A wind power prediction method based on power heat map and convolutional neural network

Abstract

The application discloses a wind power prediction method based on power heat map and convolutional neural network, comprising the following steps: 1, griding the wind power cluster distribution map; 2, filling the data in the blank cells; 3, converting the power value into the corresponding color pixel point, and then generating the power heat map; 4, using the CNN model to predict the comprehensive power of the wind power cluster, and adjusting the resolution of the power heat map according to the prediction result. The wind power cluster distribution map is gridded, the power value is filled in the corresponding cell by using the interpolation method, the power value is converted into the corresponding color pixel point by using the color parameter to generate the power heat map, and the power heat map is used as the input of the CNN model. Therefore, the application can well use the CNN model to predict the comprehensive power of the wind power cluster.

Description

Technical Field

[0001] This invention relates to the technical field of wind power prediction, specifically to a wind power prediction method based on power heatmaps and convolutional neural networks. Background Technology

[0002] With the rapid development of the world economy, environmental problems are becoming increasingly prominent, and traditional non-renewable resources such as coal and oil are becoming increasingly depleted and their supply is becoming increasingly tight. Wind energy, as a clean and inexpensive renewable energy source, plays a positive role in environmental protection and climate mitigation. In recent years, with strong support from national policies, my country's wind power industry has been developing rapidly. However, due to the intermittent nature of wind energy, wind power output exhibits strong fluctuations. Therefore, large-scale wind power integration into the power grid has a certain impact on the safe and stable operation of the power system. The key to solving this problem is the accurate prediction of wind speed and wind power output.

[0003] Currently, wind power prediction models can be broadly categorized into three types: time series models, traditional machine learning models, and combined models. For linear, stationary data, time series models perform well. To improve prediction results for non-stationary, nonlinear data, machine learning models such as CNN (Convolutional Neural Networks) and LSTM (Long Short-Term Memory) neural networks are used for wind power prediction. CNN models have good feature extraction capabilities and are commonly used in image recognition. However, simply using time series data as input to a CNN model requires further improvement in prediction results. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a wind power prediction method based on power heatmaps and convolutional neural networks, comprising the following steps:

[0005] The distribution map of wind power clusters is gridded, and the power values ​​of each wind farm are filled into the distribution map;

[0006] Fill blank cells with data;

[0007] The power values ​​are converted into pixels of corresponding colors to generate a power heatmap;

[0008] The overall power of the wind power cluster is predicted, and the resolution of the power heatmap is adjusted according to the prediction results until the prediction results meet the expected accuracy.

[0009] Based on the above scheme, the step of gridding the wind power cluster distribution map and filling in the power values ​​of each wind farm in the distribution map specifically involves:

[0010] Based on the latitude and longitude of the wind farms in the area to be identified, the DBSCAN algorithm is used to cluster the wind farms.

[0011] Calculate the geographical distance between each wind farm, compare the geographical distances of wind farms in all categories, and select the minimum distance;

[0012] The value is scaled according to the scale of the wind power cluster distribution map, and the initial sampling interval of the gridded wind power cluster distribution map is obtained by analogy with the idea of ​​image discretization.

[0013] D = min{d1, d2, ... d} i}

[0014] Where d i This represents the minimum geographical distance between wind farms in category i.

[0015] r = D*l

[0016] Where r is the initial sampling interval and l is the scale of the wind power cluster distribution map;

[0017] The wind power cluster distribution map is gridded according to the initial sampling interval;

[0018] If two wind farms appear in a cell after gridding, repeat the above steps to adjust the sampling interval, calculate the geographical distance between the two wind farms, scale the distance value according to the scale, and use the scaled distance value as the new sampling interval.

[0019] The image is then gridded again until each cell contains only one wind farm.

[0020] Based on the above scheme, the number of cells in each row and column of the wind power cluster distribution map shall not be less than 500.

[0021] Based on the above solution, the specific steps for filling blank cells with data are as follows:

[0022] Considering the correlation of wind power between adjacent wind farms, the spatiotemporal correlation weights between wind farms within a certain range adjacent to the cell are calculated based on the cell location of the wind farm.

[0023] By using the CCF cross-correlation coefficient, considering the ultra-short-term wind power prediction period, and specifying the lag time k, the cross-correlation weights between the power sequences of each pair of wind farms are calculated.

[0024] The expression for calculating the cross-correlation weight is as follows:

[0025]

[0026] Where m pi and m pj Let be the mean of the power sequence of the i-th wind farm and the power sequence of the j-th wind farm; si and s j Let p be the variance of the power sequence of the i-th wind farm and the power sequence of the j-th wind farm; it Let p be the historical power sequence at time t of wind farm i. jt+k The power sequence of wind farm j at time t+k is given; the cross-correlation weighting coefficient is based on the autocorrelation weighting coefficient and introduces a time lag parameter k.

[0027] Considering the spatiotemporal correlation of wind power, the interpolation model considering spatiotemporal correlation is constructed by using the historical statistical data of adjacent wind farms and the wind farm itself over multiple days as interpolation samples. The expression of the model is as follows:

[0028]

[0029] Where p(t) is the power value obtained by linear interpolation considering spatiotemporal correlation; p ii (t) represents the historical power data of the wind farm at time t; p ij (t) represents the wind power data of the j-th wind farm adjacent to this wind farm at time t; c is a constant term, a is the coefficient of the interpolation estimate considering its own power, and b is the coefficient of the interpolation estimate considering the power of adjacent wind farms.

[0030] The optimal parameters a, b, and c are obtained by using the least squares method.

[0031] For cells that are still missing data after the above steps, bilinear interpolation is used to fill the cells with data.

[0032] Based on the above scheme, the specific steps of converting power values ​​into pixels of corresponding colors to generate a power heatmap are as follows:

[0033] Each cell in the power data table at time t represents a pixel. The power data within each cell is converted into the color code of the corresponding pixel value, thereby generating the power contour map at time t. The steps are as follows:

[0034] Considering the power range of a typical wind farm, the power value in the power data table is defined as P. max The corresponding color code is (R) H G H B H The power value is P. min The corresponding color code is (R) L G L B L For power data P i The corresponding pixel point O in the power cloud map is calculated. i Color code (R) i G i Bi );

[0035] If the power value is greater than P max Then pixel O i Using P max Corresponding color code (R) H G H B H ), that is, R i =R H G i =G H B i =B H ;

[0036] If the power value is less than P min Then pixel O i Using P min Corresponding color code (R) L G L B L ), that is, R i =R L G i =G L B i =B L ;

[0037] If the power value is in [P] min ,P max Within the range, pixel O i The color code used (R) i G i B i The calculation formula is as follows:

[0038]

[0039]

[0040] Where P i Let be the i-th power value.

[0041] Based on the above scheme, the prediction of the comprehensive power of the wind power cluster specifically involves:

[0042] Using a CNN model, with power heatmap as input and power of all wind farms as output, we can predict the comprehensive power of wind power clusters.

[0043]

[0044] in is the predicted comprehensive power of the wind power cluster; f is the CNN model; [h,w,(R,G,B)] is the three-dimensional input matrix composed of the power cloud image height, image width, and RGB three channels;

[0045] MSE is used as an indicator to evaluate prediction accuracy. The gridded sampling interval is adjusted according to the predicted value until the prediction result meets the expected accuracy.

[0046]

[0047]

[0048] Where P i This represents the measured power output of the wind power cluster. The predicted power output of the wind power cluster is given by n; the number of prediction points is given by r; the sampling interval is given by Δr; and the sampling interval adjustment value is given by MSE. * For the expected accuracy.

[0049] On the other hand, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the computer program to implement the steps of the corresponding method based on the above-described scheme.

[0050] On the other hand, a computer-readable storage medium storing a computer program thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the corresponding method based on the above-described scheme.

[0051] This invention also provides a wind power prediction system based on power heatmaps and convolutional neural networks. The prediction system employs a wind power prediction method based on power heatmaps and convolutional neural networks, and includes:

[0052] The first module involves creating a grid of the wind power cluster distribution map and filling in the power values ​​of each wind farm in the distribution map.

[0053] Module 2: Fill blank cells with data;

[0054] The third module converts power values ​​into pixels of corresponding colors to generate a power heatmap.

[0055] The fourth module predicts the overall power of the wind power cluster and adjusts the resolution of the power heatmap based on the prediction results until the prediction results meet the expected accuracy.

[0056] The beneficial effects of this invention are:

[0057] This method not only uses historical power data of wind farms as input, but also uses a CNN model to predict the comprehensive power of wind power clusters. The resolution of the power heatmap is adjusted according to the prediction results until the prediction results meet the expected accuracy, thereby further improving the prediction accuracy. Attached Figure Description

[0058] The present invention includes the following figures:

[0059] Figure 1 A flowchart illustrating a wind power prediction method based on power cloud maps and convolutional neural networks;

[0060] Figure 2 A schematic diagram of the gridded wind power cluster distribution process. Detailed Implementation

[0061] To make the objectives, advantages, and features of the present invention more apparent, the following description is provided in conjunction with the appendix. Figure 1-2 The present invention will be further described in detail below with reference to specific embodiments.

[0062] One specific implementation method is as follows:

[0063] Step A: Grid the distribution map of wind power clusters. Analogous to the idea of ​​discretizing simulation images, fill in the power values ​​of each wind farm in the grid at the location of the wind farm in the map.

[0064] Based on the latitude and longitude of the wind farms within the desired wind power cluster area, the DBSCAN algorithm is used to cluster the wind farms. The geographical distance between each wind farm is calculated, and the geographical distances of wind farms in all categories are compared. The minimum distance is selected, and this value is scaled according to the scale of the wind power cluster distribution map. Analogous to the idea of ​​image discretization, the initial sampling interval of the gridded wind power cluster distribution map is obtained.

[0065] D = min{d1, d2, ... d} i} (1)

[0066] Where d i This represents the minimum geographical distance between wind farms in category i.

[0067] r=D*l (2)

[0068] Where r is the initial sampling interval and l is the scale of the wind power cluster distribution map.

[0069] The wind farm cluster distribution map is gridded based on the initial sampling interval. If two wind farms appear in a single cell after gridding, the above steps are repeated to adjust the sampling interval. The geographical distance between the two wind farms is calculated and scaled according to the scale. The scaled distance value is used as the new sampling interval, and the image is gridded again until only one wind farm exists in each cell. This process is as follows: Figure 1 As shown. To ensure a smooth transition in the resulting power heatmap, each row and column must contain at least 500 cells. Enter the power output of the wind farm at time t in the cell containing the wind farm's location.

[0070] Step B: Fill the blank cells with data. First, fill the cells near the known wind farms with linear interpolation that takes into account spatiotemporal correlation; fill the remaining blank cells with bilinear interpolation.

[0071] Considering the correlation of wind power between adjacent wind farms, the spatiotemporal cross-correlation weight between wind farms within a certain range adjacent to the cell is calculated based on the cell location of the wind farm; the cross-correlation weight between power sequences of each pair of wind farms is calculated by using the CCF cross-correlation coefficient, considering the ultra-short-term wind power prediction period, and specifying the lag time k.

[0072] The expression for calculating the cross-correlation weight is as follows:

[0073]

[0074] Where m pi and m pj Let be the mean of the power sequence of the i-th wind farm and the power sequence of the j-th wind farm; s i and s j Let p be the variance of the power sequence of the i-th wind farm and the power sequence of the j-th wind farm; it Let p be the historical power sequence at time t of wind farm i. jt+k The power sequence of wind farm j at time t+k is given; the cross-correlation weighting coefficient is based on the autocorrelation weighting coefficient and introduces a time lag parameter k.

[0075] Considering the spatiotemporal correlation of wind power, the interpolation model considering spatiotemporal correlation is constructed by using the historical statistical data of adjacent wind farms and the wind farm itself over multiple days as interpolation samples. The expression of the model is as follows:

[0076]

[0077] Where p(t) is the power value obtained by linear interpolation considering spatiotemporal correlation; p ii (t) represents the historical power data of the wind farm at time t; p ij(t) represents the wind power data of the j-th wind farm adjacent to this wind farm at time t; c is a constant term, a is the coefficient of the interpolation estimate considering its own power, and b is the coefficient of the interpolation estimate considering the power of adjacent wind farms.

[0078] The optimal parameters a, b, and c are obtained by using the least squares method.

[0079] For cells that are still missing data after the above steps, bilinear interpolation is used to fill the cells with data.

[0080] Step C: Based on the color parameters, convert the power values ​​into pixels of the corresponding colors to generate a power heatmap.

[0081] Each cell in the power data table at time t represents a pixel. The power data within each cell is converted into the color code of the corresponding pixel value, thereby generating the power contour map at time t. The steps are as follows:

[0082] Considering the power range of a typical wind farm, the power value in the power data table is defined as P. max The corresponding color code is (R) H G H B H The power value is P. min The corresponding color code is (R) L G L B L For power data P i The corresponding pixel point O in the power cloud map is calculated. i Color code (R) i G i B i );

[0083] If the power value is greater than P max Then pixel O i Using P max Corresponding color code (R) H G H B H ), that is, R i =R H G i =G H B i =B H ;

[0084] If the power value is less than P min Then pixel O i Using P min Corresponding color code (R) L G L B L ), that is, Ri =R L G i =G L B i =B L ;

[0085] If the power value is in [P] min ,P max Within the range, pixel O i The color code used (R) i G i B i The calculation formula is as follows:

[0086]

[0087]

[0088] Where P i Let be the i-th power value.

[0089] Step D: Using the obtained power heatmap as input, a CNN model is used to predict the comprehensive power of the wind power cluster. The resolution of the power heatmap is adjusted according to the prediction results until the prediction results meet the expected accuracy.

[0090] Using a CNN model, with power heatmap as input and power of all wind farms as output, we can predict the comprehensive power of wind power clusters.

[0091]

[0092] in is the predicted comprehensive power of the wind power cluster; f is the CNN model; [h,w,(R,G,B)] is the three-dimensional input matrix composed of the power cloud image height, image width, and RGB three channels;

[0093] MSE is used as an indicator to evaluate prediction accuracy. The gridded sampling interval is adjusted according to the predicted value until the prediction result meets the expected accuracy.

[0094]

[0095]

[0096] Where P i This represents the measured power output of the wind power cluster. The predicted power output of the wind power cluster is given by n; the number of prediction points is given by r; the sampling interval is given by Δr; and the sampling interval adjustment value is given by MSE. * For the expected accuracy.

[0097] The above embodiments are only used to illustrate the present invention and are not intended to limit the present invention. Those skilled in the art can make various changes and modifications without departing from the essence and scope of the present invention. Therefore, all equivalent technical solutions also fall within the scope of the present invention. The scope of patent protection of the present invention should be defined by the claims.

[0098] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

Claims

1. A wind power prediction method based on power heatmaps and convolutional neural networks, characterized in that, Includes the following steps: Step A: The wind power cluster distribution map is gridded to divide it into multiple cells, and the power values ​​of each wind farm are entered into the distribution map. Specifically, this includes: Based on the latitude and longitude of the wind farms in the area to be identified, the DBSCAN algorithm is used to cluster the wind farms. Calculate the geographical distance between each wind farm, compare the geographical distances of wind farms in all categories, and select the minimum distance; The power values ​​are scaled according to the scale of the wind power cluster distribution map, and the initial sampling interval of the gridded wind power cluster distribution map is obtained by analogy with the idea of ​​image discretization. ; in d i For the first i Minimum geographical distance between wind farms within the category; ; in r This is the initial sampling interval. l The scale of the wind power cluster distribution map; The wind power cluster distribution map is gridded according to the initial sampling interval; If two wind farms appear in a cell after gridding, repeat the above steps to adjust the sampling interval, calculate the geographical distance between the two wind farms, scale the distance value according to the scale, and use the scaled distance value as the new sampling interval. The image is then gridded again until each cell contains only one wind farm; Step B: Fill blank cells that do not contain wind farms with data to generate power distribution data that includes the entire wind power cluster area; the specific steps of filling blank cells with data are: considering the correlation of wind power between adjacent wind farms, and calculating the spatiotemporal correlation weight between wind farms within a certain range adjacent to the cell based on the location of the wind farm. By using the CCF cross-correlation coefficient, considering the ultra-short-term wind power forecast period, and specifying the lag time... k Calculate the cross-correlation weights between the power sequences of each pair of wind farms; The expression for calculating the cross-correlation weight is as follows: ; in p it For wind farm i History t Time-major power sequence p jt+k For wind farm j history t+k The power sequence at time step; the cross-correlation weighting coefficient is based on the autocorrelation weighting coefficient by introducing a time lag parameter k; μ pi and μ pj The first i The and the first j The mean of the power series of a wind farm; Considering the spatiotemporal correlation of wind power, the interpolation model considering the spatiotemporal correlation is constructed by using the historical statistical data of adjacent wind farms and itself over multiple days as interpolation samples. For cells that are still missing data after the above steps, bilinear interpolation is used to fill the data in the cells. Step C: Convert the power values ​​into pixels of corresponding colors to generate a power heatmap; Step D: Using the power heatmap as input, predict the overall power of the wind power cluster. Adjust the resolution of the power heatmap based on the prediction results until the prediction results meet the expected accuracy. Specifically, this includes: Using a CNN model, with power heatmap as input and power of all wind farms as output, we can predict the comprehensive power of wind power clusters. ; in This represents the predicted total power output of the wind power cluster. f For CNN models; h , w ,( R , G , B [)] is a three-dimensional input matrix consisting of the power heatmap image height, image width, and RGB three channels; MSE is used as an indicator to evaluate prediction accuracy. The gridded sampling interval is adjusted according to the predicted value until the prediction result meets the expected accuracy. ; ; in This represents the measured power output of the wind power cluster. This refers to the predicted power output of the wind power cluster. n The number of prediction points; r The sampling interval; This is the sampling interval adjustment value; For the expected accuracy.

2. The wind power prediction method based on power heatmap and convolutional neural network according to claim 1, characterized in that, The number of cells in each row and column of the wind power cluster distribution map shall not be less than 500.

3. The wind power prediction method based on power heatmap and convolutional neural network according to claim 1, characterized in that, The generation of the power heatmap specifically includes: Will t Each cell in the time-based power data table represents a pixel. The power data within each cell is converted into a color code corresponding to the pixel value, thereby generating... t The steps to obtain the power heatmap at any given time are as follows: Considering the power range of a typical wind farm, the power values ​​in the power data table are specified as follows: P max The corresponding color code is ( R H , G H , B H The power value is P min The corresponding color code is ( R L , G L , B L For power data P i The corresponding pixel points in the power heat map are calculated. O i Color codes ( R i , G i , B i ); If the power value is greater than P max Then the pixel O i use P max Corresponding color code ( R H , G H , B H ),Right now R i = R H , G i = G H , B i = B H ; If the power value is less than P min Then the pixel O i use P min Corresponding color code ( R L , G L , B L ),Right now R i = R L , G i = G L , B i = B L ; If the power value is in [ P min , P max Within the range, then the pixel point O i The color code used ( R i , G i , B i The calculation formula is as follows: ; ; in P i For the first i Power values.

4. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 3.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 3.

6. A wind power prediction system based on power heatmaps and convolutional neural networks, characterized in that, The wind power prediction method based on power heatmap and convolutional neural network as described in claim 1 includes the following prediction system: The first module involves creating a grid of the wind power cluster distribution map and filling in the power values ​​of each wind farm in the distribution map. Module 2: Fill blank cells with data; The third module converts power values ​​into pixels of corresponding colors to generate a power heatmap. The fourth module predicts the overall power of the wind power cluster and adjusts the resolution of the power heatmap based on the prediction results until the prediction results meet the expected accuracy.

Images