quickly finding the nearest neighbor The nearest neighbor model (this model is different from the traditional classification or regression model. It is a nearest neighbor search model that is specifically used to find a sample nearest neighbors, is the number of nearest neighbors), use K-Dimensional Tree (KD Tree) is used as a feature space partitioning algorithm ( The dimension tree is a binary tree structure that can directly exclude areas that are far away from the query point, thereby quickly searching for the nearest neighbor). The Euclidean distance is used to calculate the distance between samples (sample and The Euclidean distance between ,remember is the number of features), using right The nearest neighbor model is fitted. This process will be A spatial index is built on the , to facilitate the subsequent rapid search for the nearest neighbor. After the fitting is completed, any given sample can be directly queried in In this embodiment The nearest neighbor algorithm is implemented using the Sklearn toolbox in the Python environment.

[0020] 3) Remember The The sample at each moment is , using the fitting Nearest neighbor model Perform a nearest neighbor search, starting from Find and The closest Historical meteorological data, and record the historical meteorological data index. The corresponding actual power data set is obtained from ,calculate Inside The average value of actual power data is obtained The average actual power under the nearest neighbor .

[0021] 4) Repeat step 3) until the traversal is complete All samples, the average actual power data set of all samples is recorded as ,Will and The difference as Prediction error caused by , as follows: (10).

[0022] S204, the total wind farm prediction error is obtained Prediction error caused by power correction and Prediction error caused by Do the difference and get Prediction error caused by , as follows: (11).

[0023] Step S3: Links, The prediction errors caused by the wind power forecasting link and the power correction link are normalized respectively to obtain the error contribution rate of the corresponding link, and the error contribution rates of the three links are compared to determine the key links that cause the prediction error in the whole process of wind power forecasting. 1) Calculate separately middle The sum of the absolute values ​​of samples , as follows: (3) (4) (5) In the formula, Respectively Middle The sample value at each moment, the total number of samples is .

[0024] 2) Respectively After normalization, we get Error contribution rate of each link , Error contribution rate of each link , Error contribution rate of power correction link , as follows: (6) (7) (8) In this example =0.80%, =43.16%, =56.04%.

[0025] 3) Comparison Error contribution rate of each link Error contribution rate of each link , Error contribution rate of power correction link The numerical value of can be used to further analyze the cause of the overall wind power forecast error according to the error contribution rate, so as to locate the weak links in the whole process of wind power forecasting and guide the subsequent improvement of wind power forecasting accuracy. In this example, The prediction error caused by the link is the largest, followed by The prediction error caused by the power correction link is the smallest. The accuracy of the link is the main reason for the prediction error. The prediction performance of the wind power conversion model itself used in the link is an important reason for the prediction error. Compared with the above two links, the power correction link has less impact on the wind power prediction error.

[0026] The above embodiments are preferred implementation schemes of the present invention. In addition, the present invention may also be implemented in other ways. Any obvious replacement without departing from the concept of the present technical solution is within the protection scope of the present invention.

[0027] In order to make it easier for ordinary technicians in the field to understand the improvements of the present invention over the prior art, some drawings and descriptions of the present invention have been simplified, and for the sake of clarity, some other elements are omitted in this application document. Ordinary technicians in the field should realize that these omitted elements may also constitute the content of the present invention.

Claims

1. Based on The wind power prediction error decoupling evaluation method based on nearest neighbor search is characterized by: include: Step S1, collecting wind farm weather, capacity, and power data; Step S2, the whole process of wind power prediction includes Links, Link, power correction link, calculate the total prediction error of the wind farm , quantitatively evaluate the prediction errors caused by each link; S201, predict the wind farm power The actual power corresponding to Make a difference and get the total prediction error of the wind farm ; S202, first calculate the equivalent power prediction value under the condition of accurate startup capacity , and then the wind farm predicted power Equivalent power prediction value The difference between ; S203, build a method for quickly finding the nearest neighbor Nearest Neighbor Model, using historical weather data right The nearest neighbor model is fitted to form historical meteorological data With historical power data The index relationship of Nearest neighbor model Dataset Perform a nearest neighbor search for each sample in the The average actual power under the nearest neighbor , get the average actual power data set composed of all samples , the equivalent power prediction value The average actual power data set The difference is Prediction error caused by ; S204, the total wind farm prediction error is obtained Prediction error caused by power correction and Prediction error caused by Do the wrong thing and get Prediction error caused by ; Step S3: Links, The prediction errors caused by the wind power correction link and the power correction link are normalized respectively to obtain the error contribution rate of the corresponding link. The error contribution rates of the three links are compared to determine the key links that cause prediction errors in the entire process of wind power prediction.

2. According to claim 1, based on The wind power prediction error decoupling evaluation method based on nearest neighbor search is characterized by: Step S203 includes: 1) Obtain the actual power of the wind farm At the same time Dataset , and by Historical meteorological data collected by the system With historical power data ; 2) Initialize a method for quickly finding the nearest neighbor The nearest neighbor model uses The dimensional tree structure is used as the feature space partitioning algorithm, and the Euclidean distance is used to calculate the distance between samples. right The nearest neighbor model is fitted. Build a spatial index for subsequent nearest neighbor search; 3) Remember The The sample at the moment is , using the fitting Nearest neighbor model Perform a nearest neighbor search from Find and The closest Historical meteorological data, and record the historical meteorological data index. The corresponding actual power data set is obtained from ,calculate Inside The average value of actual power data is obtained The average actual power under the nearest neighbor ; 4) Repeat step 3) until the traversal is complete All samples, the average actual power data set of all samples is recorded as ,Will and The difference as Prediction error caused by .

3. According to claim 2 based on The wind power prediction error decoupling evaluation method based on nearest neighbor search is characterized by: In step S202, the planned startup capacity of the wind farm is recorded. , actual startup capacity , Indicates a moment, is the total number of samples. First, calculate the equivalent power prediction value under the condition of accurate startup capacity. , and then calculate the prediction error caused by the power correction link , as follows: （1） （2）。 4. The method according to claim 3 The wind power prediction error decoupling evaluation method based on nearest neighbor search is characterized by: Step S3 includes: 1) Calculate separately The sum of the absolute values ​​of all samples in , as follows: （3） （4） （5） In the formula, Respectively Middle The sample value at each moment, the total number of samples is ; 2) Respectively After normalization, we get Error contribution rate of each link , Error contribution rate of each link , Error contribution rate of power correction link , as follows: （6） （7） （8） 3) Comparison Error contribution rate of each link , Error contribution rate of each link , Error contribution rate of power correction link The larger the value, the greater the impact of the corresponding link on the prediction error in the whole process of wind power prediction.

5. The method according to claim 1 or 4. The wind power prediction error decoupling evaluation method based on nearest neighbor search is characterized by: Said Dataset And historical weather data All of them are wind farm meteorological data, and the wind farm meteorological data include wind speed, wind direction, temperature, humidity and air pressure at different heights.

Images