Wind field downscaling simulation prediction system and method based on machine learning

Abstract

The invention provides a wind field downscaling simulation prediction system and method based on machine learning, and the system comprises a first prediction data obtaining subsystem which is used for obtaining the grid prediction data of a low-resolution mode wind field; the live data acquisition subsystem is used for acquiring live data; the downscaling model construction subsystem is used for constructing a downscaling model based on a machine learning technology; and the second forecast data acquisition subsystem is used for performing downscaling correction according to the low-resolution mode wind field grid forecast data and the live data based on a downscaling model, and outputting high-resolution mode wind field grid forecast data. According to the wind field downscaling simulation prediction system and method based on machine learning, physical parameters, historical data and real-time monitoring data in a WRF mode are optimized and adjusted by combining WRF mode output with the resolution of 5 kilometers per hour and are used for dynamically adjusting parameters of the WRF mode, the wind field is predicted in a refined mode, and the calculated amount is smaller.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular to a wind field downscaling simulation prediction system and method based on machine learning. Background Art

[0002] At present, most wind forecast products come from regional or global numerical weather forecast models. However, the forecast based on dynamical processes has initial condition deviations and model uncertainty, which makes the numerical model unable to capture the small-scale characteristics of the atmosphere and micro-scale turbulent motion. The accuracy of wind forecast under complex terrain has always been a difficulty and focus in the field of wind energy research and meteorological research. At present, the use of dynamic downscaling methods to obtain higher-resolution wind products and wind speed deviation correction has achieved some success in some specific case tests, but due to the excessive demand for computing resources, it is temporarily unable to meet the needs of refined forecasting.

[0003] The invention patent with application number: CN202011237572.3 discloses an automated operation system for mesoscale wind energy resource analysis, including parameter configuration script files, mesoscale model WRF pre-processing WPS operation script files, mesoscale model WRF operation scripts, data assimilation script files, model post-processing data assimilation script files, model product generation LRPP script files, and operation control script files; it avoids the complex job flow control software such as rocoto, ECflow, and Airflow commonly used in meteorological business, and uses simple bash scripts to implement job submission, operation, monitoring, and compensation functions; all parameter configurations are set in the parameter configuration script file, and the remaining scripts adopt a modular approach to achieve arbitrary operation; it realizes the division of a regional range with a large number of grid points into several relatively small grid areas for mesoscale simulation, and performs regional splicing in the product stage, but the above-mentioned prior art only realizes the automated operation of model simulation, and does not evaluate or improve or enhance the model simulation effect.

[0004] In addition, the hourly grid-corrected forecast products generated by the forecast correction technical solutions in other existing technologies cannot be completely consistent with the extreme grid-corrected forecast products. Secondly, the error direction of the hourly numerical model products at the beginning of the forecast time zero and the hourly multi-source grid fusion products may be opposite to the error direction of the hourly numerical model products at the future forecast time. If the correction scheme is not appropriate, the error will increase. In addition, dynamic downscaling consumes huge computing resources because all grid points at all forecast times need to be modeled, which greatly reduces the efficiency of rolling correction of numerical model products and cannot meet the needs of refined forecasting for the time being.

[0005] In view of this, there is an urgent need for a wind field downscaling simulation and prediction system and method based on machine learning to at least solve the above-mentioned deficiencies. Summary of the invention

[0006] One of the purposes of the present invention is to provide a wind field downscaling simulation prediction system and method based on machine learning, which optimizes and adjusts the physical parameters in the WRF model by combining the WRF model output with a resolution of 5 kilometers per hour, and realizes the mapping from low resolution to high resolution. Historical data and real-time monitoring data are used to dynamically adjust the parameters of the WRF model, helping the model to better simulate the variation law of the local wind field, and to refine the prediction of the wind field with less calculation.

[0007] An embodiment of the present invention provides a wind field downscaling simulation prediction system based on machine learning, comprising:

[0008] The first forecast data acquisition subsystem is used to obtain 5km*5km model wind field grid forecast data;

[0009] A live data acquisition subsystem, used for acquiring live data;

[0010] A downscaling model construction subsystem is used to construct a downscaling model based on machine learning technology;

[0011] The second forecast data acquisition subsystem is used to perform downscaling correction based on the 5km*5km model wind field grid forecast data and actual data based on the downscaling model, and output 1km*1km model wind field grid forecast data.

[0012] Preferably, the first forecast data acquisition subsystem acquires 5km*5km model wind field grid forecast data, including:

[0013] Configure simulation domain parameters;

[0014] After the simulation domain parameters are configured, run the WPS program;

[0015] After the WPS program is executed, the WRF model is used to perform meteorological simulation;

[0016] When the WRF model performs meteorological simulation, data assimilation is performed;

[0017] After data assimilation, UPP / LRPP is run to obtain the output 5km*5km model wind field grid forecast data.

[0018] Preferably, the downscaling model construction subsystem constructs a downscaling model based on machine learning technology, including:

[0019] Based on the gradient boosting decision tree, a downscaling model is constructed;

[0020] The input of the downscaling model is the wind field features extracted from the WRF model output. Through the wind field features, the GBDT model learns the mapping relationship between the low-resolution model output and the high-resolution wind field observation data. The basic form of the model prediction is:

[0021]

[0022] Among them, F(x) is the final model prediction output, M is the total number of decision trees, and h m (x) is the mth decision tree, α m is the learning rate of m decision trees, and x is the input wind field feature;

[0023] Among them, the GBDT model learns the model by minimizing the loss function, and the loss function is:

[0024]

[0025] Among them, y i is the actual observed wind speed of the ith data sample, F(x i ) is the wind speed predicted by the model for the ith data sample, where n is the number of data samples;

[0026] The GBDT model fits a new decision tree by using negative gradients. The formula for calculating the negative gradient in the mth round is as follows:

[0027]

[0028] After fitting the new decision tree, the overall model is updated according to the update formula, which is specifically:

[0029] F m (x) = F m-1 (x)+α m h m (x)

[0030] When the preset number of trees M is reached or the negative gradient converges, the GBDT model completes learning and obtains the final downscaling model.

[0031] Preferably, the live data acquisition subsystem acquires live data, including:

[0032] Get the set of live data monitoring nodes;

[0033] Determine a first monitoring site of a live data monitoring node;

[0034] Obtaining a monitoring value of a live data monitoring node corresponding to a first monitoring site;

[0035] If the number of monitoring values ​​of the live data monitoring node corresponding to the first monitoring site is greater than or equal to the preset target number, the corresponding first monitoring site is used as the second monitoring site, otherwise, the first monitoring site is used as the third monitoring site;

[0036] Based on the preset outlier elimination rule, the outliers of the second monitoring site are eliminated, and the average value of the monitoring values ​​of the remaining live data monitoring nodes is calculated, and the average value is used as the live data of the second monitoring site;

[0037] Calculate the simulated value of the third monitoring station based on the average value of the second monitoring station;

[0038] Attempt to obtain a monitoring value of a live data monitoring node of a third monitoring site, and if the acquisition attempt is successful, calculate a deviation value between the monitoring value of the corresponding live data monitoring node and a simulated value;

[0039] If the deviation value is less than or equal to the preset deviation value threshold, the monitoring value of the corresponding live data monitoring node is used as the live data of the third monitoring site;

[0040] If the deviation values ​​are all greater than the preset deviation value threshold, the simulated values ​​are used as the actual data of the third monitoring site;

[0041] Calculating the simulated value of the third monitoring station according to the average value of the second monitoring station includes:

[0042] According to the wind field characteristics of the first monitoring station, a spatial wind field model is constructed;

[0043] Construct an input matrix based on the average value of the second monitoring station;

[0044] Determine the dynamic simulation model of the space wind field based on the space wind field model and the knowledge information of the space wind field mechanism;

[0045] Input the input matrix into the spatial wind field dynamic simulation model to obtain the output matrix;

[0046] The simulation value of the third monitoring site is determined according to the simulation result interpretation template corresponding to the output matrix.

[0047] An embodiment of the present invention provides a wind field downscaling simulation prediction system based on machine learning, further comprising:

[0048] Environmental early warning subsystem, used to provide environmental early warning based on 1km*1km model wind field grid forecast data;

[0049] Among them, the environmental warning subsystem performs environmental warning based on the 1km*1km model wind field grid forecast data, including:

[0050] According to the simulation domain parameters, the distribution information of wind farm-related pollution sources is obtained;

[0051] Based on the wind farm associated pollution source characterization template and according to the wind farm associated pollution source distribution information, the wind farm associated pollution source characteristics are extracted, wherein the wind farm associated pollution source characteristics include: type, location and movement characteristic information of the wind farm associated pollution source;

[0052] Based on the diffusion description model, the pollution source diffusion description is determined according to the characteristics of wind field-associated pollution sources and the 1km*1km model wind field grid forecast data;

[0053] Obtain the pollution level evaluation template corresponding to the pollution source;

[0054] Predict the pollution level in the pre-diffusion area based on the pollution source diffusion description and pollution level evaluation template;

[0055] Obtaining the preset pollution level index of the pre-diffusion area;

[0056] If the pollution level is greater than the pollution level index, an environmental warning for the corresponding pre-diffusion area is generated.

[0057] An embodiment of the present invention provides a wind field downscaling simulation prediction method based on machine learning, comprising:

[0058] Step 1: Obtain 5km*5km model wind field grid forecast data;

[0059] Step 2: Get live data;

[0060] Step 3: Build a downscaling model based on machine learning technology;

[0061] Step 4: Based on the downscaling model, downscaling correction is performed according to the 5km*5km model wind field grid forecast data and the actual data, and the 1km*1km model wind field grid forecast data is output.

[0062] Preferably, step 1: obtaining 5km*5km model wind field grid forecast data includes:

[0063] Configure simulation domain parameters;

[0064] After the simulation domain parameters are configured, run the WPS program;

[0065] After the WPS program is executed, the WRF model is used to perform meteorological simulation;

[0066] When the WRF model performs meteorological simulation, data assimilation is performed;

[0067] After data assimilation, UPP / LRPP is run to obtain the output 5km*5km model wind field grid forecast data.

[0068] Preferably, step 3: constructing a downscaling model based on machine learning technology, including:

[0069] Based on the gradient boosting decision tree, a downscaling model is constructed;

[0070] The input of the downscaling model is the wind field features extracted from the WRF model output. Through the wind field features, the GBDT model learns the mapping relationship between the low-resolution model output and the high-resolution wind field observation data. The basic form of the model prediction is:

[0071]

[0072] Among them, F(x) is the final model prediction output, M is the total number of decision trees, and h m (x) is the mth decision tree, α m is the learning rate of m decision trees, and x is the input wind field feature;

[0073] Among them, the GBDT model learns the model by minimizing the loss function, and the loss function is:

[0074]

[0075] Among them, y i is the actual observed wind speed of the ith data sample, F(x i ) is the wind speed predicted by the model for the ith data sample, where n is the number of data samples;

[0076] The GBDT model fits a new decision tree by using negative gradients. The formula for calculating the negative gradient in the mth round is as follows:

[0077]

[0078] After fitting the new decision tree, the overall model is updated according to the update formula, which is specifically:

[0079] F m (x) = F m-1 (x)+α m h m (x)

[0080] When the preset number of trees M is reached or the negative gradient converges, the GBDT model completes learning and obtains the final downscaling model.

[0081] Preferably, step 2: obtaining live data includes:

[0082] Get the set of live data monitoring nodes;

[0083] Determine a first monitoring site of a live data monitoring node;

[0084] Obtaining a monitoring value of a live data monitoring node corresponding to a first monitoring site;

[0085] If the number of monitoring values ​​of the live data monitoring node corresponding to the first monitoring site is greater than or equal to the preset target number, the corresponding first monitoring site is used as the second monitoring site, otherwise, the first monitoring site is used as the third monitoring site;

[0086] Based on the preset outlier elimination rule, the outliers of the second monitoring site are eliminated, and the average value of the monitoring values ​​of the remaining live data monitoring nodes is calculated, and the average value is used as the live data of the second monitoring site;

[0087] Calculate the simulated value of the third monitoring station based on the average value of the second monitoring station;

[0088] Attempt to obtain a monitoring value of a live data monitoring node of a third monitoring site, and if the acquisition attempt is successful, calculate a deviation value between the monitoring value of the corresponding live data monitoring node and a simulated value;

[0089] If the deviation value is less than or equal to the preset deviation value threshold, the monitoring value of the corresponding live data monitoring node is used as the live data of the third monitoring site;

[0090] If the deviation values ​​are all greater than the preset deviation value threshold, the simulated values ​​are used as the actual data of the third monitoring site;

[0091] Calculating the simulated value of the third monitoring station according to the average value of the second monitoring station includes:

[0092] According to the wind field characteristics of the first monitoring station, a spatial wind field model is constructed;

[0093] Construct an input matrix based on the average value of the second monitoring station;

[0094] Determine the dynamic simulation model of the space wind field based on the space wind field model and the knowledge information of the space wind field mechanism;

[0095] Input the input matrix into the spatial wind field dynamic simulation model to obtain the output matrix;

[0096] The simulation value of the third monitoring site is determined according to the simulation result interpretation template corresponding to the output matrix.

[0097] An embodiment of the present invention provides a wind field downscaling simulation prediction method based on machine learning, further comprising:

[0098] Environmental early warning based on 1km*1km model wind field grid forecast data;

[0099] Among them, environmental warning is carried out based on the 1km*1km model wind field grid forecast data, including:

[0100] According to the simulation domain parameters, the distribution information of wind farm-related pollution sources is obtained;

[0101] Based on the wind farm associated pollution source characterization template and according to the wind farm associated pollution source distribution information, the wind farm associated pollution source characteristics are extracted, wherein the wind farm associated pollution source characteristics include: type, location and movement characteristic information of the wind farm associated pollution source;

[0102] Based on the diffusion description model, the pollution source diffusion description is determined according to the characteristics of wind field-associated pollution sources and the 1km*1km model wind field grid forecast data;

[0103] Obtain the pollution level evaluation template corresponding to the pollution source;

[0104] Predict the pollution level in the pre-diffusion area based on the pollution source diffusion description and pollution level evaluation template;

[0105] Obtaining the preset pollution level index of the pre-diffusion area;

[0106] If the pollution level is greater than the pollution level index, an environmental warning for the corresponding pre-diffusion area is generated.

[0107] The beneficial effects of the present invention are:

[0108] The present invention optimizes and adjusts the physical parameters in the WRF model by combining the output of the WRF model with a resolution of 5 kilometers per hour, thus achieving mapping from low resolution to high resolution. Historical data and real-time monitoring data are used to dynamically adjust the parameters of the WRF model, helping the model to better simulate the variation of local wind fields, and to make more refined predictions of wind fields with less calculations.

[0109] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures specifically pointed out in the present application documents.

[0110] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0111] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0112] Figure 1 Schematic diagram of a wind field downscaling simulation and prediction system based on machine learning in an embodiment of the present invention;

[0113] Figure 2 It is a schematic diagram of a wind field downscaling simulation prediction method based on machine learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0114] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0115] The embodiment of the present invention provides a wind field downscaling simulation prediction system based on machine learning, such as Figure 1 As shown, including:

[0116] The first forecast data acquisition subsystem 1 is used to acquire 5km*5km model wind field grid forecast data;

[0117] A live data acquisition subsystem 2, used for acquiring live data;

[0118] A downscaling model construction subsystem 3 is used to construct a downscaling model based on machine learning technology;

[0119] The second forecast data acquisition subsystem 4 is used to perform downscaling correction based on the 5km*5km model wind field grid forecast data and the actual data based on the downscaling model, and output 1km*1km model wind field grid forecast data;

[0120] The first forecast data acquisition subsystem acquires 5km*5km model wind field grid forecast data, including:

[0121] Configure simulation domain parameters; where simulation domain parameters are: parameters of the defined wind farm simulation area;

[0122] After the simulation domain parameters are configured, run the WPS program;

[0123] After the WPS program is executed, the WRF model is used to perform meteorological simulation;

[0124] When the WRF model performs meteorological simulation, data assimilation is performed;

[0125] After data assimilation, UPP / LRPP is run to obtain the output 5km*5km model wind field grid forecast data.

[0126] The working principle and beneficial effects of the above technical solution are:

[0127] In order to improve the accuracy of the WRF model wind field prediction, a downscaling correction method based on machine learning is introduced, with the gradient boosting machine (GBDT) as the core algorithm. The input of the downscaling model mainly includes a variety of features closely related to the wind field extracted from the WRF model output, mainly including: temperature, humidity, surface type and other information, to help the model better understand the spatiotemporal distribution of wind fields under different conditions. The output results of the WRF model are used as the input of the model, and the wind field data of the observation station are used as the label / target output. The gradient boosting machine is selected to establish the downscaling model, and the machine learning model is trained to learn the mapping relationship from coarse resolution to high resolution. For the trained model, an independent historical data set is used to verify the model and evaluate the prediction performance of the model.

[0128] In the process of building a downscaling model based on the gradient boosting decision tree (GBDT), the input of the model is mainly the wind field-related features extracted from the WRF model output. Through the input features, the GBDT model learns the mapping relationship between the low-resolution model output and the high-resolution wind field observation data. The basic form of the model prediction is:

[0129]

[0130] Among them, F(x) is the final model prediction output, M is the total number of decision trees, and h m (x) is the mth decision tree, x is the input wind field feature, α m The weights (or learning rates) of the m decision trees control the impact of each tree on the final prediction;

[0131] The model is learned by minimizing the loss function. Typically, GBDT uses the squared error loss function, which is defined as follows:

[0132]

[0133] Among them, y i is the actual observed wind speed (station monitoring data) of the ith data sample, F(x i ) is the wind speed predicted by the model for the ith data sample, where n is the number of data samples;

[0134] GBDT fits a new decision tree by using negative gradients. The gradient of each round is calculated as the negative gradient of the current model to the loss function. For the squared error loss function, the residual (negative gradient) calculation formula for the mth round is as follows:

[0135]

[0136] The update rule for each round of GBDT is to fit a new decision tree h based on the residual m (x), and then update the overall model by the following formula:

[0137] F m (x) = F m-1 (x)+α m h m (x)

[0138] In the downscaling process of GBDT, different features are used to learn the spatiotemporal distribution of wind speed through decision trees. Each tree constructs nodes based on the splitting of different features. The splitting feature of each node is determined by maximizing the reduction of prediction error. Feature selection is performed based on information gain, and then a decision tree is recursively constructed. The feature with the largest information gain value is the optimal feature, that is, the larger the information gain, the higher it should be placed in the upper layer of the decision tree.

[0139] In this way, GBDT learns the mapping relationship from low-resolution model output to high-resolution wind field observation data.

[0140] Finally, the model is further optimized through an iterative process. After obtaining the preliminary results, it is evaluated by the root mean square error and deviation, and then multiple rounds of iterations are performed. After each round of iteration, the model output results will be evaluated, and the model will be further adjusted and updated according to the evaluation: until the preset number of trees M is reached or the residual converges. Multiple rounds of repeated training and adjustment are repeated to continuously improve the prediction results. As the number of iterations of the model expands, the wind speed in the area can achieve a more natural transition. By gradually fitting the residuals and multiple rounds of iterations, the model fitting ability is optimized, and by learning the spatiotemporal characteristics of historical data, the real observation values ​​are gradually approached.

[0141] In the result verification, the RMSE (root mean square error) and Bias (bias) indicators are used to measure the prediction performance of the model. If the verification results of the model show a high prediction error (such as RMSE is too large, or Bias shows a systematic deviation), the model is further adjusted, the feature selection is changed according to the information gain, and the variables that are not significantly related to the wind speed forecast are eliminated to further improve the accuracy of the model.

[0142] The present invention constructs a wind field prediction model with a resolution of 1 km based on the WRF model simulation results combined with a machine learning method, develops a wind field simulation forecasting automatic operation system based on the Linux system bash script control according to the wind field prediction model, realizes the automatic operation of the wind field prediction system, and provides convenience for scientific research and business development applications; by combining the WRF model output with a resolution of 5 km per hour, the physical parameters (such as meteorological elements, terrain, etc.) in the WRF model are optimized and adjusted to achieve mapping from low resolution to high resolution. Historical data and real-time monitoring data are used to dynamically adjust the parameters of the WRF model, helping the model to better simulate the variation law of the local wind field, and to refine the prediction of the wind field with less calculation.

[0143] In one embodiment, the live data acquisition subsystem acquires live data, including:

[0144] Acquire a set of live data monitoring nodes; wherein the live data monitoring nodes are network nodes of different types of live data monitoring parties, such as: a globally distributed ground weather station network, a local automated weather monitoring station;

[0145] Determine a first monitoring site of the live data monitoring node; wherein the first monitoring site is: a meteorological monitoring point of the live data monitoring node;

[0146] Obtaining a monitoring value of a live data monitoring node corresponding to the first monitoring site; wherein the number of monitoring values ​​of the live data monitoring node corresponding to the first monitoring site may be one or more; the monitoring value of the live data monitoring node is a meteorological measurement value of the live data monitoring node corresponding to the first monitoring site, such as wind speed and precipitation;

[0147] If the number of monitoring values ​​of the live data monitoring node corresponding to the first monitoring site is greater than or equal to the preset target number, the corresponding first monitoring site is used as the second monitoring site, otherwise, the first monitoring site is used as the third monitoring site; wherein the preset target number is manually preset, for example: 3;

[0148] Based on the preset outlier elimination rule, the outliers of the second monitoring site are eliminated, and the average value of the monitoring values ​​of the remaining live data monitoring nodes is calculated, and the average value is used as the live data of the second monitoring site; wherein, the outliers of the second monitoring site are eliminated based on the preset outlier elimination rule, including: traversing the monitoring values ​​of the live data monitoring nodes of the second monitoring site, calculating the difference data between the monitoring value of the live data monitoring node being traversed and the monitoring values ​​of other live data monitoring nodes corresponding to the second monitoring site, if the data characteristics of the difference data meet the outlier difference data characteristics (each corresponding calculated difference is greater than the preset difference threshold), then the monitoring value of the live data monitoring node being traversed is used as an outlier;

[0149] Calculate the simulated value of the third monitoring station based on the average value of the second monitoring station;

[0150] Attempt to obtain a monitoring value of a live data monitoring node of a third monitoring site, and if the acquisition attempt is successful, calculate a deviation value between the monitoring value of the corresponding live data monitoring node and a simulated value;

[0151] If the deviation value is less than or equal to a preset deviation value threshold, the monitoring value of the corresponding live data monitoring node is used as the live data of the third monitoring site; wherein the preset deviation value threshold is manually preset;

[0152] If the deviation values ​​are all greater than the preset deviation value threshold, the simulated values ​​are used as the actual data of the third monitoring site;

[0153] Calculating the simulated value of the third monitoring station according to the average value of the second monitoring station includes:

[0154] According to the wind field characteristics of the first monitoring station, a spatial wind field model is constructed; wherein the spatial wind field model is: a three-dimensional model of the simulation area constructed based on the wind field characteristics;

[0155] According to the average value of the second monitoring station, an input matrix is ​​constructed; the input matrix is: a wind speed description matrix of known simulation points in the spatial wind field model;

[0156] According to the spatial wind field model and the spatial wind field mechanism knowledge information, the spatial wind field dynamic simulation model is determined; wherein the spatial wind field dynamic simulation model is: a mathematical model that describes and simulates the distribution and change of wind in space;

[0157] Input the input matrix into the spatial wind field dynamic simulation model to obtain the output matrix; wherein the output matrix is: the wind speed description matrix of all simulation points in the spatial wind field model;

[0158] The simulation value of the third monitoring site is determined according to the simulation result interpretation template corresponding to the output matrix.

[0159] The working principle and beneficial effects of the above technical solution are:

[0160] The present invention connects to live data monitoring nodes of different node types, obtains the live data monitoring node monitoring value corresponding to the first monitoring site, and one first monitoring site may correspond to one or more live data monitoring node monitoring values. The monitoring methods of live data monitoring nodes of different node types are different. Therefore, in order to further improve the accuracy of the measured value, if the number of live data monitoring node monitoring values ​​corresponding to the first monitoring site is greater than or equal to the preset target number, it is used as the second monitoring site;

[0161] Traversing the monitoring values ​​of the live data monitoring nodes of the second monitoring site, calculating the difference data between the monitoring value of the live data monitoring node being traversed and the monitoring values ​​of other live data monitoring nodes corresponding to the second monitoring site, if each corresponding calculated difference is greater than a preset difference threshold, taking the monitoring value of the live data monitoring node being traversed as an outlier, eliminating the outlier of the second monitoring site, and calculating the average value of the monitoring values ​​of the remaining live data monitoring nodes;

[0162] A spatial wind field model is constructed according to the wind field characteristics of the first monitoring station, and then a wind speed description matrix of known simulation points in the spatial wind field model is constructed according to the average value of the second monitoring station; a spatial wind field dynamic simulation model is determined according to the spatial wind field model and the spatial wind field mechanism knowledge information. The spatial wind field dynamic simulation model can simulate the wind field according to the known wind data of each point. The description rule of the output matrix is ​​the same as the input matrix. Each matrix element position corresponds to a point in the simulation area, and the element value at the matrix element position is the wind speed and wind direction; the simulation result interpretation template is a template for interpreting the output matrix.

[0163] An attempt is made to obtain the monitoring value of the live data monitoring node of the third monitoring site. If the attempt is successful, the deviation value between the monitoring value of the corresponding live data monitoring node and the simulation value is calculated. If the deviation value is less than or equal to the preset deviation value threshold, the monitoring value of the corresponding live data monitoring node is used as the live data of the third monitoring site; if the deviation values ​​are all greater than the preset deviation value threshold, the simulation value is used as the live data of the third monitoring site, thereby improving the reliability and accuracy of obtaining the live data of the third monitoring site with fewer monitoring data sources.

[0164] The embodiment of the present invention provides a wind field downscaling simulation prediction system based on machine learning, further comprising:

[0165] Environmental early warning subsystem, used to provide environmental early warning based on 1km*1km model wind field grid forecast data;

[0166] Among them, the environmental warning subsystem performs environmental warning based on the 1km*1km model wind field grid forecast data, including:

[0167] According to the simulation domain parameters, the wind farm-associated pollution source distribution information is obtained; wherein the wind farm-associated pollution source distribution information is: the simulation domain parameters correspond to the pollution source distribution information of the actual area, such as: where there are chemical plants, where there are thermal power plants;

[0168] Based on the wind farm associated pollution source characterization template and according to the wind farm associated pollution source distribution information, the wind farm associated pollution source characteristics are extracted, wherein the wind farm associated pollution source characterization template is: a template for extracting wind farm associated pollution source characteristics against the wind farm associated pollution source distribution information, for example: an information extraction template for extracting type information of wind farm associated pollution sources from the wind farm associated pollution source distribution information; the wind farm associated pollution source characteristics include: type (for example: PM10, PM2.5), location and movement characteristic information (for example: particle velocity) of wind farm associated pollution sources;

[0169] Based on the diffusion description model, the pollution source diffusion description is determined according to the characteristics of wind farm-associated pollution sources and the 1km*1km model wind farm grid forecast data; the diffusion description model is constructed by machine learning based on the pollution diffusion law description information described by the pollution control personnel; the pollution source diffusion description is: the pollution source diffusion description semantics determined by combining the 1km*1km model wind farm grid forecast data and the wind farm-associated pollution source characteristics;

[0170] Obtain a pollution level evaluation template corresponding to the pollution source; wherein the pollution level evaluation template includes: a preset description of the pollutant type, a description of the pollution behavior, and a description of the pollutant level;

[0171] The pollution level of the pre-diffusion area is predicted based on the pollution source diffusion description and the pollution level evaluation template; when predicting the pollution level of the pre-diffusion area based on the pollution source diffusion description and the pollution level evaluation template, if the pollutant type described in the pollution source diffusion description is consistent with the pollutant type described in the pollution level evaluation template, the similarity between the pollution behavior described in the pollution source diffusion description and the pollution behavior described in the pollution level evaluation template is calculated, and the pollutant level with the greatest pollution behavior similarity is used as the predicted pollution level; the pre-diffusion area is determined based on the geographical parameters described in the pollution source diffusion description;

[0172] Obtaining the preset pollution level index of the pre-diffusion area;

[0173] If the pollution level is greater than the pollution level index, an environmental warning for the corresponding pre-diffusion area is generated.

[0174] The working principle and beneficial effects of the above technical solution are:

[0175] The present invention obtains the distribution information of wind farm-associated pollution sources, and characterizes the distribution information of wind farm-associated pollution sources to obtain the type, location and mobility characteristic information of wind farm-associated pollution sources; introduces a diffusion description model constructed by machine learning based on the pollution diffusion law description information described by pollution control personnel, and determines the pollution source diffusion description according to the characteristics of wind farm-associated pollution sources and 1km*1km model wind farm grid forecast data; then, the pollution source diffusion description is compared with the pollution level evaluation template to predict the pollution level of the pre-diffusion area, and if the pollution level is greater than the pollution level index preset in the pre-diffusion area, an environmental early warning is generated, thereby improving the early warning accuracy and timeliness.

[0176] The embodiment of the present invention provides a wind field downscaling simulation prediction method based on machine learning, such as Figure 2 As shown, including:

[0177] Step 1: Obtain 5km*5km model wind field grid forecast data;

[0178] Step 2: Get live data;

[0179] Step 3: Build a downscaling model based on machine learning technology;

[0180] Step 4: Based on the downscaling model, downscaling correction is performed according to the 5km*5km model wind field grid forecast data and the actual data, and the 1km*1km model wind field grid forecast data is output.

[0181] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A wind field downscaling simulation and prediction system based on machine learning, characterized in that: include: The first forecast data acquisition subsystem is used to obtain 5km*5km model wind field grid forecast data; A live data acquisition subsystem, used for acquiring live data; A downscaling model construction subsystem is used to construct a downscaling model based on machine learning technology; The second forecast data acquisition subsystem is used to perform downscaling correction based on the 5km*5km model wind field grid forecast data and actual data based on the downscaling model, and output 1km*1km model wind field grid forecast data.

2. A wind field downscaling simulation and prediction system based on machine learning as claimed in claim 1, characterized in that: The first forecast data acquisition subsystem acquires 5km*5km model wind field grid forecast data, including: Configure simulation domain parameters; After the simulation domain parameters are configured, run the WPS program; After the WPS program is executed, the WRF model is used to perform meteorological simulation; When the WRF model performs meteorological simulation, data assimilation is performed; After data assimilation, UPP / LRPP is run to obtain the output 5km*5km model wind field grid forecast data.

3. The wind field downscaling simulation and prediction system based on machine learning according to claim 1, characterized in that: The downscaling model construction subsystem builds a downscaling model based on machine learning technology, including: Based on the gradient boosting decision tree, a downscaling model is constructed; The input of the downscaling model is the wind field features extracted from the WRF model output. Through the wind field features, the GBDT model learns the mapping relationship between the low-resolution model output and the high-resolution wind field observation data. The basic form of the model prediction is: Among them, F(x) is the final model prediction output, M is the total number of decision trees, and h m (x) is the mth decision tree, α m is the learning rate of m decision trees, and x is the input wind field feature; Among them, the GBDT model learns the model by minimizing the loss function, and the loss function is: Among them, y i is the actual observed wind speed of the ith data sample, F(x i ) is the wind speed predicted by the model for the ith data sample, where n is the number of data samples; The GBDT model fits a new decision tree by using negative gradients. The formula for calculating the negative gradient in the mth round is as follows: After fitting the new decision tree, the overall model is updated according to the update formula, which is specifically: F m (x)=F m-1 (x)+α m h m (x) When the preset number of trees M is reached or the negative gradient converges, the GBDT model completes learning and obtains the final downscaling model.

4. The wind field downscaling simulation and prediction system based on machine learning according to claim 1, characterized in that: The live data acquisition subsystem acquires live data, including: Get the set of live data monitoring nodes; Determine a first monitoring site of a live data monitoring node; Obtaining a monitoring value of a live data monitoring node corresponding to a first monitoring site; If the number of monitoring values ​​of the live data monitoring node corresponding to the first monitoring site is greater than or equal to the preset target number, the corresponding first monitoring site is used as the second monitoring site, otherwise, the first monitoring site is used as the third monitoring site; Based on the preset outlier elimination rule, the outliers of the second monitoring site are eliminated, and the average value of the monitoring values ​​of the remaining live data monitoring nodes is calculated, and the average value is used as the live data of the second monitoring site; Calculate the simulated value of the third monitoring station based on the average value of the second monitoring station; Attempt to obtain a monitoring value of a live data monitoring node of a third monitoring site, and if the acquisition attempt is successful, calculate a deviation value between the monitoring value of the corresponding live data monitoring node and a simulated value; If the deviation value is less than or equal to the preset deviation value threshold, the monitoring value of the corresponding live data monitoring node is used as the live data of the third monitoring site; If the deviation values ​​are all greater than the preset deviation value threshold, the simulated values ​​are used as the actual data of the third monitoring site; Calculating the simulated value of the third monitoring station according to the average value of the second monitoring station includes: According to the wind field characteristics of the first monitoring station, a spatial wind field model is constructed; Construct an input matrix based on the average value of the second monitoring station; Determine the dynamic simulation model of the space wind field based on the space wind field model and the knowledge information of the space wind field mechanism; Input the input matrix into the spatial wind field dynamic simulation model to obtain the output matrix; The simulation value of the third monitoring site is determined according to the simulation result interpretation template corresponding to the output matrix.

5. The wind field downscaling simulation prediction system based on machine learning according to claim 1, characterized in that: Also includes: Environmental early warning subsystem, used to provide environmental early warning based on 1km*1km model wind field grid forecast data; Among them, the environmental warning subsystem performs environmental warning based on the 1km*1km model wind field grid forecast data, including: According to the simulation domain parameters, the distribution information of wind farm-related pollution sources is obtained; Based on the wind farm associated pollution source characterization template and according to the wind farm associated pollution source distribution information, the wind farm associated pollution source characteristics are extracted, wherein the wind farm associated pollution source characteristics include: type, location and movement characteristic information of the wind farm associated pollution source; Based on the diffusion description model, the pollution source diffusion description is determined according to the characteristics of wind field-associated pollution sources and the 1km*1km model wind field grid forecast data; Obtain the pollution level evaluation template corresponding to the pollution source; Predict the pollution level in the pre-diffusion area based on the pollution source diffusion description and pollution level evaluation template; Obtaining the preset pollution level index of the pre-diffusion area; If the pollution level is greater than the pollution level index, an environmental warning for the corresponding pre-diffusion area is generated.

6. A wind field downscaling simulation prediction method based on machine learning, characterized in that: include: Step 1: Obtain 5km*5km model wind field grid forecast data; Step 2: Get live data; Step 3: Build a downscaling model based on machine learning technology; Step 4: Based on the downscaling model, downscaling correction is performed according to the 5km*5km model wind field grid forecast data and the actual data, and the 1km*1km model wind field grid forecast data is output.

7. A wind field downscaling simulation prediction method based on machine learning as claimed in claim 6, characterized in that: Step 1: Obtain 5km*5km model wind field grid forecast data, including: Configure simulation domain parameters; After the simulation domain parameters are configured, run the WPS program; After the WPS program is executed, the WRF model is used to perform meteorological simulation; When the WRF model performs meteorological simulation, data assimilation is performed; After data assimilation, UPP / LRPP is run to obtain the output 5km*5km model wind field grid forecast data.

8. The wind field downscaling simulation prediction method based on machine learning according to claim 6, characterized in that: Step 3: Build a downscaling model based on machine learning technology, including: Based on the gradient boosting decision tree, a downscaling model is constructed; The input of the downscaling model is the wind field features extracted from the WRF model output. Through the wind field features, the GBDT model learns the mapping relationship between the low-resolution model output and the high-resolution wind field observation data. The basic form of the model prediction is: Among them, F(x) is the final model prediction output, M is the total number of decision trees, and h m (x) is the mth decision tree, α m is the learning rate of m decision trees, and x is the input wind field feature; Among them, the GBDT model learns the model by minimizing the loss function, and the loss function is: Among them, y i is the actual observed wind speed of the ith data sample, F(x i ) is the wind speed predicted by the model for the ith data sample, where n is the number of data samples; The GBDT model fits a new decision tree by using negative gradients. The formula for calculating the negative gradient in the mth round is as follows: After fitting the new decision tree, the overall model is updated according to the update formula, which is specifically: F m (x)=F m-1 (x)+α m h m (x) When the preset number of trees M is reached or the negative gradient converges, the GBDT model completes learning and obtains the final downscaling model.

9. The wind field downscaling simulation prediction method based on machine learning according to claim 6, characterized in that: Step 2: Obtain live data, including: Get the set of live data monitoring nodes; Determine a first monitoring site of a live data monitoring node; Obtaining a monitoring value of a live data monitoring node corresponding to a first monitoring site; If the number of monitoring values ​​of the live data monitoring node corresponding to the first monitoring site is greater than or equal to the preset target number, the corresponding first monitoring site is used as the second monitoring site, otherwise, the first monitoring site is used as the third monitoring site; Based on the preset outlier elimination rule, the outliers of the second monitoring site are eliminated, and the average value of the monitoring values ​​of the remaining live data monitoring nodes is calculated, and the average value is used as the live data of the second monitoring site; Calculate the simulated value of the third monitoring station based on the average value of the second monitoring station; Attempt to obtain a monitoring value of a live data monitoring node of a third monitoring site, and if the acquisition attempt is successful, calculate a deviation value between the monitoring value of the corresponding live data monitoring node and a simulated value; If the deviation value is less than or equal to the preset deviation value threshold, the monitoring value of the corresponding live data monitoring node is used as the live data of the third monitoring site; If the deviation values ​​are all greater than the preset deviation value threshold, the simulated values ​​are used as the actual data of the third monitoring site; Calculating the simulated value of the third monitoring station according to the average value of the second monitoring station includes: According to the wind field characteristics of the first monitoring station, a spatial wind field model is constructed; Construct an input matrix based on the average value of the second monitoring station; Determine the dynamic simulation model of the space wind field based on the space wind field model and the knowledge information of the space wind field mechanism; Input the input matrix into the spatial wind field dynamic simulation model to obtain the output matrix; The simulation value of the third monitoring site is determined according to the simulation result interpretation template corresponding to the output matrix.

10. The wind field downscaling simulation prediction method based on machine learning according to claim 6, characterized in that: Also includes: Environmental early warning based on 1km*1km model wind field grid forecast data; Among them, environmental warning is carried out based on the 1km*1km model wind field grid forecast data, including: According to the simulation domain parameters, the distribution information of wind farm-related pollution sources is obtained; Based on the wind farm associated pollution source characterization template and according to the wind farm associated pollution source distribution information, the wind farm associated pollution source characteristics are extracted, wherein the wind farm associated pollution source characteristics include: type, location and movement characteristic information of the wind farm associated pollution source; Based on the diffusion description model, the pollution source diffusion description is determined according to the characteristics of wind field-associated pollution sources and the 1km*1km model wind field grid forecast data; Obtain the pollution level evaluation template corresponding to the pollution source; Predict the pollution level in the pre-diffusion area based on the pollution source diffusion description and pollution level evaluation template; Obtaining the preset pollution level index of the pre-diffusion area; If the pollution level is greater than the pollution level index, an environmental warning for the corresponding pre-diffusion area is generated.

Images