A deep learning-based progressive short-term wind power prediction method and system

Abstract

The application discloses a kind of progressive short-term wind power prediction method and system based on deep learning, comprising: according to the latitude and longitude of wind field's wind tower and numerical weather prediction grid, obtain the predicted wind speed sequence in the wind tower prediction grid, observed power sequence and observed wind speed sequence;In turn according to spatial correction MLP network, time difference value MLP network, GRU network and the fourth MLP network, obtain predicted wind speed sequence, difference value predicted power sequence, second predicted power sequence and accurate predicted power sequence;Further, according to the network after training, wind power is predicted.The application connects the structure after predicting power by two paths, greatly improves the prediction accuracy;At the same time, using the black box of neural network, using deep learning modeling, is conducive to breaking through the existing meteorological science development restriction. Only the numerical weather prediction disclosed can be used to model the wind field specifically, without spending more computing power.

Description

Technical Field

[0001] This invention relates to the field of wind power prediction technology, and in particular to a progressive short-term wind power prediction method and system based on deep learning. Background Technology

[0002] Wind power generation is crucial. Studies show that wind power consumes only 1% of the carbon emissions of thermal power generation per kilowatt-hour. For grid dispatching departments to make reasonable power generation plans and improve the economy, safety, and reliability of grid operation, wind power forecasting is a vital component. Wind speed forecasting is the foundation of wind power forecasting.

[0003] Traditional wind speed forecasts have often relied on numerical weather prediction (NMR), which simplifies and approximates the fundamental equations of atmospheric motion and incorporates meteorological and physical laws and boundary constraints. NMR wind speed forecasts typically have a temporal resolution of 3 hours or 1 hour and a spatial resolution of 9 km x 9 km or 12.5 km x 12.5 km. For wind farms, a power forecast temporal resolution of 15 minutes is generally required, and wind speed forecasts must be located at the wind turbine locations. This limits forecast accuracy to existing meteorological theories; the temporal and spatial resolutions of NMR generally do not meet the requirements for wind farm power forecasting, resulting in low accuracy. Improving both temporal and spatial resolution requires significantly more computing power and time, which in turn means consuming more carbon. Summary of the Invention

[0004] This invention provides a progressive short-term wind power prediction method and system based on deep learning to overcome the above-mentioned technical problems.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] A progressive short-term wind power prediction method and system based on deep learning includes the following steps:

[0007] S1: Based on the latitude and longitude of the wind measurement tower and the numerical weather prediction grid, determine the wind measurement tower location grid in the numerical weather prediction grid, and obtain the wind measurement tower prediction grid required for wind power prediction based on the wind measurement tower location grid.

[0008] S2: Take the first time point T1, the second time point T2, and the third time point T3 in chronological order.

[0009] in, definition:

[0010] The time interval corresponding to the first time series is [T1, T3], and the time interval between adjacent time points in the first time series is set as the first time threshold.

[0011] The time interval corresponding to the second time series is [T2, T3], and the time interval between adjacent time points in the second time series is the first time threshold.

[0012] The time interval corresponding to the third time series is [T2, T3], and the time interval between adjacent time points in the third time series is the second time threshold;

[0013] Among them, the

[0014] Where 'a' is the time conversion factor;

[0015] The fourth time series corresponds to the time interval [T4, T2], where, Wherein, T4 is the fourth time point, and T4 is earlier than T2 and later than T1; the time interval between adjacent time points in the fourth time series is the first time threshold;

[0016] S3: Obtain the predicted wind speed of the latitude and longitude grid points in the wind measurement tower prediction grid; and normalize the predicted wind speed to obtain the normalized predicted wind speed;

[0017] The system acquires observation data from latitude and longitude grid points within the meteorological tower's prediction grid; and normalizes the observation data to obtain normalized observation data. The observation data includes observed power and observed wind speed, and the normalized observation data includes normalized observed power and normalized observed wind speed. Based on the normalized predicted wind speed, the system acquires the predicted wind speed sequence corresponding to a first time series for all latitude and longitude grid points within the meteorological tower's prediction grid; based on the normalized observed power, the system acquires the observed power sequence corresponding to a third time series at the meteorological tower's location; and based on the normalized observed wind speed, the system acquires the observed wind speed sequence corresponding to a fourth time series for all latitude and longitude grid points within the meteorological tower's prediction grid.

[0018] The predicted wind speed includes the i-th height H of the latitude and longitude grid point from the ground. i The wind speed at the location; where i = 1, 2, ...; the observed power includes the i-th height H of the latitude and longitude grid point from the ground. i The observed power at the location; the observed wind speed includes the i-th height H of the latitude and longitude grid point from the ground. i Observed wind speed at the location;

[0019] S4: Input the predicted wind speed sequence into the spatial correction MLP network to obtain the predicted wind speed sequence corresponding to the second time series for the wind tower location grid.

[0020] S5: Input the predicted wind speed sequence into the time difference MLP network to obtain the difference predicted power sequence corresponding to the third time sequence of the wind tower location;

[0021] S6: Based on the observed wind speed sequence, obtain the second predicted power sequence corresponding to the third time sequence of the wind measurement tower location using the GRU network;

[0022] S7: Concat the difference predicted power sequence and the second predicted power sequence to obtain the concatenated predicted power sequence corresponding to the third time series of the wind tower location;

[0023] S8: Input the connected predicted power sequence into the fourth MLP network to obtain the accurate predicted power sequence corresponding to the third time series of the wind tower location;

[0024] S9: Based on the observed power sequence and the accurate predicted power sequence, train the spatial correction MLP network, temporal difference MLP network, GRU network and fourth MLP network based on the loss function; then predict the wind power based on the trained spatial correction MLP network, temporal difference MLP network, GRU network and fourth MLP network.

[0025] Furthermore, the first time threshold = 1 hour = 60 minutes; the second time threshold = 15 minutes.

[0026] Furthermore, the method for obtaining the second predicted power sequence corresponding to the third time series is as follows:

[0027] The observed wind speed sequence corresponding to the fourth time series is input into the GRU network to obtain the first-stage predicted power sequence corresponding to the third time series; the first-stage predicted power sequence includes the first time interval. The first and second predicted power sequences within the time interval. The first and second predicted power sequences within the first stage; the time interval between adjacent time points in the first stage predicted power sequence is the second time threshold;

[0028] Time interval The observed wind speed sequence within the specified time interval is input into the GRU network to obtain the wind speed in the first and second time intervals. The second stage predicts the power sequence within the time interval, and the first, second time intervals are included. The first,1 predicted power sequence within the time interval is replaced with the second-stage predicted power sequence; the second-stage predicted power sequence includes the second,1 time interval. The second and first predicted power sequences within the second and second time intervals. The second, second predicted power sequence within the second stage; the time interval between adjacent time points in the second stage predicted power sequence is the second time threshold;

[0029] Time interval The observed wind speed sequence within the specified time interval is input into the GRU network to obtain the wind speed in the 2nd and 1st time intervals. The third stage predicted power sequence within the time interval, and the second and first time intervals. The second and first predicted power sequences within the time interval are replaced with the third-stage predicted power sequence; the third-stage predicted power sequence includes the third and first time intervals. The predicted power sequences of time intervals 3 and 1, and time intervals 3 and 2. The third and second predicted power sequences within the third stage; the time interval between adjacent time points in the predicted power sequence of the third stage is the second time threshold;

[0030] And so on, until...

[0031] Time interval The observed wind speed sequence within the specified time interval is input into the GRU network to obtain the wind speed in the (n-1)th, 1st time interval. The predicted power sequence for the nth stage within the time interval, and the (n-1)th, 1st time interval The (n-1)th, 1st predicted power sequence is replaced with the nth stage predicted power sequence to obtain the final second predicted power sequence;

[0032] The predicted power sequence for the nth stage includes the nth, 1st time interval. The nth,1st predicted power sequence and the nth,2nd time interval. The nth and 2nd predicted power sequences within the nth stage; the time interval between adjacent time points in the nth stage predicted power sequence is the second time threshold; where n is the set number of iterations.

[0033] Furthermore, after S2, the method further includes: obtaining the power values ​​of the wind measurement towers corresponding to all time points in a time series with a time interval of [T1, T3] and a time interval between adjacent time points equal to a second time threshold, and changing data with power values ​​less than 0 to equal to 0.

[0034] Furthermore, after S2, the method further includes: obtaining the wind speed values ​​within the wind tower prediction grid corresponding to all time points in a time series with a time interval of [T1, T3] and a time interval between adjacent time points equal to a second time threshold.

[0035] If the wind speed value satisfies:

[0036] The wind speed value is greater than the cut-in wind speed and less than the cut-out wind speed, and the power of the fan is 0;

[0037] Or the wind speed value is greater than the rated wind speed of the fan, and the power of the fan is lower than the set power threshold.

[0038] Or the wind speed value is less than the rated wind speed of the fan, and the power of the fan is not equal to 0;

[0039] If the wind speed value is abnormal, the first time point T1, the second time point T2, and the third time point T3 are taken again in chronological order until the wind speed value is no longer abnormal, so that all values ​​within the time interval [T1, T3] are valid.

[0040] A prediction system based on a deep learning-based progressive short-term wind power prediction method includes: an observation data acquisition module, a numerical weather prediction request module, a power prediction training module, a power prediction module, and a data transmission module.

[0041] The observation data acquisition module is used to acquire the observed power sequence and observed wind speed sequence corresponding to the time series with the second time interval within the prediction grid of the wind tower; to acquire the observed power sequence corresponding to the third time series at the location of the wind tower; the observed wind speed sequence of all latitude and longitude grid points within the prediction grid of the wind tower; and the observed wind speed sequence with the interval from the Nth historical day to the current time as the first time interval.

[0042] The numerical weather prediction request module is used to obtain the predicted wind speed of the latitude and longitude grid points in the wind tower prediction grid;

[0043] The power prediction training module is used to train the spatial correction MLP network, the temporal difference MLP network, the GRU network, and the fourth MLP network based on the observed power sequence corresponding to the third time series at the location of the wind tower, the observed wind speed sequence of all latitude and longitude grid points within the wind tower prediction grid, and the predicted wind speed.

[0044] The power prediction module is used to obtain the future predicted power sequence from the current time to the Nth future time, with the interval between the historical Nth day and the current time as the first time interval, and the predicted wind speed sequence between the historical Nth day and the Nth future time, with the interval between adjacent times as the first time threshold, from the latitude and longitude grid points of the meteorological tower prediction grid, in order to predict the power value of the second time threshold for the time interval of the Nth future time.

[0045] The data transmission module is used to upload the future predicted power sequence from the current time to the Nth day in the future, with the interval between adjacent times being a second time threshold, from the current time in the wind measurement tower location grid to the wind farm dispatch center.

[0046] Beneficial Effects: This invention provides a progressive short-term wind power prediction method and system based on deep learning. Using only publicly available numerical weather prediction results, it predicts the power of the wind measurement tower by measuring the wind speed at different heights of all latitude and longitude grid points within the tower's prediction grid, obtaining a difference prediction power sequence. Simultaneously, a GRU network is used for progressive prediction to obtain a second prediction power sequence. The two prediction power sequences are then connected and predicted again to obtain a precise prediction power sequence. The neural network is trained based on the precise prediction power sequence and the actual observed power sequence, and finally, the trained network is used to predict the future wind power of the wind measurement tower. By connecting the structures after predicting the power along two paths, the prediction accuracy is greatly improved, which is beneficial for improving wind field power prediction. Furthermore, utilizing the black-box nature of neural networks and employing deep learning modeling helps overcome the limitations of current meteorological science development. Only publicly available numerical weather predictions are needed for specific wind field modeling, without requiring excessive computing power. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a flowchart of the deep learning-based progressive short-term wind power prediction method of the present invention.

[0049] Figure 2 This is a flowchart of a deep learning-based progressive short-term wind power prediction method in an embodiment of the present invention.

[0050] Figure 3 This is a structural block diagram of a deep learning-based progressive short-term wind power prediction system in an embodiment of the present invention.

[0051] Figure 4 This is a schematic diagram of the wind speed-power curve under ideal conditions in an embodiment of the present invention;

[0052] Figure 5 This is a schematic diagram of the wind speed-power curve under factors such as sensor failure and power grid rationing in an embodiment of the present invention;

[0053] Figure 6This is a schematic diagram of the MLP network structure in an embodiment of the present invention;

[0054] Figure 7 This is a schematic diagram of the GRU network prediction process in an embodiment of the present invention;

[0055] Figure 8 This is a schematic diagram of the input and output of a GRU network in an embodiment of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] This embodiment provides a progressive short-term wind power prediction method and system based on deep learning, such as... Figure 1 and Figure 2 As shown, it includes the following steps:

[0058] S1: Based on the latitude and longitude of the wind measurement tower and the numerical weather prediction grid, determine the wind measurement tower location grid in the numerical weather prediction grid, and obtain the wind measurement tower prediction grid required for wind power prediction based on the wind measurement tower location grid.

[0059] Specifically, numerical weather prediction is presented in the form of a latitude and longitude grid. This grid contains multiple latitude and longitude grid points. The latitude and longitude grid point corresponding to the wind turbine's location is the obtained wind tower location grid. Generally, this wind tower location grid is extended upwards, downwards, left, and right by no more than three latitude and longitude grid points to obtain the final wind tower prediction grid (m, n), where m is the length of the wind tower prediction grid and n is the width. The specific number of extended grids depends on the actual terrain of the wind tower's location, such as the height of the surrounding hills and the distance between the hills and the wind tower. The wind tower prediction grid includes multiple latitude and longitude grid points.

[0060] S2: Select the first time point T1, the second time point T2, and the third time point T3 in chronological order.

[0061] in, That is, the duration between the first time point T1 and the second time point T2 is equal to the duration between the second time point T2 and the third time point T3; Definition:

[0062] The time interval corresponding to the first time series is [T1, T3], and the time interval between adjacent time points in the first time series is set as the first time threshold.

[0063] The time interval corresponding to the second time series is [T2, T3], and the time interval between adjacent time points in the second time series is the first time threshold.

[0064] The time interval corresponding to the third time series is [T2, T3], and the time interval between adjacent time points in the third time series is the second time threshold;

[0065] Among them, the

[0066] Where 'a' is the time conversion factor;

[0067] The fourth time series corresponds to the time interval [T4, T2], where, Wherein, T4 is the fourth time point, and T4 is earlier than T2 and later than T1; the time interval between adjacent time points in the fourth time series is the first time threshold;

[0068] Specifically, the time interval between the fourth time point T4 and the second time point T2 is the time series of the first time threshold; wherein, the duration between the first time point T1 and the fourth time point T4 is equal to the duration between the fourth time point T4 and the second time point T2, that is, the fourth time point T4 is between the first time point T1 and the second time point T2, and the duration between the first time point T1 and the second time point T2 is equal;

[0069] Specifically, the duration between the start and end times of the second time series is 0.5 times the duration between the start and end times of the first time series.

[0070] S3: Obtain the predicted wind speed of the latitude and longitude grid points in the wind measurement tower prediction grid; and normalize the predicted wind speed to obtain the normalized predicted wind speed;

[0071] The system acquires observation data from latitude and longitude grid points within the meteorological tower's prediction grid; and normalizes the observation data to obtain normalized observation data. The observation data includes observed power and observed wind speed, and the normalized observation data includes normalized observed power and normalized observed wind speed. Based on the normalized predicted wind speed, the system acquires the predicted wind speed sequence corresponding to a first time series for all latitude and longitude grid points within the meteorological tower's prediction grid; based on the normalized observed power, the system acquires the observed power sequence corresponding to a third time series at the meteorological tower's location; and based on the normalized observed wind speed, the system acquires the observed wind speed sequence corresponding to a fourth time series for all latitude and longitude grid points within the meteorological tower's prediction grid.

[0072] The predicted wind speed includes the i-th height H of the latitude and longitude grid point from the ground. i The wind speed at the location; where i = 1, 2, ...; the observation data includes the i-th altitude H of the latitude and longitude grid point from the ground. i The data is located at a specific point; specifically, the observation power includes the i-th altitude H of the latitude and longitude grid point from the ground. i The observed power at the location; the observed wind speed includes the i-th height H of the latitude and longitude grid point from the ground. i Observed wind speed at the location;

[0073] Specifically, the forecast wind speeds at different heights above the ground are extracted from the numerical weather forecasts on the forecast grid (m,n) of the wind tower. For example, the forecast wind speeds at heights of 10m, 100m, and 200m above the ground are extracted from the latitude and longitude grid points in the forecast grid of the wind tower, and then normalized to obtain the normalized forecast wind speeds.

[0074]

[0075] Where, x y Here, x represents the normalized forecast wind speed; x is the forecast wind speed at each grid point. min The minimum wind speed is set, typically 0; x max The maximum wind speed is set, which is generally the maximum wind speed of the past 5-10 years;

[0076] Specifically, the method for obtaining normalized observation data is the same as the method for obtaining normalized forecast wind speed.

[0077] S4: Input the predicted wind speed sequence into the spatial correction MLP network to obtain the predicted wind speed sequence corresponding to the second time series for the wind tower location grid.

[0078] Preferably, the predicted wind speed sequence corresponding to the first historical time series of all latitude and longitude grid points within the wind measurement tower prediction grid is input into the spatial correction MLP network to obtain the predicted wind speed sequence corresponding to the second time series.

[0079] Specifically, in this embodiment, the historical observation data within the wind field prediction grid of the wind measurement tower, namely the observed wind speed sequence and the observed power sequence, are all selected from the time series observation data within a complete year. This can minimize the interference of seasonal differences on the prediction results and ensure the accuracy of the prediction results of the spatially corrected MLP network.

[0080] The predicted wind speed sequence includes predicted wind speed sequences at different ground heights corresponding to the second historical time series of the meteorological tower location grid.

[0081] Specifically, during training, the spatial correction MLP network, when correcting forecasts for the next N days, inputs several t*s matrices, where t represents the number of times in the first time series, and s represents the total number of heights from which wind speeds are extracted in the meteorological tower location grid. In this embodiment, historical wind speeds are extracted at heights of 10m, 100m, and 200m above the ground, therefore s is set to 3, resulting in a t*3 matrix as the input. Each t*3 matrix represents a sample data point input to the spatial correction MLP network, and multiple sample data points form the training set of the network. This sample data includes data from all latitude and longitude grid points within the meteorological tower prediction grid. The output is a 0.5*t*3 matrix, containing only data from the meteorological tower location grid. The spatial correction MLP network corrects the data from all latitude and longitude grid points within the historical meteorological tower prediction grid to the data from a single meteorological tower location grid. The meteorological tower location grid is a latitude and longitude grid point within the meteorological tower prediction grid. The spatial correction MLP network is an existing neural network that can implement the technical solution of this patent, and will not be described in detail here.

[0082] S5: Input the predicted wind speed sequence into the time difference MLP network to obtain the difference predicted power sequence corresponding to the third time sequence of the wind tower location;

[0083] Preferably, the second time threshold is 15 minutes;

[0084] Specifically, the spatial correction MLP network and the temporal interpolation MLP network in this embodiment are both conventional techniques in the field, and their structure is a two-layer MLP network (fully connected neural network), such as... Figure 6 As shown; therefore, it will not be described in detail here.

[0085] S6: Based on the observed wind speed sequence, obtain the second predicted power sequence corresponding to the third time sequence of the wind measurement tower location using the GRU network;

[0086] The fourth time series is the time series between the fourth time point T4 and the second time point T2, with the time interval being the first time threshold.

[0087] Preferably, the method for obtaining the second predicted power sequence corresponding to the third time series is as follows: Figure 7 As shown:

[0088] The observed wind speed sequence corresponding to the fourth time series is input into the GRU network for the first time to obtain the first-stage predicted power sequence corresponding to the third time series; the first-stage predicted power sequence includes the first time interval. The first and second predicted power sequences within the time interval. The first and second predicted power sequences within the first stage; the time interval between adjacent time points in the first stage predicted power sequence is the second time threshold;

[0089] In essence, it inputs the observed wind speed sequence corresponding to the fourth time series within the time interval [T4,T2] into the GRU network and outputs the predicted power sequence within the time interval [T2,T3].

[0090] Time interval The observed wind speed sequence within the time interval is input into the GRU network for the second time to obtain the wind speed in the first and second time intervals. The second stage predicts the power sequence within the time interval, and the first, second time intervals are included. The first,1 predicted power sequence within the time interval is replaced with the second-stage predicted power sequence; the second-stage predicted power sequence includes the second,1 time interval. The second and first predicted power sequences within the second and second time intervals. The second, second predicted power sequence within the second stage; the time interval between adjacent time points in the second stage predicted power sequence is the second time threshold;

[0091] Time interval The observed wind speed sequence within the time interval is input into the GRU network for the third time to obtain the wind speed in the second and first time intervals. The third stage predicted power sequence within the time interval, and the second and first time intervals. The second and first predicted power sequences within the time interval are replaced with the third-stage predicted power sequence; the third-stage predicted power sequence includes the third and first time intervals. The predicted power sequences of time intervals 3 and 1, and time intervals 3 and 2. The third and second predicted power sequences within the third stage; the time interval between adjacent time points in the predicted power sequence of the third stage is the second time threshold;

[0092] And so on, until...

[0093] Time interval The observed wind speed sequence within the specified time interval is input into the GRU network for the nth time to obtain the wind speed in the (n-1)th time interval. The predicted power sequence for the nth stage within the time interval, and the (n-1)th, 1st time interval The (n-1)th, 1st predicted power sequence is replaced with the nth stage predicted power sequence to obtain the final second predicted power sequence;

[0094] The predicted power sequence for the nth stage includes the nth, 1st time interval. The nth,1st predicted power sequence and the nth,2nd time interval. The nth and 2nd predicted power sequences within the nth stage; the time interval between adjacent time points in the nth stage predicted power sequence is the second time threshold; where n is the set number of iterations;

[0095] This embodiment progressively fuses the predicted structure by increasing the temporal granularity from coarse to fine, resulting in more accurate predictions. The GRU network used is a standard GRU structure, which is existing technology in the field; its structure will not be described in detail here. This embodiment only uses a GRU network to implement the technical solution of this invention. Figure 8 As mentioned above, x in the figure t For the input at this current moment, y t For the predicted result, h t-1 h represents the state at the previous time step. t Output the current state;

[0096] S7: Concat the difference predicted power sequence and the second predicted power sequence to obtain the third time series corresponding to the location of the wind tower;

[0097] Specifically, the difference predicted power sequence and the final second predicted power sequence output by the GRU network are concatenated in the last dimension to obtain the concatenated predicted power sequence; the concatenation method is a conventional technique in the field and will not be described in detail here.

[0098] S8: Input the connected predicted power sequence into the fourth MLP network to obtain the accurate predicted power sequence corresponding to the third time series of the wind tower location;

[0099] Specifically, the fourth MLP network of the present invention is an MLP network using a linear layer, which is a conventional MLP network in the field.

[0100] S9: Based on the observed power sequence (historical true power sequence) and the accurate predicted power sequence, train the spatially corrected MLP network, temporal difference MLP network, GRU network, and fourth MLP network using a loss function; the loss function is the MSE function. Backpropagate the network errors using gradients to update the network parameters; this yields the final trained spatially corrected MLP network, temporal difference MLP network, GRU network, and fourth MLP network. Use these trained networks to predict wind power.

[0101] Specifically, based on the predicted wind speed sequence from the historical Nth day to the future Nth day, with the interval between adjacent times being the first time threshold, and the observed wind speed sequence from the historical Nth day to the current time, using the trained spatial correction MLP network, temporal difference MLP network, GRU network, and fourth MLP network, the future predicted power sequence from the current time to the future Nth day, with the interval between adjacent times being the second time threshold, is obtained in the wind tower location grid, so as to predict the power value of the second time threshold for the time interval of the future N days.

[0102] This invention addresses the inherent characteristics of wind fields by using the GRU model to progressively predict future power based on historical wind speed observations. It also utilizes historical wind speed forecasts from numerical weather prediction for prediction and finally integrates the two results to improve prediction accuracy.

[0103] Specifically, wind turbines have cut-in wind speed, cut-out wind speed, and rated wind speed, all of which are marked at the factory. A wind turbine can only generate electricity normally when the wind speed is between the cut-in and cut-out wind speeds. When the wind speed is greater than or equal to the rated wind speed but less than the cut-out wind speed, the wind turbine's power output is the rated power. Therefore, the wind speed-power curve under ideal conditions is as follows: Figure 4 As shown; however, due to a series of practical problems such as sensor failure and power grid rationing, the wind speed-power curve often presents as shown in the figure. Figure 5 As shown:

[0104] In the embodiments of the present invention, the target wind turbine to be repaired for power anomalies has the following specific characteristics: when the wind speed of the target wind turbine is between the cut-in and cut-out wind speeds, the power is 0; when the wind speed is greater than the rated wind speed, the power is relatively small; when the wind speed is less than the cut-in wind speed, the power is not 0; the anomaly value is denoted as NAN.

[0105] When the predicted power value is abnormal, if the power used for network training contains anomalies, the network will learn based on the abnormal data, affecting the final prediction accuracy. Therefore, it is necessary to eliminate abnormal situations. The method used in this embodiment is as follows:

[0106] Preferably, after step S2, the method further includes: acquiring the wind tower power value corresponding to a time point in a fifth time series with a time interval of [T1, T3] and a time interval between adjacent time points equal to a second time threshold, and changing data with power values ​​less than 0 to equal 0. The wind tower power value includes the i-th height H of the latitude and longitude grid point from the ground. i The power value at that location.

[0107] Following S2, the method further includes: acquiring the wind speed values ​​within the wind tower prediction grid corresponding to all time points in a time series with a time interval of [T1, T3] and a time interval between adjacent time points equal to a second time threshold.

[0108] If the wind speed value satisfies:

[0109] The wind speed value is greater than the cut-in wind speed and less than the cut-out wind speed, and the power of the fan is 0;

[0110] Or the wind speed value is greater than the rated wind speed of the fan, and the power of the fan is lower than the set power threshold.

[0111] Or the wind speed value is less than the rated wind speed of the fan, and the power of the fan is not equal to 0;

[0112] If the wind speed value is abnormal, the first time point T1, the second time point T2, and the third time point T3 are taken again in chronological order until the wind speed value is no longer abnormal, so that all values ​​within the time interval [T1, T3] are valid.

[0113] Specifically, the abnormal data mentioned above is cleaned up; finally, clean wind power data is obtained, and wind power is predicted to improve the accuracy of the prediction results.

[0114] This embodiment also discloses a progressive short-term wind power prediction system based on deep learning. Figure 3 As shown, it includes: an observation data acquisition module, a numerical weather prediction request module, a power prediction training module, a power prediction module, and a data transmission module.

[0115] The observation data acquisition module is used to acquire the observed power sequence and observed wind speed sequence corresponding to the time series with the second time interval within the prediction grid of the wind tower; to acquire the observed power sequence corresponding to the third time series at the location of the wind tower; the observed wind speed sequence of all latitude and longitude grid points within the prediction grid of the wind tower; and the observed wind speed sequence with the interval from the Nth historical day to the current time as the first time interval.

[0116] Specifically, the observation data acquisition module in this embodiment records and stores the wind speed and power observed by the wind tower at 15-minute intervals;

[0117] The numerical weather forecast request module is used to download and store numerical weather forecasts, namely the forecast wind speeds of latitude and longitude grid points in the wind tower prediction grid.

[0118] The power prediction training module is used to train the spatial correction MLP network, the temporal difference MLP network, the GRU network, and the fourth MLP network based on the observed power sequence corresponding to the third time series at the location of the wind tower, the observed wind speed sequence of all latitude and longitude grid points within the wind tower prediction grid, and the predicted wind speed.

[0119] Specifically, the power prediction training module is used to receive observed power and wind speed data as well as forecast wind speed data from numerical weather prediction, train the prediction model, and at regular intervals, add the latest observed data and numerical weather prediction data to further optimize the model.

[0120] The power prediction module is used to obtain the future predicted power sequence from the current time to the Nth future time in the meteorological tower location grid, based on the observed wind speed sequence of all latitude and longitude grid points in the meteorological tower prediction grid, with the interval between the historical Nth day and the current time as the first time interval, and the predicted wind speed sequence between the historical Nth day and the future Nth day, with the interval between adjacent times as the first time threshold, in order to predict the power value of the second time threshold time interval for the next N days.

[0121] The data transmission module is used to upload the future predicted power sequence from the current time to the Nth day in the future, with the interval between adjacent times being the second time threshold, from the current time in the wind measurement tower location grid to the wind farm dispatch center.

[0122] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A progressive short-term wind power prediction method based on deep learning, characterized in that, Includes the following steps: S1: Based on the latitude and longitude of the wind measurement tower and the numerical weather prediction grid, determine the wind measurement tower location grid in the numerical weather prediction grid, and obtain the wind measurement tower prediction grid required for wind power prediction based on the wind measurement tower location grid. S2: Take the first time point T1, the second time point T2, and the third time point T3 in chronological order. in, definition: The time interval corresponding to the first time series is [T1, T3], and the time interval between adjacent time points in the first time series is set as the first time threshold. The time interval corresponding to the second time series is [T2, T3], and the time interval between adjacent time points in the second time series is the first time threshold. The time interval corresponding to the third time series is [T2, T3], and the time interval between adjacent time points in the third time series is the second time threshold; Among them, the Where 'a' is the time conversion factor; The fourth time series corresponds to the time interval [T4, T2], where, Wherein, T4 is the fourth time point, and T4 is earlier than T2 and later than T1; the time interval between adjacent time points in the fourth time series is the first time threshold; S3: Obtain the predicted wind speed of the latitude and longitude grid points in the wind measurement tower prediction grid; and normalize the predicted wind speed to obtain the normalized predicted wind speed; The system acquires observation data from latitude and longitude grid points within the meteorological tower's prediction grid; and normalizes the observation data to obtain normalized observation data. The observation data includes observed power and observed wind speed, and the normalized observation data includes normalized observed power and normalized observed wind speed. Based on the normalized predicted wind speed, the system acquires the predicted wind speed sequence corresponding to a first time series for all latitude and longitude grid points within the meteorological tower's prediction grid; based on the normalized observed power, the system acquires the observed power sequence corresponding to a third time series at the meteorological tower's location; and based on the normalized observed wind speed, the system acquires the observed wind speed sequence corresponding to a fourth time series for all latitude and longitude grid points within the meteorological tower's prediction grid. The predicted wind speed includes the i-th height H of the latitude and longitude grid point above the ground. i The wind speed at the location; where i = 1, 2, ...; the observed power includes the i-th height H of the latitude and longitude grid point from the ground. i The observed power at the location; the observed wind speed includes the i-th height H of the latitude and longitude grid point from the ground. i Observed wind speed at the location; S4: Input the predicted wind speed sequence into the spatial correction MLP network to obtain the predicted wind speed sequence corresponding to the second time series for the wind tower location grid. S5: Input the predicted wind speed sequence into the time difference MLP network to obtain the difference predicted power sequence corresponding to the third time sequence of the wind tower location; S6: Based on the observed wind speed sequence, obtain the second predicted power sequence corresponding to the third time sequence of the wind measurement tower location using the GRU network; S7: Concat the difference predicted power sequence and the second predicted power sequence to obtain the concatenated predicted power sequence corresponding to the third time series of the wind tower location; S8: Input the connected predicted power sequence into the fourth MLP network to obtain the accurate predicted power sequence corresponding to the third time series of the wind tower location; S9: Based on the observed power sequence and the accurate predicted power sequence, train the spatial correction MLP network, temporal difference MLP network, GRU network and fourth MLP network based on the loss function; then predict the wind power based on the trained spatial correction MLP network, temporal difference MLP network, GRU network and fourth MLP network.

2. The progressive short-term wind power prediction method based on deep learning according to claim 1, characterized in that, The first time threshold = 1 hour = 60 minutes; the second time threshold = 15 minutes.

3. The progressive short-term wind power prediction method based on deep learning according to claim 1, characterized in that, The method for obtaining the second predicted power sequence corresponding to the third time series is as follows: The observed wind speed sequence corresponding to the fourth time series is input into the GRU network to obtain the first-stage predicted power sequence corresponding to the third time series; the first-stage predicted power sequence includes the first time interval. The first and second predicted power sequences within the time interval. The first and second predicted power sequences within the first stage; the time interval between adjacent time points in the first stage predicted power sequence is the second time threshold; Time interval The observed wind speed sequence within the specified time interval is input into the GRU network to obtain the wind speed in the first and second time intervals. The second stage predicts the power sequence within the time interval, and the first, second time intervals are included. The first,1 predicted power sequence within the time interval is replaced with the second-stage predicted power sequence; the second-stage predicted power sequence includes the second,1 time interval. The second and first predicted power sequences within the second and second time intervals. The second, second predicted power sequence within the second stage; the time interval between adjacent time points in the second stage predicted power sequence is the second time threshold; Time interval The observed wind speed sequence within the specified time interval is input into the GRU network to obtain the wind speed in the 2nd and 1st time intervals. The third stage predicted power sequence within the time interval, and the second and first time intervals. The second and first predicted power sequences within the time interval are replaced with the third-stage predicted power sequence; the third-stage predicted power sequence includes the third and first time intervals. The predicted power sequences of time intervals 3 and 1, and time intervals 3 and 2. The third and second predicted power sequences within the third stage; the time interval between adjacent time points in the predicted power sequence of the third stage is the second time threshold; And so on, until... Time interval The observed wind speed sequence within the specified time interval is input into the GRU network to obtain the wind speed in the (n-1)th, 1st time interval. The predicted power sequence for the nth stage within the time interval, and the (n-1)th, 1st time interval The (n-1)th, 1st predicted power sequence is replaced with the nth stage predicted power sequence to obtain the final second predicted power sequence; The predicted power sequence for the nth stage includes the nth, 1st time interval. The predicted power sequence of the nth,1st time interval and the nth,2nd time interval. The nth and 2nd predicted power sequences within the nth stage; the time interval between adjacent time points in the nth stage predicted power sequence is the second time threshold; where n is the set number of iterations.

4. The progressive short-term wind power prediction method based on deep learning according to claim 1, characterized in that, After S2, the method further includes: obtaining the power values ​​of the wind towers corresponding to all time points in a time series with a time interval of [T1, T3] and a time interval between adjacent time points equal to the second time threshold, and changing the data with power values ​​less than 0 to equal 0.

5. The progressive short-term wind power prediction method based on deep learning according to claim 1, characterized in that, Following S2, the method further includes: acquiring the wind speed values ​​within the wind tower prediction grid corresponding to all time points in a time series with a time interval of [T1, T3] and a time interval between adjacent time points equal to a second time threshold. If the wind speed value satisfies: The wind speed value is greater than the cut-in wind speed and less than the cut-out wind speed, and the power of the fan is 0; Or the wind speed value is greater than the rated wind speed of the fan, and the power of the fan is lower than the set power threshold. Or the wind speed value is less than the rated wind speed of the fan, and the power of the fan is not equal to 0; If the wind speed value is abnormal, the first time point T1, the second time point T2, and the third time point T3 are taken again in chronological order until the wind speed value is no longer abnormal, so that all values ​​within the time interval [T1, T3] are valid.

6. A prediction system based on the deep learning-based progressive short-term wind power prediction method according to any one of claims 1 to 5, characterized in that, include: The system includes: observation data acquisition module, numerical weather prediction request module, power prediction training module, power prediction module, and data transmission module. The observation data acquisition module is used to acquire the observed power sequence and observed wind speed sequence corresponding to the time series with the second time interval within the prediction grid of the wind tower; to acquire the observed power sequence corresponding to the third time series at the location of the wind tower; the observed wind speed sequence of all latitude and longitude grid points within the prediction grid of the wind tower; and the observed wind speed sequence with the interval from the Nth historical day to the current time as the first time interval. The numerical weather prediction request module is used to obtain the predicted wind speed of the latitude and longitude grid points in the wind tower prediction grid; The power prediction training module is used to train the spatial correction MLP network, the temporal difference MLP network, the GRU network, and the fourth MLP network based on the observed power sequence corresponding to the third time series at the location of the wind tower, the observed wind speed sequence of all latitude and longitude grid points within the wind tower prediction grid, and the predicted wind speed. The power prediction module is used to obtain the future predicted power sequence from the current time to the Nth future time, with the interval between the historical Nth day and the current time as the first time interval, and the predicted wind speed sequence between the historical Nth day and the Nth future time, with the interval between adjacent times as the first time threshold, from the latitude and longitude grid points of the meteorological tower prediction grid, in order to predict the power value of the second time threshold for the time interval of the Nth future time. The data transmission module is used to upload the future predicted power sequence from the current time to the Nth day in the future, with the interval between adjacent times being the second time threshold, from the current time in the wind measurement tower location grid to the wind farm dispatch center.

Images