JP2023085635A - Wind state estimation device and wind state estimation method - Google Patents

Abstract

To provide a wind state estimation device capable of estimating a wind state including uncertainties in a non-observation period.SOLUTION: A wind state estimation device 100 comprises: a storage unit 20 that stores calibration data 21 indicating a wind state measured near an installation site of a wind power generation facility, and reference data 22 indicating a wind state obtained by a homogeneous measurement method or analysis method; a regression model learning unit 12 that learns a regression model based on the calibration data 21 and the reference data 22 when a period included in the calibration data 21 is defined as a first period, a period included in the reference data 22 is defined as a second period, and a period obtained by subtracting the first period from the second period is defined as a third period; and a wind state estimation unit that estimates a wind state including uncertainties of estimation results based on the reference data 22 and the regression model in the third period.SELECTED DRAWING: Figure 1

Description

The present invention relates to a wind condition estimation device and a wind condition estimation method.

In recent years, carbon neutrality has attracted attention. Carbon neutrality is an international effort to conserve the global environment by reducing emissions of carbon dioxide, a greenhouse gas. In October 2020, the Japanese government also set a goal of "carbon neutrality by 2050," which aims to reduce greenhouse gas emissions to zero by 2050. Some companies are starting new businesses or changing their management policies to reduce greenhouse gas emissions. Especially in the field of power generation, further introduction of renewable energy power generation equipment is expected for low carbonization.

When introducing renewable energy power generation equipment, business operators conduct business feasibility evaluations at the business planning stage. In the business feasibility evaluation, the cash flow of the power generation facility is predicted, and the predicted cash flow is evaluated from the perspective of profitability and certainty of loan repayment. In the case of renewable energy power generation facilities, cash flow income is electricity sales income, and cash flow expenditure includes construction costs and maintenance and inspection costs.

The business period of renewable energy power generation facilities is about 20 years, and the weather tends to fluctuate from year to year. Since it is difficult to accurately predict the weather 10 to 20 years from now, instead of using weather forecasts, electricity sales revenues are predicted on the assumption that past weather will be repeated. However, it is not realistic to conduct weather observations for 10-20 years for one power plant. Therefore, we conducted meteorological observation for about one year at the power generation site, prepared suitable reference data expressing the weather for 10 to 20 years, calculated the correlation between the meteorological observation results and the reference data, and calculated the correlation. and reference data, it is common to generate weather data for 10-20 years. Calculate the amount of power generated based on the generated weather data and predict the income from power sales.

A method of calculating the correlation between the weather observation result and the reference data and estimating the wind conditions during the non-observation period based on the correlation and the reference data is disclosed in Patent Document 1, for example.

JP 2020-159725 A

In the business feasibility evaluation, it is necessary to consider not only the predicted value but also the possible range and distribution of the predicted value. For example, it may be confirmed that the loan repayment will not be hindered even if the electricity sales revenue falls below the predicted value. If the uncertainty of the predicted value can also be calculated, it will be useful to consider measures such as commercializing only power generation facilities that will not hinder loan repayment even if the revenue from electricity sales declines, or preparing a reserve fund.

Patent Literature 1 discloses a method of calculating one predicted value at each time. However, there is no disclosure of a method for calculating uncertainties such as the range and distribution of predicted values.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a wind condition estimating apparatus for estimating wind conditions including uncertainties during non-observation periods.

In order to solve the above-mentioned problems, the wind resource estimation device of the present invention provides calibration data indicating the wind resource measured in the vicinity of the installation site of the wind power generation facility, and indicates the wind resource obtained by a homogeneous measurement method or analysis method. a storage unit storing reference data; a period included in the calibration data as a first period; a period included in the reference data as a second period; and from the second period to the first period. a regression model learning unit that learns a regression model based on the calibration data and the reference data when the period excluding is a third period; and the reference data and the regression model in the third period and a wind condition estimating unit for estimating the wind condition including the uncertainty of the estimation result. Other aspects of the present invention are described in embodiments below.

ADVANTAGE OF THE INVENTION According to this invention, the wind condition estimation apparatus which estimates the wind condition in a non-observation period including uncertainty can be provided.

It is a figure which shows schematic structure of the wind condition estimation apparatus which concerns on embodiment. It is a figure which shows the processing flow by a wind condition estimation apparatus. FIG. 10 is a diagram showing accuracy comparison of regression models; FIG. 4 is a flow chart showing uncertainty estimation processing for estimating uncertainty by Monte Carlo dropout using a regression model as a neural network; FIG. 10 is a diagram showing an example of applying Monte Carlo dropout when estimating wind speed, and showing an example of displaying the uncertainty of wind speed estimation results. It is a figure which shows the example of a display of the display part of a wind condition estimation apparatus.

Embodiments of the present invention will be described below with reference to the drawings. Although the characteristic configuration and processing of the wind condition estimation device will be described, the wind condition estimation device may include configurations and processing that are omitted below.
(overall structure)
FIG. 1 is a diagram showing a schematic configuration of a wind condition estimation device 100 according to an embodiment. The wind condition estimation device 100 has a processing unit 10 , a storage unit 20 , an input unit 30 , a display unit 40 and a communication unit 50 . Storage unit 20 stores calibration data 21, reference data 22, learned regression model 23, wind condition estimation result 24, etc. Processing unit 10 includes data input unit 11, regression model learning unit 12, wind condition estimation It includes a unit 13, a screen display processing unit 14, and the like.

(Operating environment)
The wind condition estimation device 100 is an electronic computer having a communication function. The processing unit 10 is a central processing unit (CPU) and executes various programs stored in memory. The memory holds various programs and temporary data for the wind condition estimation device 100 to execute processing. The display unit 40 is a display or the like, and displays the execution status, execution results, and the like of processing by the wind condition estimation device 100 . The input unit 30 is a device such as a keyboard and a mouse for inputting instructions to the computer, and inputs instructions such as program activation. The communication unit 50 exchanges various data and commands with other devices. The storage unit 20 stores various data for the wind condition estimation device 100 to execute processing. When the regression model includes matrix calculation, it is preferable to have a GPU from the viewpoint of speeding up the calculation. As an interface, a touch panel as an input/output unit and voice input may be provided.

(data for calibration)
The calibration data 21 is time-series data of wind conditions measured in the vicinity of the wind power generation site. The data input section 11 saves the calibration data 21 in the storage section 20 .

When estimating wind conditions with emphasis on immediacy, it is preferable that the data input unit 11 has a data acquisition function for acquiring the calibration data 21 so that the measuring device and the data input unit 11 can communicate with each other. Wind conditions include at least wind speed.

Instruments include wind towers, Doppler Lidar, and Doppler Sodar. A wind condition observation tower is a measuring instrument equipped with an anemometer by protruding an arm from a steel tower of about 60m. A Doppler lidar is a measuring device that emits laser light and receives scattered light from aerosols to measure wind speed and direction. The Doppler lidar may be installed on the ground and emit laser light into the sky, installed on the nacelle of the wind power generation facility and emitted forward, or installed on an unmanned aerial vehicle (drone). A Doppler soda is a measuring instrument that emits sound waves to measure wind speed and direction.

It is preferable to include the wind direction in addition to the wind speed. Seventy percent of Japan's land is mountainous, and the number of wind power generation facilities being built on mountainous terrain is increasing. In the power generation calculation of wind power generation in mountainous terrain, it is necessary to consider not only the wind speed but also the wind direction. Compared to flat terrain such as plains and coasts, it is more susceptible to the undulations and roughness of the surrounding area. Also, wind farms in mountainous terrain are often installed with multiple units instead of one, and are more susceptible to wakes. A wake is a phenomenon in which the wind received by the wind power generation equipment on the downwind side is reduced compared to the free air flow. Since the amount of power generated by the wind power generation equipment on the leeward side decreases, even if the wind speed is the same, if the wind direction is different, the total amount of power generated by the multiple wind power generation equipment will be different. Additionally, temperature, humidity, precipitation, and solar radiation may be included.

It is preferable that the wind condition has a value for each height above the ground. It is known that the wind speed tends to increase with increasing height above the ground. In addition, the nacelle of wind power generation equipment is installed about 100m from the ground, but the blade has a length of about 50m, so the lowest end of the trajectory that the tip of the blade passes is about 50m from the ground, and the highest end is It is about 150m above the ground. Therefore, it is more preferable to have values from 50m to 150m above the ground.

The data input unit 11 preferably has a function of verifying the calibration data 21 before storing the calibration data 21 in the storage unit 20 . For example, when the calibration data 21 represents the wind speed, the wind speed is a numerical value of 0 or more, so it may have a function to return an error when there is a numerical value less than 0.

(reference data)
The reference data 22 is time-series data of wind conditions obtained by a uniform measurement method or analysis method (measurement method or analysis method). The data input unit 11 saves the reference data 22 in the storage unit 20 .

The period included in the calibration data 21 is defined as the first period, the period included in the reference data 22 is defined as the second period, and the period obtained by subtracting the first period from the second period is defined as the third period. Consider learning a regression model based on the data 21 and the reference data 22 and estimating based on the reference data 22 and the regression model in the third period.

It is desirable that the reference data 22 has the same measurement method and analysis method, or a negligibly small change in the second period. If data with different measurement methods or analysis methods are included, there is a concern that the regression model learned in the first period cannot be applied in the third period. Factors for different measurement methods include, for example, different observation locations and observation equipment. Wind conditions are affected by the surroundings of the observation location, so the values will change if the observation location changes. Factors for the different analysis methods include, for example, different spatial resolutions and different data assimilation techniques.

(Reference data that is compatible with each application)
Suitable reference data 22 differs depending on the application. There are mainly three types of applications of the present invention. There are three types: backcast, forecast, and interpolation.

(Reference data suitable for backcasting)
In the business feasibility evaluation conducted when introducing renewable energy power generation facilities, instead of forecasting wind conditions for the next 10 to 20 years, it is assumed that the weather for the past 10 to 20 years will be repeated, and the past 10 to 20 years to calculate wind conditions.

Backcast estimates past wind conditions to evaluate the feasibility of wind power generation. Assuming that the period included in the calibration data 21 is the first period and the period included in the reference data 22 is the second period, the first period and the second period overlap, and the second period is The start date is earlier than the start date of the first period.

In backcasting, the longer the period of the reference data 22, the better, preferably ten years or more. In addition, data issued by a third-party organization is desirable for the business feasibility evaluation because it is conducted as part of a loan discussion between a business operator and a financial institution, and from the viewpoint of data falsification prevention.

Reference data 22 that satisfies these requirements and is suitable for backcasting includes, for example, observation results from the Meteorological Agency and meteorological companies, and long-term reanalysis data. The Japan Meteorological Agency publishes weather data observed by the Regional Meteorological Observation System (AMEDAS) on its website. Long-term reanalysis data are meteorological data generated with the same spatial resolution and data assimilation methods, primarily to capture long-term trends in climate change. Examples include MERRA-2 of the US National Aeronautics and Space Administration, ERA5 of the European Center for Medium-Term Weather Forecasts, and JRA-55 of the Japan Meteorological Agency.

(Reference data suitable for forecasting)
In the case of a power generation business that uses the feed-in premium system (hereinafter referred to as FIP), which will come into effect in April 2022, power generation companies will bid based on forecasts of the amount of power to be generated.

Forecast estimates future wind conditions for forecasting and controlling the amount of power generated by wind power generation facilities. Assuming that the period included in the calibration data 21 is the first period and the period included in the reference data 22 is the second period, the first period and the second period overlap, and the second period is The end date is later than the end date of the first period.

In the forecast, the reference data 22 are future forecast values. The period of the reference data 22 is shorter than the backcast, and may be from one day to several days.

Examples of reference data 22 suitable for forecasting include GPV (Grid Point Value) and MSM (Meso scale model), which are numerical forecast models that predict areas including Japan with a three-dimensional grid, published by the Meteorological Agency. .

(reference data suitable for interpolation)
Meteorological observations at the power generation project site may be lost due to lightning strikes, freezing, and interference from animals and plants. Interpolation interpolates missing wind condition data. When the period included in the calibration data 21 is the first period and the period included in the reference data 22 is the second period, the period to be interpolated should be included in both periods. The period of the reference data 22 may be shorter than the backcast.

Reference data suitable for interpolation may be objective analysis data in addition to the observation results and long-term reanalysis data of the Meteorological Agency and meteorological companies cited as reference data suitable for backcasting. For example, FNL of the US National Center for Environmental Prediction.

(Supplement 1 regarding reference data)
If the reference data 22 is an analysis method that expresses the state of the atmosphere as grid point values, the reference data 22 may be values that express the latitude and longitude of grid points instead of the wind power generation site. As a method of selecting grid points, there are a method of selecting grid points close to the wind power generation project site, and a method of selecting grid points with similar topographical conditions from a group of neighboring grid points. One or more grid points may be adopted as the reference data 22 .

(Supplement 2 regarding reference data)
The data input unit 11 preferably has a function of verifying the reference data 22 before storing the reference data 22 in the storage unit 20 . For example, if the reference data 22 is an analysis value, it is preferable to have a function to check whether the time zones of the reference data 22 and the calibration data 21 are aligned. This is because the time zone of the reference data 22 may be registered in universal standard time even in Japan.

(processing flow)
FIG. 2 is a diagram showing a processing flow by the wind condition estimation device 100. As shown in FIG. The process flow will be described below. Although FIG. 2 shows a series of processing flows, it is possible to start or end the processing flow in the middle of the processing flow.

(Regression model learning process)
The regression model learning process S<b>1 is executed by the regression model learning unit 12 . The regression model learning unit 12 acquires the calibration data 21 and the reference data 22 from the storage unit 20 , learns the regression model, and stores the learned regression model 23 in the storage unit 20 .

Various methods can be applied to the regression model as long as it is a model that can output a desired item using the reference data 22 as an input, and can calculate parameters from the calibration data 21 . Here, numerical analysis using linear regression and neural networks for one year of wind condition data measured at two locations is exemplified.

Wind conditions are 10-minute values of wind speed (m/s) and wind direction (degrees). The wind speed is given as a value greater than or equal to 0. The wind direction is given in degrees of 0 degrees or more and less than 360 degrees, with north being 0 degrees. One of the wind condition data is used as reference data 22 . Of the other, the three-fourths continuous period from the end is used as calibration data 21, and the remaining period is used as evaluation data.

A period included in the calibration data 21 is defined as a first period, a period included in the reference data 22 is defined as a second period, and a period obtained by subtracting the first period from the second period is defined as a third period. At this time, the period included in the evaluation data is the third period. Also, since the start date of the reference data 22 (second period) is earlier than the start date of the calibration data 21 (first period), backcasting is performed.

The wind condition estimation result 24 and evaluation data are compared to verify the accuracy of each regression model. Although there are various accuracy indexes, the root mean square error (hereinafter referred to as RMSE, which is an acronym for Root Mean Square Error) is used for verification. RMSE can be calculated by taking the square root of the mean of the squared errors at each time. Calculate the RMSE for each wind speed and wind direction. When calculating the wind direction error, it is processed so that the error is less than 180 degrees. For example, if the wind condition estimation result 24 is 359 degrees and the evaluation data is 0 degrees, the error is not 359 degrees, which is the difference between 359 degrees and 0 degrees. do.

FIG. 3 is a diagram showing a comparison of accuracy of regression models. FIG. 3 shows the results of accuracy verification while changing the regression model (points a to g). The smaller the RMSE, the better the accuracy, and 0 means the highest accuracy. Therefore, in FIG. 3, the accuracy is higher toward the lower left.

(About point a)
Point a in FIG. 3 is the case where the reference data 22 is the wind condition estimation result, that is, the case where no correction is made. The wind direction RMSE is 65 (degrees) and the wind speed RMSE is 4.1 (m/s). In the subsequent numerical analysis, we will consider how much it can be reduced based on this accuracy.

(About point b)
Point b in FIG. 3 shows that the wind speed learns a linear regression equation that inputs reference data 22 and outputs calibration data 21 in the first period, and inputs reference data 22 in the third period. The obtained value is used as the estimation result. The reference data 22 is used as an estimation result for the wind direction. Since the wind direction estimation result is the same as point a in FIG. 3, the wind direction RMSE (degrees) is the same as point a in FIG. The wind speed RMSE (m/s) is 3.5 (m/s), and the wind speed estimation accuracy is higher than at point a in FIG.

(About point c)
What can be done to improve the accuracy of wind direction estimation? At point c in FIG. 3, the wind direction learns a linear regression equation with reference data 22 as input and calibration data 21 as output in the first period, and the reference data 22 as input in the third period. The value obtained by doing so is used as the estimation result, and the reference data 22 is used as the estimation result for the wind speed. Since the estimation result of the wind speed is the same as the point a in FIG. 3, the wind speed RMSE (m/s) is the same as the point a in FIG. On the other hand, the wind direction RMSE (degrees) is 94 (degrees), and the wind direction estimation accuracy is lower than at point a in FIG. Since the wind direction is discontinuous from 0 degrees to 360 degrees, it is considered that the north wind cannot be expressed well and the estimation accuracy is low.

(About point d)
Therefore, a method for treating the wind speed V and the wind direction θ by decomposing them into the north-south component Vns and the east-west component Vew of the wind speed is considered. First, a method for converting the wind speed into north-south and east-west components is shown. Let the wind speed be V (m/s). Let θ (degrees) be the wind direction in a clockwise degree system where the north is 0 (degrees) and the east is 90 (degrees). With the east-west direction as the X axis, the north-south direction as the Y axis, and the northeast as the first quadrant, the wind direction φ (rad) by the counterclockwise radian method in the orthogonal coordinate system is calculated using the formula (1) using the circular constant π can be calculated by
φ=(90−θ)×π/180 (1)

The north-south component Vns (m/s) of the wind speed can be calculated by Equation (2), and the east-west component Vew (m/s) of the wind speed can be calculated by Equation (3).
Vns=V×sin(φ) …(2)
Vew=V×cos(φ) …(3)

Next, a method for converting the north-south component Vns and the east-west component Vew of the wind speed into the wind speed V and the wind direction θ will be described. The wind direction φ (rad) by the counterclockwise radian method can be calculated by Equation (4) using the arctan function with the end region from -π/2 to π/2.
If Vew>0, φ=arctan(Vns/Vew),
If Vew<0, φ=arctan(Vns/Vew)−π,
φ=π/2 if Vew=0 and Vns>0,
If Vew=0 and Vns<0, then φ=-π/2 (4)

When the counterclockwise wind direction φ (rad) is given by the radian method, the clockwise wind direction θ (degrees) can be calculated by Equation (5).
θ=90−φ×180/π (5)

Since the possible range of φ (rad) in Equation (4) is from -3π/2 to π/2, the possible range of θ (degrees) in Equation (5) is from 0 (degrees) to 360 (degrees). .
The wind speed V can be calculated by Equation (6).
V=√(Vns ² +Vew ² ) (6)

Point d in FIG. 3 is obtained by the following procedure.
First, the wind speed and wind direction of the reference data 22 and the calibration data 21 are converted into the north-south component and the east-west component of the wind speed based on formulas (1) to (3).

Next, in the first period, the north-south component of the wind speed of the reference data 22 is input, and the north-south component of the wind speed of the calibration data 21 is output. The north-south component of the wind speed is estimated by inputting the north-south component of the wind speed of . Also, in the first period, a linear regression formula is learned in which the east-west component of the wind speed of the reference data 22 is input and the east-west component of the wind speed of the calibration data 21 is output. The east-west component of the wind speed is estimated by inputting the east-west component of the wind speed.

Furthermore, from the north-south component and the east-west component of the estimated wind speed, the estimation result of wind speed and wind direction in the third period is obtained based on Equations (4) to (6). The wind direction RMSE (degrees) is 76 (degrees), and the wind direction estimation accuracy is higher than point c in FIG. 3, but lower than point a in FIG. The wind speed RMSE (m/s) is 4.2 (m/s), and the wind speed estimation accuracy is lower than that at point a in FIG. This suggests that the north-south and east-west components of wind speed may be difficult to express linearly.

(About point e)
Point e in FIG. 3 is obtained by using the linear regression equation of point d in FIG. 3 as a neural network. A neural network is a mathematical model inspired by neurons in the brain. It is known for its ability to accurately approximate arbitrary continuous functions, and is used in various fields. Neural networks solve regression and classification problems by adjusting the parameters of the neural network. Neural network parameters include the number of neurons, how they are connected, how they are implemented, and whether or not they are used.

Connections between neurons are mainly represented by a continuous layer structure. Layers are made up of neurons or groups of neurons. There is an input layer, an intermediate layer and an output layer. There may be multiple intermediate layers. A neuron on the output layer side receives a signal emitted by a neuron on the input layer side. Input layer neurons receive information (data and arguments), and output layer neurons output answers to regression and classification problems.

The coupling strength between neurons is represented by affine transformation. Affine transformation is expressed in the form of y=Wx+b, where x is a signal output by a neuron on the input layer side and y is a signal received by a neuron on the output layer side in two adjacent layers. Among the parameters of the affine transformation, the part (W) representing the linear transformation is called the weight, and the part (b) representing the constant term is called the bias. A neuron is represented by an activation function. Activation functions include, for example, identity functions, softmax functions, sigmoid functions, hyperbolic tangent functions (tanh functions), and normalized linear functions (ReLU functions).

The wind direction RMSE (degrees) is 63 (degrees), and the wind direction estimation accuracy is higher than at point a in FIG. It suggests the possibility that the neural network can successfully learn the nonlinear properties of the north-south and east-west components of wind speed.

(About point f)
Point f in FIG. 3 is a modification of the input and output layers of the neural network of point e in FIG. At point e in FIG. 3, the north-south component and the east-west component of the wind speed are assumed to be learning. Note that the neural network is a regression model that can have multiple inputs and multiple outputs. good too.

The wind direction RMSE (degrees) is 62 (degrees), and the wind direction estimation accuracy is higher than that in FIG. 3a. In this numerical analysis example, the estimation accuracy of the wind direction is the highest. The wind speed RMSE (m/s) is 4.0 (m/s), and the wind speed estimation accuracy is higher than that at point e in FIG.

At point f in FIG. 3, although the number of parameters between the input layer and the first intermediate layer and between the last intermediate layer and the output layer increases, the intermediate layers with a large number of parameters can use the same neural network. It is expected that the amount of calculation will be less than the point e of .

Practically, it is conceivable that the reference data 22 and the calibration data 21 have different time granularities. When matching to the coarser time granularity, we discard features that appear only at the finer time granularity. Point g in FIG. 3 is obtained by changing the time granularity of the north-south component and the east-west component of the wind speed from 10-minute values to 1-minute values at point f in FIG.

Consider the case where the reference data 22 and the calibration data 21 are wind speed and wind direction. The wind speed can be arbitrarily changed in time granularity by an appropriate interpolation method. Interpolation techniques include, for example, linear interpolation, spline interpolation, and polynomial interpolation. On the other hand, wind direction, which is a discrete value, is difficult to interpolate. For example, consider interpolating between wind speeds of the same value and wind directions of 1 degree and 359 degrees. If the wind direction is linearly interpolated, it will be 180 degrees, but since 1 degree and 359 degrees are northerly winds and 180 degrees is southerly winds, it is difficult to say that the interpolation is successful. On the other hand, if the reference data 22 and the calibration data 21 are converted into the north-south component and the east-west component of the wind speed, the north-south component and the east-west component of the wind speed are linearly interpolated, and the wind speed and wind direction are calculated from the linearly interpolated north-south component and the east-west component. , the interpolated wind direction is 0 degrees, which can be interpolated successfully.

There is software for calculating the power generated by a wind power generation facility, which has a function of inputting wind condition data and calculating the corresponding power generated (kW). When the 10-minute value of the wind condition data is input, the 10-minute value of the generated power is output, and when the 1-hour value of the wind condition data is input, the 1-hour value of the generated power is output. For example, by simulating a wind farm controller (a function that suppresses the output of wind power generation equipment so that it does not exceed the interconnection capacity when the total rated output of multiple wind power generation equipment exceeds the interconnection capacity). When the generated power is post-processed, the finer the time granularity, the higher the accuracy of the process.

The wind direction RMSE (degrees) at point g in FIG. 3 is 62 (degrees) and the wind speed RMSE (m/s) is 3.9 (m/s). is also expensive. It is suggested that processing with finer time granularity has the effect of data augmentation (a method of increasing the amount of data used for neural network learning). If the wind speed estimation accuracy is improved, the power generation estimation accuracy is also improved. Since the wind energy is proportional to the cube of the wind speed, the accuracy of estimating the amount of power generated is greatly improved over the accuracy of estimating the wind speed.

(Ingenuity in inputting regression model learning processing)
Although wind speed and wind direction are input this time, other data may be input from the viewpoint of high accuracy. For example, weather conditions such as temperature, humidity, precipitation, and solar radiation may be included. Wind conditions are characterized by yearly and seasonal fluctuations, and may include calendars. For coastal wind conditions, land and sea breezes are known and may include time of day and binary variables representing pre-sunset and post-sunset. Furthermore, not only the value at a certain time, but also the difference from the previous time, the value at the previous time, and the moving average up to the previous time may be included.

The value may not be input as it is, but may be input after being normalized by an appropriate method. Normalization methods include, for example, a method of subtracting the minimum value in the first period from the value at each time and dividing by the difference between the maximum value and the minimum value in the first period (Min-Max Normalization); There is a technique (Z-Score Normalization) of subtracting the average of the period and dividing by the standard deviation of the first period.

(Ingenuity in the output of regression model learning processing)
Since the regression model takes more time to learn than to estimate, it is preferable to save the regression model from the viewpoint of speeding up calculation. If the regression model is linear regression, for example, store the linear coefficient (slope) and intercept for each variable. If the regression model is a neural network, it stores, for example, the strength of connections between neurons.

(Wind condition estimation processing)
The wind condition estimation process S<b>2 is executed by the wind condition estimation unit 13 . The wind condition estimation unit 13 acquires the learned regression model 23 and the reference data 22 from the storage unit 20 . Assuming that the period included in the calibration data 21 is the first period and the period included in the reference data 22 is the second period, the wind conditions in the third period obtained by subtracting the first period from the second period to estimate Then, the estimation result is stored in the storage unit 20 as the wind condition estimation result 24 .

For example, in the case where the learned regression model 23 inputs the wind speed of the reference data 22 and outputs the wind speed, the wind condition estimation unit 13 estimates the wind speed at each time in the third period.

The wind condition estimation result 24 may include not only values at each time but also ranges. If the regression model is linear regression, the prediction interval can be calculated using the t-distribution. If the regression model is a neural network, one method for calculating the range is Monte Carlo dropout. Dropout in neural networks is a technique to prevent overfitting. It is a method of ignoring neurons at random during training and training a different neural network each time. Monte Carlo dropout is characterized by randomly dropping neurons not only during training but also during estimation. In the case of dropout, all neurons are used when estimating, so feeding the same input to a trained model yields the same result. On the other hand, in the case of Monte Carlo dropout, neurons are randomly lost during estimation, so the same input gives different results. By outputting multiple results, the uncertainty of the neural network can be expressed. An implementation that eliminates not only neurons but also layers may be used.

FIG. 4 is a flow chart showing uncertainty estimation processing S40 for estimating uncertainty by Monte Carlo dropout using a regression model as a neural network.

First, the wind condition estimation unit 13 reads the learned regression model f from the storage unit 20 (process S41). Next, the wind condition estimation unit 13 reads the reference data X for the prediction target period from the storage unit 20 (process S42). X is an array, and a consecutive number i starting from 0 is a subscript, the i-th element of X is X[i], and the number of elements of X is length(X). Also, f is a function, the argument of f is the reference data value at a certain time, and the return value y of f is the wind condition estimation result. Note that f returns different y even if the same argument is input. A consecutive number j starting from 0 is taken as a counter variable.

The wind condition estimation unit 13 sets i to 0 (processing S43), sets j to 0 (processing S44), and stores y in a two-dimensional array Y having i and j as subscripts (processing S45). . Then, the wind condition estimation unit 13 determines whether or not j is less than 1000 (process S46). If j is less than 1000 (process S46, YES), returns to process S45. S46, NO), and proceeds to processing S47. In addition, although processing S46 exemplifies 1000 as a threshold value, the threshold value may be set arbitrarily.

Then, the wind condition estimating unit 13 aggregates the obtained Y using the aggregate function g to obtain z (process S47). where g is a function and the arguments of g are Y and i. Also, the return value z of g is a structure that stores statistics of wind condition estimation results (1,000 cases in FIG. 4) for all j with the same i in Y. FIG. Statistics include histograms representing mean, maximum, minimum, standard deviation, and probability of occurrence. It may also include kurtosis and skewness. It may also include probability density functions estimated by non-parametric techniques such as kernel density estimation.

Then, the wind condition estimation unit 13 determines whether i is less than length(X) (process S48), and if i is less than length(X) (process S48, YES), returns to process S44, and i If it is equal to or greater than length(X) (process S48, NO), the series of processes is terminated. That is, the processes S44 to S47 are performed for different i to obtain the wind condition estimation result Z including uncertainty for X. Z is an array with i as a subscript, and the length of the array is the same as X.

FIG. 5 is an example in which Monte Carlo dropout is applied at the time of estimating wind speed, and is a diagram showing an example of display of the uncertainty of wind speed estimation results. Output is performed 1,000 times at each time, the average is indicated by a solid line, and the maximum and minimum ranges are indicated by gray areas.

For example, the appearance probability at 5 o'clock was distributed with an average of 4.8 (m/s), a maximum of 5.4 (m/s), and a minimum of 4.2 (m/s). The appearance probability at 10 o'clock was distributed with an average of 9.4 (m/s), a maximum of 10.4 (m/s), and a minimum of 7.9 (m/s). The appearance probability at 18:00 was distribution with an average of 4.0 (m/s), a maximum of 4.5 (m/s), and a minimum of 3.1 (m/s). If the distribution is broad, interpret that the uncertainty is large. At 5:00, 10:00, and 18:00 in FIG. 5, it is interpreted that the uncertainty at 10:00 is greater than at 5:00 and 18:00.

A value obtained by adding the standard deviation to the average may be displayed instead of the maximum, and a value obtained by subtracting the standard deviation from the average may be displayed instead of the minimum. When outliers are included in the estimation results, the difference between the maximum/minimum and the mean becomes large compared to the spread of the distribution, which unnecessarily overestimates the uncertainty.

This time, the values for one hour per day are illustrated, and since the calculation load was small, 1,000 samples were collected at each time to calculate the average, maximum, and minimum values. When the calculation load is large, such as when estimating 20 years or when estimating 1-minute values, it is not necessary to collect the same number of samples at all times. If the wind speed is low in the first few samples, or if the standard deviation is small, the calculation may be terminated. That is, the wind resource estimator 13 has a function of terminating repeated calculations when the average or standard deviation of the estimated results is smaller than a predetermined value when calculating a plurality of wind resource estimation results. For example, if the first 100 samples have an average of 1 (m/s) and a standard deviation of 0.1 (m/s), the calculation may be terminated. The relationship between wind speed (m/s) and output (kW) in wind power generation equipment is called a power curve. On the power curve, the wind speed 0 (m/s) to 2 (m/s) may have an output of 0 (kW), and when the wind speed is around 1 (m/s), the output (kW) is all This is because it is 0 (kW) and is not affected by wind speed estimation errors.

For example, when estimating 10-minute values for 20 years, there are approximately 1,000,000 points of time (=20×365×24×60/10) to be estimated. If 1,000 samples are uniformly collected at each time, approximately 1 billion samples (=approximately 1,000,000×1,000) are required to be estimated, resulting in a large computational load. This process, which can reduce the calculation load without lowering the calculation accuracy while paying attention to the properties of the power curve, is preferable from the viewpoint of shortening the calculation time.

(Screen display processing)
FIG. 6 is a diagram showing a display example of the display unit 40 of the wind condition estimation device 100. As shown in FIG. The screen display processing S<b>3 is executed by the screen display processing unit 14 . The screen display processing unit 14 acquires the wind condition estimation result 24 from the storage unit 20 and displays it on the display unit 40 . The display unit 40 includes a reference data registration button 61 , a calibration data registration button 62 , a regression model setting form 63 , and a wind condition estimation result output button 64 . A function of displaying the registered reference data 22, the registered calibration data 21, and the estimation result 24 may be provided.

This FIG. 6 shows the situation in which the wind speed (m/s) is backcasting. Graphs 65, 66, and 67 on the right side show the reference data 22, the calibration data 21, and the wind condition estimation result 24 in order from the top. Assuming that the period included in the calibration data 21 is the first period and the period included in the reference data 22 is the second period, the first period and the second period overlap, and the second period overlaps. It can be verified that the start date is earlier than the start date of the first period. Moreover, when the period obtained by subtracting the first period from the second period is defined as the third period, it can be confirmed that the wind condition estimation result 24 includes the third period. In the wind condition estimation result 24, the predicted value is expressed by a solid line, and the possible range of the predicted value is expressed by graying.

Note that when forecasting, the end of the reference data 22 is after the end of the calibration data 21, and in FIG. 6, the right side of the calibration data 21 (the right side of the graph 66) is the third period.

Business operators of renewable energy power generation facilities are expected to use backcasting at the business planning stage before construction of power generation facilities, and forecasting at the business operation stage after construction.

As shown in FIG. 6, if wind conditions including uncertainty can be predicted, both backcasting and forecasting can take into account the uncertainty of power generation calculation and electricity sales income. In backcasting, when evaluating profitability and the certainty of loan repayment, downside risks can be evaluated, and only power generation facilities that will not hinder loan repayment will be commercialized or a reserve fund will be prepared. , can be used.

Forecasts can also be used to plan bidding volumes for renewable energy power plants that use the Feed-In-Premium Scheme (FIP). In the case of a power generation business using FIP, the power generation company needs to predict the amount of power generation and bid, but if the amount of power generated by the successful bid differs from the actual amount of power generated, it is necessary to settle the difference. Since the unit price for liquidation may differ between the upside error and the downside error, if the range and distribution that the forecast value can take can be understood, the forecast value will not be used as the bid amount, and it will be more economical. The value can be the bid amount.

This device can obtain the wind condition estimation result 24 by preparing reference data 22 and calibration data 21 for a required period regardless of the application of backcasting, forecasting, or interpolation. From the standpoint of ease of learning by business operators, it is preferable to be able to obtain wind resource estimation results with the same device from the business planning stage to the business operation stage.

The regression model form configures the regression model implemented in the device. In addition to the regression formula to be used, variables of the regression formula may be set. The wind condition estimation result 24 is output by pressing the wind condition estimation result 24 output button. In the case of FIG. 6, as the wind condition estimation result 24, there is a method of outputting the time series of the wind speed in the third period as a comma-separated character string (CSV) file.

10 processing unit 11 data input unit 12 regression model learning unit 13 wind condition estimation unit 14 screen display processing unit 20 storage unit 21 calibration data 22 reference data 23 learned regression model 24 wind condition estimation result (estimation result)
30 Input unit 40 Display unit 50 Communication unit 61 Reference data registration button 62 Calibration data registration button 63 Regression model setting form 64 Wind condition estimation result output button 100 Wind condition estimation device S1 Regression model learning processing S2 Wind condition estimation processing S3 Screen display processing S40 Uncertainty estimation processing

In order to solve the above-mentioned problems, the wind resource estimation device of the present invention provides calibration data indicating the wind resource measured in the vicinity of the installation location of the wind power generation equipment, uniform measurement, and measurement of the observation location unchanged within the observation period. a storage unit storing reference data indicating wind conditions obtained by a method or an analysis method; a period included in the calibration data as a first period; and a period included in the reference data as a second period a regression model learning unit for learning a regression model based on the calibration data and the reference data, where a period obtained by subtracting the first period from the second period is a third period; It is characterized by comprising a wind condition estimating unit for estimating the wind condition including the uncertainty of the estimation result based on the reference data and the regression model in the third period. Other aspects of the present invention are described in embodiments below.

Claims (9)

a storage unit storing calibration data indicating the wind conditions measured near the installation site of the wind power generation facility and reference data indicating the wind conditions obtained by a homogeneous measurement method or analysis method;
A period included in the calibration data is defined as a first period, a period included in the reference data is defined as a second period, and a period obtained by subtracting the first period from the second period is defined as a third period. a regression model learning unit that learns a regression model based on the calibration data and the reference data;
A wind condition estimating device comprising: a wind condition estimating unit that estimates a wind condition including uncertainty of an estimation result based on the reference data and the regression model in the third period. The wind condition estimation device according to claim 1, wherein the uncertainty of the estimation result is expressed as the appearance probability of the estimation result. 2. The wind condition estimation apparatus according to claim 1, wherein said calibration data are observation values obtained by a wind condition observation tower, a Doppler lidar, and a Doppler soda. The wind condition estimation device according to claim 1, wherein the regression model learning unit learns the regression model after converting the input wind speed and wind direction into a north-south component and an east-west component of the wind speed. The wind condition estimation device according to claim 1, wherein the regression model learning unit has a function of learning after changing the time granularity of the calibration data and the reference data. The regression model is a neural network having an input layer, an intermediate layer, and an output layer,
The wind condition according to claim 1, wherein the input layer inputs the reference data and adjusts parameters so as to reduce the error between the output of the output layer and the calibration data. estimation device. 6. The wind resource estimation unit has a function of randomly changing the parameters of the regression model, and calculates the uncertainty of the estimation result by calculating and aggregating a plurality of estimation results. The wind condition estimation device according to . The wind condition estimator according to claim 7, characterized in that, when calculating a plurality of estimation results, the wind condition estimation unit has a function of terminating repeated calculations when an average or standard deviation of the estimation results is smaller than a predetermined value. situation estimation device. A first step of learning a regression model based on calibration data indicating wind conditions measured near the installation location of the wind power generation facility and reference data indicating wind conditions obtained by a homogeneous measurement method or analysis method; ,
A period included in the calibration data is defined as a first period, a period included in the reference data is defined as a second period, and a period obtained by subtracting the first period from the second period is defined as a third period. a second step of estimating wind conditions including uncertainties of estimation results based on the reference data and the regression model in the third period.

Images