CN116992222B - Method, device, equipment and medium for migration learning of wind element correction model - Google Patents

Method, device, equipment and medium for migration learning of wind element correction model Download PDF

Info

Publication number
CN116992222B
CN116992222B CN202311256975.6A CN202311256975A CN116992222B CN 116992222 B CN116992222 B CN 116992222B CN 202311256975 A CN202311256975 A CN 202311256975A CN 116992222 B CN116992222 B CN 116992222B
Authority
CN
China
Prior art keywords
correction model
data set
wind
domain correction
source domain
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202311256975.6A
Other languages
Chinese (zh)
Other versions
CN116992222A (en
Inventor
张皓
文仁强
杜梦蛟
张子良
贾天下
王浩
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Gezhouba Electric Power Rest House
China Three Gorges Corp
Original Assignee
Beijing Gezhouba Electric Power Rest House
China Three Gorges Corp
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beijing Gezhouba Electric Power Rest House, China Three Gorges Corp filed Critical Beijing Gezhouba Electric Power Rest House
Priority to CN202311256975.6A priority Critical patent/CN116992222B/en
Publication of CN116992222A publication Critical patent/CN116992222A/en
Application granted granted Critical
Publication of CN116992222B publication Critical patent/CN116992222B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/10—Pre-processing; Data cleansing
    • G06F18/15—Statistical pre-processing, e.g. techniques for normalisation or restoring missing data
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01P—MEASURING LINEAR OR ANGULAR SPEED, ACCELERATION, DECELERATION, OR SHOCK; INDICATING PRESENCE, ABSENCE, OR DIRECTION, OF MOVEMENT
    • G01P5/00—Measuring speed of fluids, e.g. of air stream; Measuring speed of bodies relative to fluids, e.g. of ship, of aircraft
    • G01P5/001—Full-field flow measurement, e.g. determining flow velocity and direction in a whole region at the same time, flow visualisation
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01W—METEOROLOGY
    • G01W1/00—Meteorology
    • G01W1/10—Devices for predicting weather conditions
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01W—METEOROLOGY
    • G01W1/00—Meteorology
    • G01W1/18—Testing or calibrating meteorological apparatus
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/096—Transfer learning

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Environmental & Geological Engineering (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Mathematical Physics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Evolutionary Biology (AREA)
  • Software Systems (AREA)
  • Environmental Sciences (AREA)
  • Atmospheric Sciences (AREA)
  • Biodiversity & Conservation Biology (AREA)
  • Ecology (AREA)
  • Probability & Statistics with Applications (AREA)
  • Aviation & Aerospace Engineering (AREA)
  • Feedback Control In General (AREA)
  • Aerodynamic Tests, Hydrodynamic Tests, Wind Tunnels, And Water Tanks (AREA)

Abstract

The invention relates to the technical field of weather, and discloses a transfer learning method, a device, equipment and a medium of a wind element correction model, wherein the method comprises the following steps: acquiring a first data set at a preset point position and a second data set at a point position to be corrected in an offshore wind farm, and establishing a wind element source domain correction model through a self-adaptive momentum random optimization algorithm; calculating a judgment coefficient; and based on the second data set and the judging coefficient, performing migration learning on the wind element source domain correction model, generating a wind element target domain correction model, and correcting to obtain a correction result. According to the method, the trained wind element source domain correction models at other preset point positions are migrated, so that the problem of pattern correction of the point position lacking in data materials is solved; the calculation innovation of the judgment coefficient is used in the transfer learning process of the wind element correction model, and the use scene of the index is further expanded.

Description

Method, device, equipment and medium for migration learning of wind element correction model
Technical Field
The invention relates to the technical field of weather, in particular to a method, a device, equipment and a medium for migration learning of a wind element correction model.
Background
The existing weather forecast data is usually simulated based on a weather forecast mode (WRF mode), but the simulation result is usually different from the actual wind measurement data to some extent. In order to improve the accuracy of the data, a certain correction strategy is generally required to correct the weather forecast simulation data.
At present, a common correction strategy is generally based on a machine learning model, and the model has the characteristics of high fitting precision, strong learning capacity and the like. However, machine learning models such as deep learning require a large number of data samples, and insufficient sample size is likely to cause problems such as insufficient learning and over-fitting, resulting in poor model applicability, so that it is difficult to construct a proper correction model at points where data is lacking.
At present, aiming at the difficult problem of wind element correction in the data-missing material point position weather forecast mode, a mode of directly correcting by using a correction model of adjacent point positions is generally adopted, and although a certain correction effect can be obtained, the correction effect is limited.
Disclosure of Invention
In view of the above, the invention provides a method, a device, equipment and a medium for migration learning of a wind element correction model, so as to solve the problem that the existing correction model by using adjacent points has poor correction effect on wind elements in a data-missing point weather forecast mode.
In a first aspect, the present invention provides a method for learning by migration of a wind element correction model, the method for learning by migration of a wind element correction model comprising:
acquiring a first data set at a preset point position and a second data set at a point position to be corrected in the offshore wind farm when the offshore wind farm is in a preset weather forecast mode, wherein the first data set comprises a first weather forecast mode data set and a first actually measured wind speed data set, and the second data set comprises a second weather forecast mode data set and a second actually measured wind speed data set; based on the first data set, establishing a wind element source domain correction model through a self-adaptive momentum random optimization algorithm; calculating a judgment coefficient of the first measured wind speed data set and the second measured wind speed data set; based on the second data set and the judging coefficient, performing migration learning on the wind element source domain correction model to generate a wind element target domain correction model; and correcting the second weather forecast mode data set at the point to be corrected by using the wind element target domain correction model to obtain a correction result.
According to the migration learning method of the wind element correction model, the trained wind element source domain correction model at other preset point positions is migrated by adopting a certain strategy, so that the problem of pattern correction of the point position lacking in data materials is solved; the calculation innovation of the judgment coefficient is used in the transfer learning process of the wind element correction model, and the use scene of the index is further expanded.
In an alternative embodiment, based on the first data set, an adaptive momentum random optimization algorithm is used to build a wind factor source domain correction model, including:
acquiring a first neural network parameter set of an initial wind element source domain correction model; based on the first data set and the first neural network parameter set, training the initial wind element source domain correction model by utilizing a self-adaptive momentum random optimization algorithm until the wind element source domain correction model meeting a preset first condition is obtained.
According to the invention, the model training is performed by using the self-adaptive momentum random optimization algorithm, so that the accuracy of the model training is improved.
In an alternative embodiment, training the initial wind element source domain correction model using an adaptive momentum random optimization algorithm based on the first data set and the first neural network parameter set until a wind element source domain correction model satisfying a preset first condition is obtained, including:
dividing the first data set into a first training data set and a first test data set according to a time sequence; dividing the first training data set into at least one batch of training data according to the batch size order; training an initial wind element source domain correction model by utilizing a self-adaptive momentum random optimization algorithm based on each batch of training data and a first neural network parameter set; judging whether the trained initial wind element source domain correction model meets a preset first condition or not by using a first test data set; and determining a wind element source domain correction model according to the judging result.
According to the invention, the model training is performed by using the self-adaptive momentum random optimization algorithm, so that the accuracy of the model training is improved.
In an alternative embodiment, determining the wind factor source domain correction model according to the determination result includes:
when the trained initial wind element source domain correction model meets a preset first condition, determining the trained initial wind element source domain correction model as a wind element source domain correction model; when the trained initial wind element source domain correction model does not meet a preset first condition, adjusting a first neural network parameter set based on each batch of training data; based on the training data of each batch and the adjusted first neural network parameter set, training the initial wind element source domain correction model by utilizing a self-adaptive momentum random optimization algorithm until the wind element source domain correction model meeting the preset first condition is obtained.
According to the invention, model parameters are adjusted in real time according to the training result in the training process, so that the model training precision is improved.
In an alternative embodiment, calculating the decision coefficients of the first and second measured wind speed data sets comprises:
calculating a first measured data average value of a first measured wind speed data set in the first data set; calculating a second measured data average value of a second measured wind speed data set in the second data set; a determination coefficient is calculated based on the first measured data average value and the second measured data average value.
The method is used for creatively calculating the judgment coefficients in the transfer learning process of the wind element correction model, and further expands the use scene of the index.
In an alternative embodiment, based on the second data set and the decision coefficient, performing migration learning on the wind element source domain correction model to generate a wind element target domain correction model, including:
acquiring a preset first threshold value and a preset second threshold value; respectively comparing the judgment coefficient with a preset first threshold value and a preset second threshold value; and performing migration learning on the wind element source domain correction model according to the comparison result and the second data set to generate a wind element target domain correction model.
According to the invention, different migration learning strategies are set according to different values of the judgment coefficients, so that the training cost of the wind element target domain correction model can be reduced as much as possible, and the training effect is ensured.
In an alternative embodiment, the performing migration learning on the wind element source domain correction model according to the comparison result and the second data set to generate a wind element target domain correction model includes:
when the judgment coefficient is larger than a preset second threshold value, acquiring a second neural network parameter set of the wind element source domain correction model; based on the second neural network parameter set and the second data set, training the wind element source domain correction model by using a self-adaptive momentum random optimization algorithm until the wind element target domain correction model meeting the preset second condition is obtained.
According to the invention, different migration learning strategies are set according to different values of the judgment coefficients, so that the training cost of the wind element target domain correction model can be reduced as much as possible, and the training effect is ensured.
In an alternative embodiment, the method further includes performing migration learning on the wind element source domain correction model according to the comparison result and the second data set to generate a wind element target domain correction model, and further includes:
when the judgment coefficient is larger than a preset first threshold value and smaller than a preset second threshold value, acquiring an output layer weight parameter and other neural network layer weight parameters in a second neural network parameter set; initializing the output layer weight parameters to obtain target output layer weight parameters; determining a third neural network parameter set based on the target output layer weight parameter and other neural network layer weight parameters; based on the third neural network parameter set and the second data set, training the wind element source domain correction model by using a self-adaptive momentum random optimization algorithm until the wind element target domain correction model meeting the preset second condition is obtained.
According to the invention, different migration learning strategies are set according to different values of the judgment coefficients, so that the training cost of the wind element target domain correction model can be reduced as much as possible, and the training effect is ensured.
In an alternative embodiment, the method further includes performing migration learning on the wind element source domain correction model according to the comparison result and the second data set to generate a wind element target domain correction model, and further includes:
initializing weight parameters of each neural network layer contained in the second neural network parameter set when the judgment coefficient is smaller than a preset first threshold value to obtain a fourth neural network parameter set; based on the fourth neural network parameter set and the second data set, training the wind element source domain correction model by using a self-adaptive momentum random optimization algorithm until the wind element target domain correction model meeting the preset second condition is obtained.
According to the invention, different migration learning strategies are set according to different values of the judgment coefficients, so that the training cost of the wind element target domain correction model can be reduced as much as possible, and the training effect is ensured.
In an alternative embodiment, the method further comprises: inputting the second weather forecast mode data set into a wind element target domain correction model to obtain weather forecast mode target domain correction data; calculating an error index based on the weather forecast mode target domain correction data; and evaluating the correction result based on the error index.
The invention evaluates the correction result by using the error index, and improves the accuracy of the correction result.
In a second aspect, the present invention provides a transfer learning apparatus for a wind element correction model, the transfer learning apparatus comprising:
the system comprises an acquisition module, a correction module and a correction module, wherein the acquisition module is used for acquiring a first data set at a preset point position and a second data set at a point position to be corrected in the offshore wind farm when the offshore wind farm is in a preset weather forecast mode, the first data set comprises a first weather forecast mode data set and a first actually measured wind speed data set, and the second data set comprises a second weather forecast mode data set and a second actually measured wind speed data set; the building module is used for building a wind element source domain correction model based on the first data set through a self-adaptive momentum random optimization algorithm; the calculation module is used for calculating the judgment coefficients of the first measured wind speed data set and the second measured wind speed data set; and the migration learning module is used for performing migration learning on the wind element source domain correction model based on the second data set and the judgment coefficient to generate a wind element target domain correction model. And the correction module is used for correcting the second weather forecast mode data set at the position of the point to be corrected by using the wind element target domain correction model to obtain a correction result.
In an alternative embodiment, the establishing module includes:
The first acquisition sub-module is used for acquiring a first neural network parameter set of the initial wind element source domain correction model; the first training sub-module is used for training the initial wind element source domain correction model by utilizing the self-adaptive momentum random optimization algorithm based on the first data set and the first neural network parameter set until the wind element source domain correction model meeting the preset first condition is obtained.
In an alternative embodiment, a first training sub-module includes:
a first dividing unit for dividing the first data set into a first training data set and a first test data set in time sequence; the second dividing unit is used for dividing the first training data set into at least one batch of training data according to the batch size sequence; the first training unit is used for training the initial wind element source domain correction model by utilizing a self-adaptive momentum random optimization algorithm based on each batch of training data and the first neural network parameter set; the judging unit is used for judging whether the trained initial wind element source domain correction model meets a preset first condition or not by utilizing the first test data set; and the first determining unit is used for determining the wind factor source domain correction model according to the judging result.
In a third aspect, the present invention provides a computer device comprising: the wind element correction model migration learning method according to the first aspect or any one of the embodiments corresponding to the first aspect is executed by the processor.
In a fourth aspect, the present invention provides a computer-readable storage medium having stored thereon computer instructions for causing a computer to execute the method for learning and transferring a wind element correction model according to the first aspect or any one of the embodiments corresponding thereto.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings that are needed in the description of the embodiments or the prior art will be briefly described, and it is obvious that the drawings in the description below are some embodiments of the present invention, and other drawings can be obtained according to the drawings without inventive effort for a person skilled in the art.
FIG. 1 is a flow chart of a method for transition learning of a wind element correction model according to an embodiment of the present invention;
FIG. 2 is a flow chart of another method for transition learning of a wind element correction model according to an embodiment of the present invention;
FIG. 3 is a flow chart of a method of transition learning of a further wind element correction model according to an embodiment of the present invention;
FIG. 4 is a flow chart of a method for transition learning of a further wind element correction model according to an embodiment of the present invention;
FIG. 5 is a schematic diagram of a WRF correction model migration strategy according to an embodiment of the present invention;
FIG. 6 is a schematic diagram of a stacked self-encoder model according to an embodiment of the present invention;
FIG. 7 is a schematic diagram of a source domain correction effect of a WRF mode correction model according to an embodiment of the present invention;
FIG. 8 is a schematic diagram of the effect of target domain correction of a WRF mode correction model in accordance with an embodiment of the present invention;
FIG. 9 is a block diagram of a wind element correction model transfer learning apparatus according to an embodiment of the present invention;
fig. 10 is a schematic diagram of a hardware structure of a computer device according to an embodiment of the present invention.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are some embodiments of the present invention, but not all embodiments of the present invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
Existing wind factor correction for WRF modes is usually modeled based on a machine learning framework, and in order to ensure a certain correction effect, the machine learning model needs a sufficient amount of actual measurement samples for model training. When correcting the point position with relatively lacking measured sample data, if the machine learning is directly adopted for modeling, the network performance is often difficult to reach the optimal value and even training is difficult to finish. At present, a method is generally adopted in which correction is performed by directly using a correction model of a neighboring point, and although a certain correction effect can be obtained, the correction effect is limited.
According to an embodiment of the present invention, there is provided an embodiment of a method for learning the migration of wind element correction models, it should be noted that the steps illustrated in the flowcharts of the drawings may be performed in a computer system such as a set of computer executable instructions, and that although a logical order is illustrated in the flowcharts, in some cases, the steps illustrated or described may be performed in an order different from that herein.
In this embodiment, a method for learning to migrate a wind element correction model is provided, fig. 1 is a flowchart of a method for learning to migrate a wind element correction model according to an embodiment of the present invention, as shown in fig. 1, where the flowchart includes the following steps:
Step S101, acquiring a first data set at a preset point position and a second data set at a point position to be corrected in the offshore wind farm when the offshore wind farm is in a preset weather forecast mode.
The preset weather forecast mode is a WRF mode, represents a new generation of non-static balanced high-resolution mesoscale weather research and forecast numerical mode, and mainly considers the forecast of important weather from cloud scale to weather scale.
The first data set comprises a first weather forecast mode data set and a first measured wind speed data set which are simulated at preset points in the offshore wind farm in a WRF mode.
The second data set comprises a second weather forecast mode data set and a second actually measured wind speed data set which are simulated at the point to be corrected in the offshore wind farm when the WRF mode is adopted.
Further, a first weather forecast pattern datasetThe wind speed, the wind direction and other meteorological element data of each height layer of the offshore wind farm at the preset point position can be obtained; second weather forecast pattern dataset +.>The wind speed, the wind direction and other meteorological element data of each height layer of the offshore wind farm at the position of the point to be corrected can be obtained.
First measured wind speed datasetThe actual wind speed data of the offshore wind farm at the preset point position can be obtained from a wind measuring tower, laser radar wind measuring equipment and the like.
Second measured wind speed datasetThe actual wind speed data of the offshore wind farm at the point to be corrected can be obtained.
Further, the two types of data (weather forecast mode data and wind speed data) may also include other variable types: wind speed (altitude layer typically 100, 90, 80, 70 meters, or other altitude layer), value range typically 0-50m/s; wind direction (height layers are typically 100, 90, 80, 70 meters, or other height layers) with values ranging from 0-360 degrees; temperature, humidity, air pressure, etc.
Step S102, based on the first data set, a wind element source domain correction model is established through a self-adaptive momentum random optimization algorithm.
Wherein the adaptive momentum random optimization algorithm represents a deep learning optimizer algorithm for solving the optimization problem of random targets in a high-dimensional parameter space.
Specifically, the model training is performed by using the self-adaptive momentum random optimization algorithm, so that the accuracy of the established wind element source domain correction model can be improved.
Step S103, calculating a judging coefficient of the first measured wind speed data set and the second measured wind speed data set.
In this embodiment, the correlation between the first measured wind speed data set and the second measured wind speed data set is reflected by the decision coefficient, and the larger the decision coefficient is, the higher the correlation is.
Specifically, in a typical machine learning modeling process, a decision coefficient is generally used to measure the correlation between the actual output of the model and the target output, i.e. to determine to what extent the learning effect of the model is achieved.
In the embodiment, the calculation innovation of the judgment coefficient is used in the transfer learning process of the wind element correction model, so that the use scene of the index is further expanded.
And step S104, performing migration learning on the wind element source domain correction model based on the second data set and the judgment coefficient to generate a wind element target domain correction model.
Specifically, different migration learning strategies can be formed according to the value of the judgment coefficient.
Further, according to different formed transfer learning strategies, transfer learning is carried out on the wind element source domain correction model, and a wind element target domain correction model after transfer learning can be generated.
And step S105, correcting the second weather forecast mode data set at the point to be corrected by using the wind element target domain correction model to obtain a correction result.
Specifically, the second weather forecast mode data set at the point to be corrected can be corrected according to the wind element target domain correction model after transfer learning, and a corresponding correction result is obtained.
According to the migration learning method of the wind element correction model, the trained wind element source domain correction model at other preset points is migrated by adopting a certain strategy, so that the problem of pattern correction of the point lacking in data materials is solved; the calculation innovation of the judgment coefficient is used in the transfer learning process of the wind element correction model, and the use scene of the index is further expanded.
In this embodiment, a method for learning the wind element correction model by migration is provided, fig. 2 is a flowchart of a method for learning the wind element correction model by migration according to an embodiment of the present invention, as shown in fig. 2, and the flowchart includes the following steps:
step S201, acquiring a first data set at a preset point position and a second data set at a point position to be corrected in the offshore wind farm when the offshore wind farm is in a preset weather forecast mode. Please refer to step S101 in the embodiment shown in fig. 1 in detail, which is not described herein.
Step S202, based on the first data set, a wind element source domain correction model is established through a self-adaptive momentum random optimization algorithm. Please refer to step S102 in the embodiment shown in fig. 1 in detail, which is not described herein.
Step S203, calculating a determination coefficient of the first measured wind speed data set and the second measured wind speed data set.
Specifically, the step S203 includes:
step S2031, calculating a first measured data average value of a first measured wind speed dataset in the first dataset.
Specifically, a first measured wind speed data set in the first data set is calculatedMean value of stroke speed data, i.e. first measured data mean +.>。
Step S2032, calculating a second measured data average value of a second measured wind speed dataset in the second dataset.
Specifically, a second measured wind speed dataset in a second dataset is calculatedMean value of stroke speed data, second measured data mean +.>。
In step S2033, a determination coefficient is calculated based on the first measured data average value and the second measured data average value.
Specifically, the determination coefficient is calculated by the following relational expression (1)：
（1）
Wherein:representing the number of samplesAmount of the components.
Further, the determination coefficientTypically the value ranges are [0,1]Is generally regarded as->Two sets of data greater than 0.8 are highly correlated.
And step S204, performing migration learning on the wind element source domain correction model based on the second data set and the judgment coefficient to generate a wind element target domain correction model. Please refer to step S104 in the embodiment shown in fig. 1 in detail, which is not described herein.
And step S205, correcting the second weather forecast mode data set at the point to be corrected by using the wind element target domain correction model to obtain a correction result. Please refer to step S105 in the embodiment shown in fig. 1 in detail, which is not described herein.
According to the migration learning method of the wind element correction model, a data set containing weather forecast mode data and measured wind speed data is obtained, and data support is provided for migration of a follow-up wind element source domain correction model; migrating the trained wind element source domain correction model at other preset point positions by adopting a certain strategy, so as to solve the problem of pattern correction of the point position lacking in data materials; the calculation innovation of the judgment coefficient is used in the transfer learning process of the wind element correction model, and the use scene of the index is further expanded.
In this embodiment, a method for learning the wind element correction model by migration is provided, fig. 3 is a flowchart of a method for learning the wind element correction model by migration according to an embodiment of the present invention, as shown in fig. 3, and the flowchart includes the following steps:
step S301, acquiring a first data set at a preset point position and a second data set at a point position to be corrected in the offshore wind farm when the offshore wind farm is in a preset weather forecast mode. Please refer to step S101 in the embodiment shown in fig. 1 in detail, which is not described herein.
Step S302, a wind element source domain correction model is established through a self-adaptive momentum random optimization algorithm based on the first data set.
Specifically, the step S302 includes:
step S3021, obtaining a first neural network parameter set of the initial wind element source domain correction model.
Wherein the first neural network parameter set may include a number of neural network layers of the initial wind element source domain correction modellAnd associated hyper-parameters within layers, e.g. weightsWBias ofbEtc. The neural network may be a long-term memory network, a stacked self-encoder, a deep belief network, or other machine learning network.
Step S3022, training the initial wind element source domain correction model by using the adaptive momentum random optimization algorithm based on the first data set and the first neural network parameter set until the wind element source domain correction model satisfying the preset first condition is obtained.
Specifically, the first weather forecast pattern data set in the first data setDetermining the first measured wind speed data set as input of a wind factor source domain correction model +.>And determining the output of the wind element source domain correction model.
Further, based on the initial first neural network parameter set, on the basis of the input and output of the model, training the initial wind element source domain correction model by utilizing a self-adaptive momentum random optimization algorithm until the accuracy of the trained model reaches the requirement or the training frequency reaches the maximum iteration frequency, and completing the training to obtain the corresponding wind element source domain correction model.
In some alternative embodiments, step S3022 includes:
step a1, dividing the first data set into a first training data set and a first test data set according to time sequence.
Step a2, dividing the first training data set into at least one batch of training data according to the batch size order.
And a3, training an initial wind element source domain correction model by using a self-adaptive momentum random optimization algorithm based on each batch of training data and the first neural network parameter set.
And a step a4 of judging whether the trained initial wind element source domain correction model meets a preset first condition by using the first test data set.
And a step a5, determining a wind factor source domain correction model according to the judging result.
Wherein the first training data set is the first 80% of the data in the first data set; the first test dataset was 20% of the data after the first dataset.
Specifically, the first training data set is divided into a plurality of batches according to the batch size, and each batch of training data is sequentially input into an initial wind element source domain correction model and is trained by utilizing an adaptive momentum random optimization algorithm.
After all training data in the first training data set are traversed once, a first test data set is adopted to carry out model test, whether the accuracy of the initial wind element source domain correction model after training reaches a set requirement or whether the training times corresponding to the initial wind element source domain correction model after training reaches the maximum iteration times is judged, and the final wind element source domain correction model is determined according to the judging result.
In some alternative embodiments, step a5 includes:
and a step a51 of determining the trained initial wind element source domain correction model as a wind element source domain correction model when the trained initial wind element source domain correction model meets a preset first condition.
And a step a52, when the trained initial wind element source domain correction model does not meet the preset first condition, adjusting the first neural network parameter set based on each batch of training data.
And a step a53 of training the initial wind element source domain correction model by utilizing a self-adaptive momentum random optimization algorithm based on each batch of training data and the adjusted first neural network parameter set until the wind element source domain correction model meeting the preset first condition is obtained.
Specifically, when the accuracy of the trained initial wind element source domain correction model meets the requirement or the training frequency corresponding to the trained initial wind element source domain correction model is less than or equal to the maximum iteration frequency, training is completed, and the trained initial wind element source domain correction model is determined to be a final wind element source domain correction model.
When the accuracy of the initial wind element source domain correction model after training does not meet the requirement or the training frequency corresponding to the initial wind element source domain correction model after training is larger than the maximum iteration frequency, traversing each batch of training data in the first training data set again, adjusting the first neural network parameter set, then continuing training the initial wind element source domain correction model by utilizing the self-adaptive momentum random optimization algorithm based on the adjusted first neural network parameter set and each batch of data in the first training data set until the accuracy of the model after training reaches the requirement or the training frequency reaches the maximum iteration frequency, and completing training and obtaining the corresponding wind element source domain correction model.
Step S303, calculating a judgment coefficient of the first measured wind speed data set and the second measured wind speed data set. Please refer to step S203 in the embodiment shown in fig. 2 in detail, which is not described herein.
And step S304, performing migration learning on the wind element source domain correction model based on the second data set and the judgment coefficient to generate a wind element target domain correction model. Please refer to step S104 in the embodiment shown in fig. 1 in detail, which is not described herein.
And step S305, correcting the second weather forecast mode data set at the point to be corrected by using the wind element target domain correction model to obtain a correction result. Please refer to step S105 in the embodiment shown in fig. 1 in detail, which is not described herein.
Step S306, the second weather forecast mode data set is input into the wind element target domain correction model, and weather forecast mode target domain correction data are obtained.
Specifically, the second weather forecast mode data set at the point to be corrected is input into the wind element target domain correction model generated after transfer learning, so that corrected weather forecast mode target domain correction data can be obtained.
Step S307 calculates an error index based on the weather forecast mode target domain correction data.
Specifically, error indicators are calculated using the following relational expressions (2) and (3), respectively And->：
（2）
（3）
Wherein:representing the average absolute error; />Representing root mean square error; />Representing the number of samples;and representing model output, namely weather forecast mode target domain correction data.
Step S308, the correction result is evaluated based on the error index.
Specifically, according to the calculated error indexAnd->Can correct the obtained correction resultAn evaluation is performed.
According to the transfer learning method of the wind element correction model, model training is carried out by utilizing the self-adaptive momentum random optimization algorithm, model parameters are adjusted in real time according to training results in the training process, the model training accuracy is improved, further, the second weather forecast mode data set at the position to be corrected of the wind element target domain correction model obtained through training is corrected, the correction results are obtained, the problem of mode correction of the position to be corrected of the lack of data materials is solved, meanwhile, calculation innovation of the judgment coefficients is used in the transfer learning process of the wind element correction model, and the use scene of the index is further expanded.
In this embodiment, a method for learning the wind element correction model by migration is provided, fig. 4 is a flowchart of a method for learning the wind element correction model by migration according to an embodiment of the present invention, as shown in fig. 4, and the flowchart includes the following steps:
Step S401, acquiring a first data set at a preset point position and a second data set at a point position to be corrected in the offshore wind farm when the offshore wind farm is in a preset weather forecast mode. Please refer to step S101 in the embodiment shown in fig. 1 in detail, which is not described herein.
Step S402, based on the first data set, a wind element source domain correction model is established through a self-adaptive momentum random optimization algorithm. Please refer to step S302 in the embodiment shown in fig. 3 in detail, which is not described herein.
Step S403, calculating a determination coefficient of the first measured wind speed data set and the second measured wind speed data set. Please refer to step S203 in the embodiment shown in fig. 2 in detail, which is not described herein.
And step S404, performing migration learning on the wind element source domain correction model based on the second data set and the judgment coefficient to generate a wind element target domain correction model.
Specifically, the step S404 includes:
in step S4041, a preset first threshold and a preset second threshold are obtained.
Wherein the preset first threshold is 0.5; the second threshold is preset to 0.8.
In step S4042, the determination coefficients are compared with the preset first threshold and the preset second threshold, respectively.
Specifically, the determination coefficients are respectively Comparison was made with 0.5 and 0.8.
And step S4043, performing migration learning on the wind element source domain correction model according to the comparison result and the second data set to generate a wind element target domain correction model.
Specifically, according to the determination coefficientDifferent transfer learning strategies can be formed by comparing the values of the wind element source domain correction models, and further, the second data set at the point to be corrected is combined, the different transfer learning strategies are utilized to transfer learn the wind element source domain correction models, and corresponding wind element target domain correction models are generated.
In some alternative embodiments, step S4043 includes:
and b1, when the judgment coefficient is larger than a preset second threshold value, acquiring a second neural network parameter set of the wind element source domain correction model.
And b2, training the wind element source domain correction model by utilizing a self-adaptive momentum random optimization algorithm based on the second neural network parameter set and the second data set until the wind element target domain correction model meeting the preset second condition is obtained.
In particular, whenAnd when the wind speed (wind direction) at the position of the point to be corrected is consistent with the change rule of the wind speed (wind direction) at the preset point, the correlation between the first data set and the second data set is higher.
Therefore, the wind element target domain correction model can directly follow the structure and all weight parameters of the wind element source domain correction model, namely the second neural network parameter set.
Further, based on the second neural network parameter set and the second data set, training the wind element source domain correction model by utilizing a self-adaptive momentum random optimization algorithm until the accuracy of the trained model reaches the requirement or the training frequency reaches the maximum iteration frequency, and completing training to obtain a corresponding wind element target domain correction model.
In some optional embodiments, the step S4043 further includes:
and c1, when the judgment coefficient is larger than a preset first threshold value and smaller than a preset second threshold value, acquiring an output layer weight parameter and other neural network layer weight parameters in the second neural network parameter set.
And c2, initializing the output layer weight parameters to obtain target output layer weight parameters.
And c3, determining a third neural network parameter set based on the target output layer weight parameter and other neural network layer weight parameters.
And c4, training the wind element source domain correction model by utilizing a self-adaptive momentum random optimization algorithm based on the third neural network parameter set and the second data set until the wind element target domain correction model meeting the preset second condition is obtained.
In particular, whenAnd when the wind speed (wind direction) at the position of the point to be corrected and the wind speed (wind direction) at the preset point position show that the first data set and the second data set have certain correlation, namely the change of the wind speed (wind direction) at the position of the point to be corrected is consistent in part of characteristics.
Therefore, the wind element target domain correction model uses the structure of the wind element source domain correction model, and the weight parameters of the output layer of the wind element source domain correction model are reinitialized, and the weight parameters of other neural network layers are kept unchanged.
Specifically, when a threshold value is smaller than a preset second threshold value, acquiring an output layer weight parameter, and initializing the output layer weight parameter to obtain a target output layer weight parameter.
Further, according to the other neural network layer weight parameters, the initialized target output layer weight parameters and the third neural network parameter set which can be corresponding to the initialized target output layer weight parameters, the third neural network parameter set can be determined.
And finally, training the wind element source domain correction model by utilizing a self-adaptive momentum random optimization algorithm based on the third neural network parameter set and the second data set until the accuracy of the trained model reaches the requirement or the training frequency reaches the maximum iteration frequency, and completing training to obtain a corresponding wind element target domain correction model.
In some optional embodiments, the step S4043 further includes:
and d1, initializing weight parameters of each neural network layer contained in the second neural network parameter set when the judgment coefficient is smaller than a preset first threshold value, and obtaining a fourth neural network parameter set.
And d2, training the wind element source domain correction model by utilizing a self-adaptive momentum random optimization algorithm based on the fourth neural network parameter set and the second data set until the wind element target domain correction model meeting the preset second condition is obtained.
In particular, whenWhen a correlation between the first data set and the second data set is indicated, little or no apparent correlation.
Therefore, the wind element target domain correction model uses the structure of the wind element source domain correction model, and initializes all the neural network layer weight parameters of the wind element source domain correction model again, so that a fourth initialized neural network parameter set can be obtained.
And finally, training the wind element source domain correction model by utilizing a self-adaptive momentum random optimization algorithm based on the fourth neural network parameter set and the initialized fourth neural network parameter set until the accuracy of the trained model reaches the requirement or the training frequency reaches the maximum iteration frequency, and completing the training to obtain a corresponding wind element target domain correction model.
And step S405, correcting the second weather forecast mode data set at the point to be corrected by using the wind element target domain correction model to obtain a correction result. Please refer to step S105 in the embodiment shown in fig. 1 in detail, which is not described herein.
According to the migration learning method of the wind element correction model, the trained wind element source domain correction model at other preset points is migrated by adopting a certain strategy, so that the problem of pattern correction of the point lacking in data materials is solved; the calculation innovation of the judgment coefficient is used in the transfer learning process of the wind element correction model, namely, different transfer learning strategies are set according to different values of the judgment coefficient, so that the use scene of the index is further expanded, the training cost of the wind element target domain correction model can be reduced as much as possible, and the training effect is ensured.
In one example, a migration learning strategy for a WRF mode wind element correction model is provided, comprising:
1. the WRF mode wind element data simulated at the specific point is obtained, and the WRF mode wind element data comprises variables such as wind speed, wind direction, other meteorological elements (precipitation, radiation, temperature and humidity and the like) of each height layer and actual wind speed data (which can be from a wind measuring tower, laser radar wind measuring equipment and the like) at the same point are determined to be a model source domain data set.
2. Selecting each related variable (mode simulation wind direction, wind speed, temperature and humidity and the like) in the source domain data set to be determined as an input data set of a wind speed correction modelOutput data set determined as correction model based on measured wind speed data +.>The method comprises the steps of carrying out a first treatment on the surface of the The source domain correction model can select a long-term memory network, a short-term memory network, a stacked self-encoder, a deep confidence network and other machine learning frameworks.
3. Training of source domain correction models based on machine learning framework:
(1) Determining the number of layers of a neural networklAnd associated hyper-parameters such as weights within each layerWBias ofbEtc.;
(2) Dividing an input data set and an output data set into a training set and a testing set according to a time sequence, wherein the training set is the first 80% of data, and the testing set is the last 20% of data;
(3) Dividing a training set into a plurality of batches according to the batch size, sequentially inputting each batch into a source domain correction model, and training by adopting a self-adaptive momentum random optimization algorithm; after all training data are traversed once, a test set is adopted to carry out model test, and after the precision setting requirement or the maximum iteration number is reached, model training is completed; otherwise, the training data is traversed again to continue model parameter adjustment; and performing iterative training for multiple times until the model precision reaches the set requirement or the maximum iterative times, and completing the network training.
4. Setting the trained source domain correction model as a model to be migrated, and keeping the model structure and various related parameters unchanged.
5. Obtaining simulated WRF mode wind element data at the position of the point to be corrected and determining the simulated WRF mode wind element data as a target domain input data setAnd the actual variables to be corrected (wind speed, wind direction and other factors) at the point are taken as the output data set of the target domain>And dividing the target domain input/output data set into a target domain training set and a target domain testing set according to the time sequence, wherein the training set is the first 80% of data, and the testing set is the last 20% of data.
6. Performing data correlation analysis on the actual data to be corrected in the source domain data set and the actual variables to be corrected in the target domain data set, wherein the correlation degree adopts a judgment coefficientMeasuring, wherein->Reference is made to the above relation (1). />
7. Computing source domain output data setsOutput data set with target domain->Decision coefficient between two sets of data ∈ ->And according to->The values are respectively set according to the following standard to set a model migration strategy, as shown in fig. 5:
(1) When (when)And when the method shows that the correlation degree of the source domain data and the target domain data is higher, the change rule of the wind speeds (wind directions) at two positions can be embodied to be more consistent. In such cases, the target domain initial correction model directly follows the structure and all weight parameters of the source domain model. Fine-tuning model parameters by adopting a target domain data set, wherein the fine-tuning process of the target domain model refers to a 3 (2-3) part to form a target domain correction model;
(2) When (when)And when the method shows that the source domain data and the target domain data have a certain degree of correlation, the method can be embodied in that the wind speed (wind direction) changes at two positions are consistent in part of characteristics. Under such a situation, the initial target domain correction model uses the source domain model structure to reinitialize the weight parameters of the output layer of the source domain model and keep the weight parameters of the rest layers unchanged. Training the adjusted model by adopting a target domain data set, and forming a target domain correction model by referring to a 3 (2-3) part in a fine tuning process of the target domain model;
(3) When (when)And when the source domain data and the target domain data are related to a small degree or no obvious correlation is shown. Order of (A)The standard domain initial correction model uses the source domain model structure to initialize all network parameters again. Training the model after the adjustment by adopting the target domain data set, and forming a target domain correction model by referring to a 3 (1-3) part in the fine adjustment process of the target domain model;
8. and performing performance test on the trained target domain correction model by adopting test data, and indicating that model migration reaches the set requirement after meeting the requirement.
Further, there is provided a specific embodiment based on specific experimental data, comprising:
(1) Obtaining WRF mode simulation data of a sea area, and selecting 2 fixed point positions in the area, wherein the point position 1 is required to be provided with enough actual measurement data samples for training a WRF correction model, and the point position 2 is required to be provided with quantitative data correspondingly so as to support the implementation of the migration learning strategy provided by the embodiment of the invention.
(2) The acquired data set is further partitioned. Wherein the source domain input data set corresponds to the WRF simulated wind speed data of 100 m height at the point position 1 for 10 minutesThe source domain outputs the actual wind measurement data of 100 meters height at the position 1 corresponding to the data set for 10 minutes>（/>、/>The time period is 2018, 1 month, 1 day to 12 months, 31 days); WRF simulation wind speed data +.10 minutes for 100 m height at point location 2 corresponding to source domain input dataset>The source domain outputs the actual wind measurement data of 100 meters height at the position 2 corresponding to the data set for 10 minutes>（/>、/>The time period is from 1 day of 12 months to 31 days of 12 months in 2018).
(3) The source domain correction model selects the stacked self-encoder model, as shown in FIG. 6, with the network inputs set toWherein->Represents->Simulating wind speed at moment; the output is +.>Real test data corresponding to time is +.>. The network comprises 2 hidden layers, the number of neurons in each layer is 60, and the weight and bias in each layer are randomly initialized by adopting zero-mean Gaussian distribution with variance of 0.01.
(4) The training of the source domain correction stacking self-encoder model adopts a self-adaptive momentum random optimization algorithm, the training is completed after the network model precision reaches the set index, and the tested effect is shown in fig. 7 (in order to ensure a better display effect, only 1000 samples of effect are shown in the figure). And (5) reserving the weight and bias obtained after the model structure and training.
(5) Computing source domain output data setsOutput data set with target domain->Is->Because the time lengths of the two data sets are inconsistent, the source domain data sets need to be intercepted before calculation so as to be unified to the same length. Calculated +.>The value is 0.682.
(6) According toThe size 0.682, and the model migration policy shown in step 7 above, 6. Select the policy applicable to this embodiment, namely class 2: the target domain initial correction model uses a source domain model structure, re-initializes the weight parameters of the output layer of the source domain model, keeps the weight parameters of the rest layers unchanged, and determines the weight parameters as a target domain WRF correction initialization model.
(7) The training of the target domain WRF correction model also adopts a self-adaptive momentum random optimization algorithm, the training is completed after the network model precision reaches the set index, and final target domain WRF mode corrected data can be obtained, and the test effect is as shown in fig. 8 (the target domain data volume is less, and the test only adopts 1000 sample data).
(8) In order to further illustrate the effectiveness of the present invention, the error index shown in the above-mentioned relations (2) and (3) is selected to evaluate the correction effect of the final target domain, and the evaluation result is as follows: before the target domain correction:--2.387；-2.978, after target domain correction: />--1.617；/>--1.878。
In summary, the migration learning strategy for the WRF mode wind element correction model provided in this embodiment can effectively solve the problem that the appropriate WRF mode correction model is difficult to construct at the point where the data material is lacking.
The embodiment also provides a device for transferring and learning the wind factor correction model, which is used for realizing the embodiment and the preferred implementation manner, and is not described in detail. As used below, the term "module" may be a combination of software and/or hardware that implements a predetermined function. While the means described in the following embodiments are preferably implemented in software, implementation in hardware, or a combination of software and hardware, is also possible and contemplated.
The present embodiment provides a transfer learning device of a wind element correction model, as shown in fig. 9, including:
the acquiring module 901 is configured to acquire a first data set at a preset point position and a second data set at a point position to be corrected in the offshore wind farm in a preset weather forecast mode;
The building module 902 is configured to build a wind element source domain correction model based on the first data set through a self-adaptive momentum random optimization algorithm;
a calculation module 903, configured to calculate a determination coefficient of the first measured wind speed dataset and the second measured wind speed dataset;
the migration learning module 904 is configured to perform migration learning on the wind element source domain correction model based on the second data set and the decision coefficient, and generate a wind element target domain correction model.
The correction module 905 is configured to correct the second weather forecast mode data set at the point to be corrected by using the wind element target domain correction model, so as to obtain a correction result.
In some alternative embodiments, the establishing module 902 includes:
the first acquisition sub-module is used for acquiring a first neural network parameter set of the initial wind element source domain correction model.
The first training sub-module is used for training the initial wind element source domain correction model by utilizing the self-adaptive momentum random optimization algorithm based on the first data set and the first neural network parameter set until the wind element source domain correction model meeting the preset first condition is obtained.
In some alternative embodiments, the first training submodule includes:
the first dividing unit is used for dividing the first data set into a first training data set and a first test data set according to time sequence.
And the second dividing unit is used for dividing the first training data set into at least one batch of training data according to the batch size order.
The first training unit is used for training the initial wind element source domain correction model by utilizing the self-adaptive momentum random optimization algorithm based on each batch of training data and the first neural network parameter set.
The judging unit is used for judging whether the trained initial wind element source domain correction model meets a preset first condition or not by using the first test data set.
And the first determining unit is used for determining the wind factor source domain correction model according to the judging result.
In some alternative embodiments, the determining unit comprises:
and the first determining subunit is used for determining the trained initial wind element source domain correction model as a wind element source domain correction model when the trained initial wind element source domain correction model meets a preset first condition.
And the second determining subunit is used for adjusting the first neural network parameter set based on each batch of training data when the trained initial wind factor source domain correction model does not meet the preset first condition.
The training subunit is used for training the initial wind element source domain correction model by utilizing the self-adaptive momentum random optimization algorithm based on each batch of training data and the adjusted first neural network parameter set until the wind element source domain correction model meeting the preset first condition is obtained.
In some alternative embodiments, the computing module 903 includes:
the first calculation sub-module is used for calculating a first measured data average value of a first measured wind speed data set in the first data set.
And the second calculation sub-module is used for calculating a second measured data average value of a second measured wind speed data set in the second data set.
And the third calculation sub-module is used for calculating the judgment coefficient based on the first measured data average value and the second measured data average value.
In some alternative embodiments, the transfer learning module 904 includes:
the second acquisition sub-module is used for acquiring a preset first threshold value and a preset second threshold value.
And the comparison sub-module is used for respectively comparing the judgment coefficient with a preset first threshold value and a preset second threshold value.
And the generation sub-module is used for performing migration learning on the wind element source domain correction model according to the comparison result and the second data set to generate a wind element target domain correction model.
In some alternative embodiments, generating the sub-module includes:
the first acquisition unit is used for acquiring a second neural network parameter set of the wind element source domain correction model when the judgment coefficient is larger than a preset second threshold value.
The first adjusting unit is used for adjusting the second neural network parameter set based on the second data set to obtain the second neural network parameter set.
The second training unit is used for training the wind element source domain correction model by utilizing the self-adaptive momentum random optimization algorithm based on the second neural network parameter set and the second data set until the wind element target domain correction model meeting the preset second condition is obtained.
In some alternative embodiments, the generating sub-module further comprises:
the second obtaining unit is used for obtaining the output layer weight parameters and other neural network layer weight parameters in the second neural network parameter set when the judgment coefficient is larger than a preset first threshold value and smaller than a preset second threshold value.
The first initializing unit is used for initializing the output layer weight parameters to obtain target output layer weight parameters.
And the second determining unit is used for determining a third neural network parameter set based on the target output layer weight parameter and other neural network layer weight parameters.
And the third training unit is used for training the wind element source domain correction model by utilizing the self-adaptive momentum random optimization algorithm based on the third neural network parameter set and the second data set until the wind element target domain correction model meeting the preset second condition is obtained.
In some alternative embodiments, the generating sub-module further comprises:
And the second initializing unit is used for initializing the weight parameters of each neural network layer contained in the second neural network parameter set when the judgment coefficient is smaller than a preset first threshold value to obtain a fourth neural network parameter set.
And the fourth training unit is used for training the wind element source domain correction model by utilizing the self-adaptive momentum random optimization algorithm based on the fourth neural network parameter set and the second data set until the wind element target domain correction model meeting the preset second condition is obtained.
In some optional embodiments, the transfer learning device of the wind element correction model further includes:
and the input module is used for inputting the second weather forecast mode data set into the wind element target domain correction model to obtain weather forecast mode target domain correction data.
The first calculation module is used for calculating an error index based on the weather forecast mode target domain correction data.
And the evaluation module is used for evaluating the correction result based on the error index.
Further functional descriptions of the above respective modules and units are the same as those of the above corresponding embodiments, and are not repeated here.
The wind element correction model transfer learning device in this embodiment is presented in the form of a functional unit, where the unit refers to an ASIC (Application Specific Integrated Circuit ) circuit, a processor and a memory executing one or more software or fixed programs, and/or other devices that can provide the above functions.
The embodiment of the invention also provides computer equipment, which is provided with the transfer learning device of the wind element correction model shown in the figure 9.
Referring to fig. 10, fig. 10 is a schematic structural diagram of a computer device according to an alternative embodiment of the present invention, as shown in fig. 10, the computer device includes: one or more processors 10, memory 20, and interfaces for connecting the various components, including high-speed interfaces and low-speed interfaces. The various components are communicatively coupled to each other using different buses and may be mounted on a common motherboard or in other manners as desired. The processor may process instructions executing within the computer device, including instructions stored in or on memory to display graphical information of the GUI on an external input/output device, such as a display device coupled to the interface. In some alternative embodiments, multiple processors and/or multiple buses may be used, if desired, along with multiple memories and multiple memories. Also, multiple computer devices may be connected, each providing a portion of the necessary operations (e.g., determined as a server array, a set of blade servers, or a multiprocessor system). One processor 10 is illustrated in fig. 10.
The processor 10 may be a central processor, a network processor, or a combination thereof. The processor 10 may further include a hardware chip, among others. The hardware chip may be an application specific integrated circuit, a programmable logic device, or a combination thereof. The programmable logic device may be a complex programmable logic device, a field programmable gate array, a general-purpose array logic, or any combination thereof.
Wherein the memory 20 stores instructions executable by the at least one processor 10 to cause the at least one processor 10 to perform a method for implementing the embodiments described above.
The memory 20 may include a storage program area that may store an operating system, at least one application program required for functions, and a storage data area; the storage data area may store data created according to the use of the computer device, etc. In addition, the memory 20 may include high-speed random access memory, and may also include non-transitory memory, such as at least one magnetic disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, memory 20 may optionally include memory located remotely from processor 10, which may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the internet, intranets, local area networks, mobile communication networks, and combinations thereof.
Memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk, or solid state disk; the memory 20 may also comprise a combination of the above types of memories.
The computer device also includes a communication interface 30 for the computer device to communicate with other devices or communication networks.
The embodiments of the present invention also provide a computer readable storage medium, and the method according to the embodiments of the present invention described above may be implemented in hardware, firmware, or as a computer code which may be recorded on a storage medium, or as original stored in a remote storage medium or a non-transitory machine readable storage medium downloaded through a network and to be stored in a local storage medium, so that the method described herein may be stored on such software process on a storage medium using a general purpose computer, a special purpose processor, or programmable or special purpose hardware. The storage medium can be a magnetic disk, an optical disk, a read-only memory, a random access memory, a flash memory, a hard disk, a solid state disk or the like; further, the storage medium may also comprise a combination of memories of the kind described above. It will be appreciated that a computer, processor, microprocessor controller or programmable hardware includes a storage element that can store or receive software or computer code that, when accessed and executed by the computer, processor or hardware, implements the methods illustrated by the above embodiments.
Although embodiments of the present invention have been described in connection with the accompanying drawings, various modifications and variations may be made by those skilled in the art without departing from the spirit and scope of the invention, and such modifications and variations fall within the scope of the invention as defined by the appended claims.

Claims (11)

1. A method for transition learning of a wind element correction model, the method comprising:
acquiring a first data set at a preset point position and a second data set at a point position to be corrected in the offshore wind farm when the offshore wind farm is in a preset weather forecast mode, wherein the first data set comprises a first weather forecast mode data set and a first actually measured wind speed data set, and the second data set comprises a second weather forecast mode data set and a second actually measured wind speed data set;
based on the first data set, establishing a wind element source domain correction model through a self-adaptive momentum random optimization algorithm;
calculating a judgment coefficient of the first measured wind speed data set and the second measured wind speed data set;
based on the second data set and the judging coefficient, performing migration learning on the wind element source domain correction model to generate a wind element target domain correction model;
Correcting the second weather forecast mode data set at the position of the point to be corrected by using the wind element target domain correction model to obtain a correction result;
based on the second data set and the determination coefficient, performing migration learning on the wind element source domain correction model to generate a wind element target domain correction model, including:
acquiring a preset first threshold value and a preset second threshold value;
comparing the judgment coefficient with the preset first threshold value and the preset second threshold value respectively;
performing migration learning on the wind element source domain correction model according to the comparison result and the second data set to generate the wind element target domain correction model, wherein the method comprises the following steps:
when the judging coefficient is larger than the preset second threshold value, acquiring a second neural network parameter set of the wind element source domain correction model;
training the wind element source domain correction model by utilizing a self-adaptive momentum random optimization algorithm based on the second neural network parameter set and the second data set until the wind element target domain correction model meeting a preset second condition is obtained;
when the judging coefficient is larger than the preset first threshold value and smaller than the preset second threshold value, acquiring an output layer weight parameter and other neural network layer weight parameters in the second neural network parameter set;
Initializing the output layer weight parameters to obtain target output layer weight parameters;
determining a third neural network parameter set based on the target output layer weight parameter and the other neural network layer weight parameters;
training the wind element source domain correction model by utilizing the self-adaptive momentum random optimization algorithm based on the third neural network parameter set and the second data set until the wind element target domain correction model meeting the preset second condition is obtained;
initializing each neural network layer weight parameter contained in the second neural network parameter set when the judgment coefficient is smaller than the preset first threshold value to obtain a fourth neural network parameter set;
and training the wind element source domain correction model by utilizing the self-adaptive momentum random optimization algorithm based on the fourth neural network parameter set and the second data set until the wind element target domain correction model meeting the preset second condition is obtained.
2. The method of claim 1, wherein establishing a wind factor source domain correction model based on the first dataset via an adaptive momentum random optimization algorithm comprises:
Acquiring a first neural network parameter set of an initial wind element source domain correction model;
based on the first data set and the first neural network parameter set, training the initial wind element source domain correction model by utilizing a self-adaptive momentum random optimization algorithm until the wind element source domain correction model meeting a preset first condition is obtained.
3. The method of claim 2, wherein training the initial wind element source domain correction model using an adaptive momentum random optimization algorithm based on the first data set and the first neural network parameter set until the wind element source domain correction model satisfying a preset first condition is obtained, comprising:
dividing the first data set into a first training data set and a first test data set according to time sequence;
dividing the first training data set into at least one batch of training data according to the batch size order;
training the initial wind element source domain correction model by utilizing a self-adaptive momentum random optimization algorithm based on each batch of training data and the first neural network parameter set;
judging whether the trained initial wind element source domain correction model meets the preset first condition or not by utilizing the first test data set;
And determining the wind factor source domain correction model according to the judging result.
4. A method according to claim 3, wherein determining the wind element source domain correction model based on the determination result comprises:
when the trained initial wind element source domain correction model meets the preset first condition, determining the trained initial wind element source domain correction model as the wind element source domain correction model;
when the trained initial wind element source domain correction model does not meet the preset first condition, adjusting the first neural network parameter set based on each batch of training data;
based on each batch of training data and the adjusted first neural network parameter set, training the initial wind element source domain correction model by using a self-adaptive momentum random optimization algorithm until the wind element source domain correction model meeting the preset first condition is obtained.
5. The method of claim 1, wherein calculating the decision coefficients for the first and second measured wind speed datasets comprises:
calculating a first measured data average value of the first measured wind speed dataset in the first dataset;
Calculating a second measured data average value of the second measured wind speed data set in the second data set;
and calculating the judgment coefficient based on the first measured data average value and the second measured data average value.
6. The method according to claim 1, wherein the method further comprises:
inputting the second weather forecast mode data set into the wind element target domain correction model to obtain weather forecast mode target domain correction data;
calculating an error index based on the weather forecast mode target domain correction data;
and evaluating the correction result based on the error index.
7. A wind factor correction model migration learning apparatus, the apparatus comprising:
the system comprises an acquisition module, a correction module and a correction module, wherein the acquisition module is used for acquiring a first data set at a preset point position and a second data set at a point position to be corrected in the offshore wind farm when the offshore wind farm is in a preset weather forecast mode, the first data set comprises a first weather forecast mode data set and a first actual measurement wind speed data set, and the second data set comprises a second weather forecast mode data set and a second actual measurement wind speed data set;
the building module is used for building a wind element source domain correction model through a self-adaptive momentum random optimization algorithm based on the first data set;
The calculation module is used for calculating the judgment coefficients of the first measured wind speed data set and the second measured wind speed data set;
the transfer learning module is used for performing transfer learning on the wind element source domain correction model based on the second data set and the judging coefficient to generate a wind element target domain correction model;
the correction module is used for correcting the second weather forecast mode data set at the position of the point to be corrected by using the wind element target domain correction model to obtain a correction result;
the migration learning module comprises:
the second acquisition submodule is used for acquiring a preset first threshold value and a preset second threshold value;
the comparison sub-module is used for respectively comparing the judgment coefficient with a preset first threshold value and a preset second threshold value;
the generation sub-module is used for performing migration learning on the wind element source domain correction model according to the comparison result and the second data set to generate a wind element target domain correction model;
the generating sub-module includes:
the first acquisition unit is used for acquiring a second neural network parameter set of the wind element source domain correction model when the judgment coefficient is larger than a preset second threshold value;
the first adjusting unit is used for adjusting the second neural network parameter set based on the second data set to obtain a second neural network parameter set;
The second training unit is used for training the wind element source domain correction model by utilizing a self-adaptive momentum random optimization algorithm based on the second neural network parameter set and the second data set until a wind element target domain correction model meeting a preset second condition is obtained;
the second acquisition unit is used for acquiring the output layer weight parameters and other neural network layer weight parameters in the second neural network parameter set when the judgment coefficient is larger than a preset first threshold value and smaller than a preset second threshold value;
the first initializing unit is used for initializing the output layer weight parameters to obtain target output layer weight parameters;
the second determining unit is used for determining a third neural network parameter set based on the target output layer weight parameter and other neural network layer weight parameters;
the third training unit is used for training the wind element source domain correction model by utilizing a self-adaptive momentum random optimization algorithm based on a third neural network parameter set and a second data set until a wind element target domain correction model meeting a preset second condition is obtained;
the second initializing unit is used for initializing weight parameters of each neural network layer contained in the second neural network parameter set when the judgment coefficient is smaller than a preset first threshold value to obtain a fourth neural network parameter set;
And the fourth training unit is used for training the wind element source domain correction model by utilizing the self-adaptive momentum random optimization algorithm based on the fourth neural network parameter set and the second data set until the wind element target domain correction model meeting the preset second condition is obtained.
8. The apparatus of claim 7, wherein the means for establishing comprises:
the first acquisition sub-module is used for acquiring a first neural network parameter set of the initial wind element source domain correction model;
the first training sub-module is used for training the initial wind element source domain correction model by utilizing a self-adaptive momentum random optimization algorithm based on the first data set and the first neural network parameter set until the wind element source domain correction model meeting a preset first condition is obtained.
9. The apparatus of claim 8, wherein the first training submodule comprises:
a first dividing unit for dividing the first data set into a first training data set and a first test data set according to a time sequence;
a second dividing unit, configured to divide the first training data set into at least one batch of training data according to a batch size order;
The first training unit is used for training the initial wind element source domain correction model by utilizing a self-adaptive momentum random optimization algorithm based on each batch of training data and the first neural network parameter set;
the judging unit is used for judging whether the trained initial wind element source domain correction model meets the preset first condition or not by utilizing the first test data set;
and the first determining unit is used for determining the wind factor source domain correction model according to the judging result.
10. A computer device, comprising:
a memory and a processor, the memory and the processor being communicatively connected to each other, the memory having stored therein computer instructions, the processor executing the computer instructions to perform the method of migrating a wind element correction model according to any one of claims 1 to 6.
11. A computer-readable storage medium having stored thereon computer instructions for causing a computer to execute the method of shift learning of the wind element correction model according to any one of claims 1 to 6.
CN202311256975.6A 2023-09-27 2023-09-27 Method, device, equipment and medium for migration learning of wind element correction model Active CN116992222B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202311256975.6A CN116992222B (en) 2023-09-27 2023-09-27 Method, device, equipment and medium for migration learning of wind element correction model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202311256975.6A CN116992222B (en) 2023-09-27 2023-09-27 Method, device, equipment and medium for migration learning of wind element correction model

Publications (2)

Publication Number Publication Date
CN116992222A CN116992222A (en) 2023-11-03
CN116992222B true CN116992222B (en) 2024-01-26

Family

ID=88525223

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202311256975.6A Active CN116992222B (en) 2023-09-27 2023-09-27 Method, device, equipment and medium for migration learning of wind element correction model

Country Status (1)

Country Link
CN (1) CN116992222B (en)

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104268627A (en) * 2014-09-10 2015-01-07 天津大学 Short-term wind speed forecasting method based on deep neural network transfer model
CN112288164A (en) * 2020-10-29 2021-01-29 四川大学 Wind power combined prediction method considering spatial correlation and correcting numerical weather forecast
KR20220102329A (en) * 2021-01-13 2022-07-20 대한민국(산림청 국립산림과학원장) Apparatus and method for forecasting the seasonal forest fire risk
CN115983494A (en) * 2023-02-10 2023-04-18 广东工业大学 Short-term wind power prediction method and system for newly-built small-sample wind power plant
CN116227554A (en) * 2023-03-06 2023-06-06 中国长江三峡集团有限公司 Analog data correction method and device for meteorological data and electronic equipment
CN116777039A (en) * 2023-04-24 2023-09-19 华南理工大学 Double-layer neural network wind speed prediction method based on training set segmentation and error correction

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104268627A (en) * 2014-09-10 2015-01-07 天津大学 Short-term wind speed forecasting method based on deep neural network transfer model
CN112288164A (en) * 2020-10-29 2021-01-29 四川大学 Wind power combined prediction method considering spatial correlation and correcting numerical weather forecast
KR20220102329A (en) * 2021-01-13 2022-07-20 대한민국(산림청 국립산림과학원장) Apparatus and method for forecasting the seasonal forest fire risk
CN115983494A (en) * 2023-02-10 2023-04-18 广东工业大学 Short-term wind power prediction method and system for newly-built small-sample wind power plant
CN116227554A (en) * 2023-03-06 2023-06-06 中国长江三峡集团有限公司 Analog data correction method and device for meteorological data and electronic equipment
CN116777039A (en) * 2023-04-24 2023-09-19 华南理工大学 Double-layer neural network wind speed prediction method based on training set segmentation and error correction

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
Improved daily SMAP satelitte soil moisture prediction over China using Deep learning model with transfer learning;Qingliang Li等;Journal of Hydrology;第1-14页 *
Two-Layer Transfer-Learning-Based Architecture for Short-Term Load Forecasting;Long Cai等;IEEE Transactions on Industrial Informatics;第1722-1732页 *
基于深度学习的风电爬坡识别及短期功率预测技术研究;李文泽;中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑;第5章 *
基于特征选择与多层级深度迁移学习的风电场短期功率预测;程凯等;高电压技术;第497-503页 *

Also Published As

Publication number Publication date
CN116992222A (en) 2023-11-03

Similar Documents

Publication Publication Date Title
Zhang et al. Short-term rainfall forecasting using multi-layer perceptron
CN111563706A (en) Multivariable logistics freight volume prediction method based on LSTM network
US11231519B2 (en) Method and device for simulating discharge, and computer device
CN109001736B (en) Radar echo extrapolation method based on deep space-time prediction neural network
CN112288164B (en) Wind power combined prediction method considering spatial correlation and correcting numerical weather forecast
CN112287294B (en) Space-time bidirectional soil water content interpolation method based on deep learning
WO2023179167A1 (en) Crop irrigation water demand prediction method based on aquacrop model and svr
CN111310965A (en) Aircraft track prediction method based on LSTM network
KR102396290B1 (en) Method for providing ultra low altitude wind prediction information
CN112180369B (en) Depth learning-based sea surface wind speed inversion method for one-dimensional synthetic aperture radiometer
CN116227554A (en) Analog data correction method and device for meteorological data and electronic equipment
Nevtipilova et al. Testing artificial neural network (ANN) for spatial interpolation
CN114512191A (en) Penicillin concentration prediction method based on migration component analysis
CN113095547B (en) Short-term wind power prediction method based on GRA-LSTM-ICE model
CN113486591B (en) Gravity multi-parameter data density weighted inversion method for convolutional neural network result
CN116992222B (en) Method, device, equipment and medium for migration learning of wind element correction model
CN116976227B (en) Storm water increasing forecasting method and system based on LSTM machine learning
CN111914487B (en) Data-free regional hydrological parameter calibration method based on antagonistic neural network
CN110852475B (en) Extreme gradient lifting algorithm-based vegetation index prediction method, system and equipment
CN110852415A (en) Vegetation index prediction method, system and equipment based on neural network algorithm
CN115392594B (en) Electrical load model training method based on neural network and feature screening
CN115062551B (en) Wet physical process parameterization method based on time sequence neural network
CN107590346B (en) Downscaling correction model based on spatial multi-correlation solution set algorithm
CN116774316A (en) Prediction method and device for short-term precipitation, storage medium and computer equipment
CN116663395A (en) Rainfall equivalent surface generation method based on regression of support vector machine of parameter optimization

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant