Offshore wind power operation environment scene construction method considering typhoon forecast error

Abstract

The invention relates to an offshore wind power operation environment scene construction method considering typhoon forecast errors. The method comprises the steps of selecting similar historical typhoon forecast data according to current typhoon forecast information; based on wind speeds of similar historical typhoon forecast and current typhoon forecast, combining marine environment element information to construct a marine environment scene information space; a kernel density estimation model of historical typhoon wind speed forecast errors is constructed, corresponding kernel density bandwidths are calculated and updated according to different typhoon forecast timeliness, and historical forecast wind speed error distribution is obtained; and inputting the marine environment scene information space and the historical forecast wind speed error distribution into the improved generative adversarial network, and generating a marine environment scene of which the wind speed interval meets the historical typhoon wind speed forecast error distribution through the wind speed interval constraint loss. Compared with the prior art, the marine environment scene is updated according to the wind speed error distribution under different forecast time periods of the typhoon, and the scene generation quality is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of generating scenarios for the operation environment of offshore wind power, and in particular, to a method for constructing short-term ocean environment scenarios considering typhoon forecast errors. Background Art

[0002] The coastal waters where offshore wind power develops on a large scale and in clusters are often areas severely affected by typhoon and hurricane weather. Against the backdrop of global climate change, extreme climate events such as typhoons show an increasing trend of occurrence, posing severe challenges to the high-reliability operation, high-quality maintenance, and power output of offshore wind farms. Accurately constructing the ocean environment scenario of an offshore wind farm during a typhoon can provide crucial support for the operation and maintenance decisions of offshore wind power, and reduce the uncertain impacts on power grid dispatching, wind farm operation, maintenance decisions, and other aspects during a typhoon.

[0003] In terms of typhoon scenario modeling, from the perspective of scenario application, it can be divided into single-variable and multi-variable scenario modeling. In single-variable scenario modeling, the empirical typhoon wind field models such as Batts, Holland, and Yan Meng are mainly used to describe the wind speed magnitude, relying on historical typhoon data to depict the distribution and evolution characteristics of the wind speed. In multi-variable scenario modeling, the Monte Carlo simulation, Copula method, nested stochastic composite distribution (NSCD), and linear superposition offset model are mainly adopted, which can reflect the coupling relationship between multi-dimensional ocean environment variable characteristics such as wave height and wave peak period. The above models and methods provide scenario support for the power output and operation and maintenance decisions of wind farms. However, the scenarios based on historical data cannot accurately reflect the actual process of a single typhoon affecting a wind farm. From the forecast information released by meteorological departments, due to the randomness of typhoon paths and intensities, it is difficult to accurately forecast them, and the deviation of the time-limited forecast is relatively large. How to utilize typhoon forecast information to reduce the uncertain impact of the forecast time-limited deviation on scenario generation and improve the quality of scenario generation is of great significance for the large-scale development and high-reliability operation of offshore wind power.

[0004] Chinese Patent Application CN202311863996.4 discloses a typhoon disaster simulation method, device, terminal, and medium. The technical solution provided by this method is based on historical typhoon data. First, static wind field models of multiple different time sections are established for multiple key parameters of the typhoon, such as the typhoon landing point, moving direction, and real-time speed, and then combined with the land attenuation coefficient of the typhoon to construct an evolution model considering the spatio-temporal dynamic characteristics of the typhoon. Although this model can better simulate the spatial distribution characteristics of typhoon wind speed, the generated scenario has not combined typhoon forecast data and is difficult to reflect the real process of future typhoons.

[0005] Chinese Patent Application CN202411259930.9 discloses a short-term prediction method for the marine environment. The method includes constructing a time-series variational autoencoder model. The encoder uses a long short-term memory network to obtain a compressed representation of the marine environment state, and the decoder uses a long short-term memory network to reconstruct the original marine environment state sequence from the compressed representation of the marine environment state. This method has not considered the typhoon process and cannot describe the marine environment scenario under the influence of typhoons. Summary of the Invention

[0006] The object of the present invention is to overcome the defects of the above-mentioned prior art, such as being difficult to reflect the real process of future typhoons and unable to describe the marine environment scenario under the influence of typhoons, and to provide a method for constructing a marine wind power operation environment scenario considering typhoon prediction errors.

[0007] The object of the present invention can be achieved by the following technical solutions:

[0008] A method for constructing a marine wind power operation environment scenario considering typhoon prediction errors, the steps include:

[0009] Select similar historical typhoon prediction data according to the current typhoon prediction information;

[0010] Based on the wind speeds of the similar historical typhoon predictions and the current typhoon prediction, combined with the marine environment element information, construct a marine environment scenario information space;

[0011] Construct a kernel density estimation model for the historical typhoon wind speed prediction error. The estimation model calculates and updates the corresponding kernel density bandwidth according to different typhoon prediction time periods, and obtains the historical prediction wind speed error distribution;

[0012] Input the marine environment scenario information space and the historical prediction wind speed error distribution into an improved generative adversarial network. The improved generative adversarial network generates a marine environment scenario whose wind speed interval satisfies the historical typhoon wind speed prediction error distribution through a wind speed interval constraint loss.

[0013] As a preferred technical solution, the similar historical typhoon prediction data is selected according to the similarity of typhoon intensity and path information. The total similarity S total is composed of the similarity S pres of the central minimum pressure of the typhoon and the path similarity S path obtained by weighted average:

[0014] S total = w pres ·S pres + w path ·S path

[0015] where: w pres and w pathThey are the weight coefficients of the central lowest pressure similarity and the path similarity respectively;

[0016] The similarity of the typhoon forecast intensity in this instance to the central lowest pressure of the i-th historical forecast is as follows:

[0017]

[0018] In the formula: P cur , P i his are the predicted values of the central lowest pressure of the typhoon in this instance and the i-th historical typhoon respectively;

[0019] The path similarity is the minimum total distance of all possible path alignment methods calculated by the dynamic time warping algorithm between the path of the typhoon forecast in this instance and the path of the historical forecast typhoon. The path similarity between the set of typhoon path sequence points at the current moment and the set of typhoon path sequence points of the q-th historical typhoon is as follows:

[0020]

[0021]

[0022]

[0023]

[0024] In the formula: is the DTW distance between the set of typhoon path sequence points T cur at the current moment and the set of typhoon path sequence points of the q-th historical typhoon; c(m,n) is the element in the lower right corner of the cumulative distance matrix C(m,n). The cumulative distance matrix C(m,n) represents the optimal alignment path between each pair of coordinate points of T cur and , indicating the minimum alignment cost from the first point to the i-th point of T cur and from the first point to the j-th point of the path ; d(i,j) is the Euclidean distance between any point in the set of typhoon sequence points in this instance and any point in the set of historical typhoon sequence points.

[0025] As an optimal technical solution, the path similarity weight coefficient w path is obtained by the angle change between adjacent points on the path, i.e., the path curvature:

[0026]

[0027]

[0028] In the formula: D vit is the path curvature difference between the current predicted typhoon and the historical predicted typhoon, and respectively represent the path curvatures of the current predicted and historical typhoon prediction sequence points; d vit is the threshold of the curvature difference.

[0029] As an optimal technical solution, the construction of the marine environmental scenario information space is specifically as follows:

[0030] Arrange the similarity S total values of the historical predicted typhoon and the current predicted typhoon from large to small, and select the historical predicted typhoon wind speed dataset v thr higher than the similarity threshold S for and the corresponding actual typhoon wind speed dataset v act to form a typhoon wind speed matrix

[0031] Combine the wave height H sp and the wave peak period T sp during the typhoon and related to the scene accessibility, to obtain the marine environmental scenario information space R sp =[V sp ,H sp ,T sp .

[0032] As an optimal technical solution, for the kernel density estimation model of the historical typhoon wind speed prediction error, the wind speed error probability density estimated by KDE at the t prediction time limit is as shown in the following formula:

[0033]

[0034]

[0035] In the formula: h g is the bandwidth of the Gaussian kernel function, which is updated in two stages: outside the 24-hour warning line and within the 24-hour warning line; N is the total number of wind speed samples; is the nth wind speed error in the dataset at different prediction time limits t, is the root mean square error of the wind speed error.

[0036] As an optimal technical solution, for the typhoon error data outside the 24-hour warning line, the optimal theoretical window width theory is used to minimize the integrated mean square error MISE(·) between the wind speed error estimation density and the true density, and the optimal bandwidth is calculated as shown below:

[0037]

[0038] Where: N out is the number of wind speed error samples outside the 24-hour warning line, and σ N represents the variance of the wind speed error sample data set; f(ε t ) is the probability density function of the wind speed forecast error; is the kernel density estimation result of the probability density function f(ε t ).

[0039] As a preferred technical solution, for the typhoon error data within the 24-hour warning line, the local density weight d of the data point is introduced n to quantify the local density characteristics around the nth data point, thereby updating the bandwidth value, as shown in the following formula:

[0040]

[0041] Where: N in is the number of error samples within the 24-hour warning line, represents the nth wind speed error data of the mth typhoon at this forecast time, is the mean value of the error data sample; ζ g is the local density scale parameter, which is used to control the attenuation speed of the local density weight.

[0042] As a preferred technical solution, the improved generative adversarial network includes a generator and a discriminator. Specifically:

[0043] The historical typhoon wind speed error distribution, wave height, and wave peak period data are used as the conditional vector c, which is connected to the low-dimensional latent variable noise z to form the joint input G(z, c) of the generator G;

[0044] The generator G projects the joint input into a high-dimensional space through a fully connected layer for shape transformation; a multi-scale convolutional network is used to capture the local details and global patterns of the ocean environment scene variables through convolutional kernels of different sizes; multi-layer transposed convolutions are used for upsampling step by step to generate a high-resolution ocean scene including the uncertain wind speed interval, wave height, and wave peak period;

[0045] The discriminator D(s) receives the original sample data and the generated ocean environment scene data respectively, and discriminates the authenticity of the generated data.

[0046] As a preferred technical solution, the improved generative adversarial network sets the wind speed interval constraint loss L W to make the wind speed interval in the generated ocean environment scene satisfy the estimated probability density distribution f(ε t ) of the historical typhoon forecast wind speed error. The loss functions L G and L D are as shown in the following formula:

[0047]

[0048] where: E(·) represents the mathematical expectation; p s (s) is the distribution of the generated ocean environmental scenario data of wind speed, wave height, and wave peak period; f w [G(s)] is the evaluation output of the discriminator for the scenario G(s) generated by the generator; λ w is the weight coefficient for balancing the adversarial loss of the generator and the wind speed interval constraint loss; p data (x) is the distribution of the historical ocean environmental scenario data of wind speed, wave height, and wave peak period, f w (x) is the evaluation output of the discriminator for the historical data samples;

[0049] The generator and the discriminator jointly optimize the objective function through an adversarial game, as shown in the following formula:

[0050]

[0051] As a preferred technical solution, the improved generative adversarial network introduces the Wasserstein distance to replace the JS divergence of the generative adversarial network, which is defined as follows:

[0052]

[0053] where: f(x) and f(s) are the probability distributions of historical data and generated data respectively; sup is the supremum; ||f|| L ≤1 indicates that the function f should satisfy 1-Lipschitz continuity.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] 1) Aiming at the problem of quantifying the uncertainty of the ocean environmental scenario under the influence of typhoons, the present invention combines the historical sample information of typhoons, the current forecast information, and the information of the ocean environmental elements affecting the wind farm under normal conditions, and proposes a short-term ocean environmental scenario generation method based on an improved generative adversarial network. The improved generative adversarial network also introduces a wind speed interval constraint loss, so that the wind speed interval in the generated short-term ocean environmental scenario satisfies the estimated probability density distribution of the wind speed error, and can update the ocean environmental scenario according to the wind speed error distribution at different typhoon forecast times, which can reflect the real process of future typhoons and improve the quality of scenario generation.

[0056] 2) The improved generative adversarial network proposed by the present invention takes the wind speed error distribution, significant wave height, and wave peak period data of similar historical typhoon forecasts as conditional vectors for input. First, the joint input is projected into a high-dimensional space for shape transformation, and then a multi-scale convolutional network kernel is used to capture the local details and global patterns of ocean environmental scene variables. The generated high-resolution ocean scenes containing uncertain wind speed intervals, significant wave heights, and wave peak periods can reflect the temporal coupling relationship between multi-dimensional ocean environmental variable characteristics considering typhoon forecast errors, and describe the ocean environmental scenes under the influence of typhoons, improving the quality of short-term ocean environmental scene generation considering typhoons.

[0057] 3) The present invention uses the kernel density estimation method to estimate the distribution characteristics of typhoon forecast wind speed errors. And for the multi-modal distribution characteristics of typhoon wind speed forecast errors, the probability density bandwidth of typhoon forecast wind speed errors is updated in two stages: outside the 24-hour warning line and within the 24-hour warning line. It is applicable to the complex and irregular data distribution of typhoon wind speeds. And for the differences in the distribution characteristics of wind speed errors caused by different forecast lead times, KDE can capture the possible multi-peak distribution characteristics in the data and find a balance between smoothness and retaining data characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 is a flowchart of the method for constructing the offshore wind power operation environment scene considering typhoon forecast errors according to the present invention;

[0059] Figure 2 is a flowchart of the scene generation training based on TFE-C-DCGAN in the present invention;

[0060] Figure 3 is a graph showing the change of the loss functions of the generator and discriminator in a specific embodiment of the present invention;

[0061] Figure 4 is a schematic diagram of the generated result and the weather accessible window in a specific embodiment of the present invention;

[0062] Figure 5 is a schematic diagram of the scene wind speed results generated by different prediction methods in a specific embodiment of the present invention;

[0063] Figure 6 is a schematic diagram of the KDE estimation result of the wind speed error at a 24-hour forecast lead time in a specific embodiment of the present invention;

[0064] Figure 7 is a schematic diagram of the KDE estimation result of the wind speed error outside the 24-hour forecast lead time in a specific embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0065] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented on the premise of the technical solution of the present invention, and provides a detailed implementation manner and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0066] Embodiment 1

[0067] Aiming at the problem of quantifying the uncertainty of the marine environment scenario under the influence of typhoons, the present invention combines the historical sample information of typhoons (including basic information and historical forecast information), the current forecast information, and the marine environment element information (such as wind speed, wave height, wave peak period, etc.) affecting the wind farm under normal conditions, and proposes a short-term marine environment scenario generation method based on an improved generative adversarial network to improve the quality of scenario generation. The historical typhoon forecast data with high similarity is selected through a typhoon similarity model. Combining other element information of the marine environment such as wave height and wave peak period, a scenario information space is constructed. And a kernel density estimation model of the historical typhoon wind speed forecast error is constructed. The marine environment scenario considering the typhoon forecast error is generated through an improved generative adversarial network, and the kernel density bandwidth window is updated according to different typhoon forecast time limits to obtain the updated scenario. The specific process is as Figure 1 shown.

[0068] Step 1: Construct a scenario information space considering typhoon intensity and path similarity

[0069] According to the current typhoon forecast information, select the historical typhoon forecast information similar to it, and then combine the marine environment element information (such as wind speed, wave height, wave peak period, etc.) under normal conditions to form a marine environment scenario information space.

[0070] The present invention constructs a similarity index based on typhoon intensity and path information. Typhoon intensity includes the central maximum wind speed and the lowest air pressure. Since the change trends of the two are basically the same, the present invention uses the central lowest air pressure. The similarity between the current typhoon forecast intensity and the i-th historical forecast is shown in formula (1):

[0071]

[0072] In the formula: P cur , P i his are the predicted values of the central lowest air pressure of the current typhoon and the i-th historical typhoon respectively.

[0073] The moving path of the typhoon is time series data, which is represented by the longitude and latitude coordinates of the typhoon at different time points. Assume that the set of predicted path coordinate points of the current typhoon is The predicted path coordinate sequence point of the q-th historical typhoon is Since the set of path coordinate sequence points of the current typhoon forecast has the characteristic of non-linear alignment with the set of path coordinate sequence points of historical typhoon forecasts, the dynamic variation law of the sequence points on the time axis is considered, and the dynamic time warping algorithm (DTW) is used to calculate the minimum distance between the two forecast paths. Here, the minimum distance refers to the minimum total distance considering all possible path alignment methods. Then, the set of typhoon path sequence points T at the current moment cur and the set of path sequence points of the q-th historical typhoon The DTW distance and the path similarity are as follows:

[0074]

[0075]

[0076]

[0077]

[0078] In the formula: c(m,n) is the lower right corner element of the cumulative distance matrix C(m,n). The cumulative distance matrix C(m,n) represents the optimal alignment path between each pair of coordinate points of T cur and , indicating the minimum alignment cost from the first point to the i-th point of T cur and the first point to the j-th point of the path ; d(i,j) is the Euclidean distance between any point in the set of current typhoon sequence points and any point in the set of historical typhoon sequence points.

[0079] The total similarity S is formed by weighted averaging the similarity of the lowest central pressure of the typhoon center and the path total , as shown in the following formula:

[0080] S total = w pres ·S pres + w path ·S path (6)

[0081] In the formula: w pres and w path are the weight coefficients of the central pressure similarity and the path similarity respectively, and w pres + w path = 1. The selection of the path similarity weight coefficient w path is obtained from the angular change between adjacent points on the path, that is, the path curvature, as shown in formula (7):

[0082]

[0083]

[0084]

[0085] In the formula: D vit is the path curvature difference between the typhoon forecasted this time and the historical typhoon forecasts, and respectively represent the path curvatures of the current forecast and the historical forecasts; k i is the path curvature of the typhoon forecast sequence points; d vit is the threshold of the curvature difference, and its value is related to the typhoon intensity. Strong typhoons (low pressure, high wind speed) are usually accompanied by large path curvature changes, while the path changes of weak typhoons are relatively gentle.

[0086] Arrange the similarity S total values of the historical typhoon forecasts and the current typhoon forecast from largest to smallest, and select the historical typhoon wind speed dataset v thr higher than the similarity threshold S for and the corresponding actual typhoon wind speed dataset v act to form a typhoon wind speed matrix Then, combined with the wave height H sp and wave peak period T sp related to the scene accessibility during the typhoon, the ocean environmental scene information space R sp =[V sp , H sp , T sp .

[0087] Step 2: Initial scene generation based on the typhoon forecast error TFE-C-DCGAN

[0088] In order to reflect the temporal coupling relationship between the multi-dimensional ocean environmental variable characteristics considering the typhoon forecast error and improve the quality of short-term ocean environmental scene generation considering typhoons, the present invention proposes a method for short-term ocean environmental scene generation based on the typhoon forecast error deep convolutional generative adversarial network TFE-C-DCGAN (Typhoon forecast error-condition-deep convolution generative adversarial networks, TFE-C-DCGAN), and the model structure and training process are as Figure 2 shown.

[0089] First, the historical typhoon wind speed error distribution, wave height, and wave peak period data are used as the conditional vector c, which is concatenated with the low-dimensional latent variable noise z to form the combined input G(z, c) of the generator G. To enable the generator to capture the non-linear relationships among the wind speed error, wave height, and wave peak period variables in a higher-dimensional feature space, the combined input is projected into a high-dimensional space through a fully connected layer for shape transformation. A multi-scale convolutional network is used instead of the traditional convolutional network, and convolutional kernels of different sizes are set to capture the local details and global patterns of the ocean environment scene variables. Then, multi-layer transposed convolutions are used for upsampling step by step to generate a high-resolution ocean scene s that includes the uncertain wind speed interval, wave height, and wave peak period. Finally, the discriminator D(s) receives the original sample data and the generated ocean environment scene data respectively and discriminates the authenticity of the generated data, as shown in the following formula:

[0090]

[0091] In the formula: Conv trans and Conv represent transposed convolution and convolution operations; represents the concatenation operation, is to process the concatenated vector through a fully connected layer; H pre and T pre represent the historical wind speed error, wave height, and wave peak period data of the ocean environment; J represents the output of the multi-scale convolutional layer, which is used to extract the multi-scale features of the combined input vector G(z, c); Concat(·) represents the concatenation of the features of the convolutional outputs at different scales, W k represents the k-th size convolutional kernel, where b k is the trainable parameter corresponding to the convolutional operation, which is used to adjust the value of the convolutional layer output, σ is the ReLU (Rectified Linear Unit) activation function, defined as σ(x) = max(0, x); where H gen and T gen represent the generated wind speed interval, wave height, and wave peak period respectively; the discriminator D(s) ranges from 0 to 1, which is the probability of judging that the input data is real data.

[0092] To make the wind speed interval in the generated short-term ocean environment scene satisfy the estimated probability density distribution f(ε t ) of the wind speed error, the wind speed interval constraint loss L W is introduced. Then, the loss functions L G and L D of the generator and the discriminator are as shown in Equation (11):

[0093]

[0094] where: E(·) represents the mathematical expectation; p s (s) is the distribution of marine environmental data such as the generated wind speed, wave height, and wave crest period, indicating that s follows p s (s) when taking the expectation; f w [G(s)] is the evaluation output of the discriminator for the generated scenario G(s); λ w is the weight coefficient for balancing the adversarial loss of the generator and the wind speed interval constraint loss, and the optimal value is determined using the grid search method and cross-validation; p data (x) is the distribution of historical marine environmental data such as wind speed, wave height, and wave crest period, indicating that the expectation x follows p data (x) when taking the expectation; f w (x) is the evaluation output of the discriminator for the historical data samples.

[0095] The generator and the discriminator jointly optimize the objective function through an adversarial game, as shown in the following formula:

[0096]

[0097] When the output distribution of the generator is significantly different from the true data distribution, the output distribution of the discriminator may become very small and there may be a problem of vanishing gradients. Therefore, the Wasserstein distance is introduced to replace the JS divergence of the traditional GAN, and the definition is as follows:

[0098]

[0099] where: f(x) and f(s) are the probability distributions of historical data and generated data respectively; sup is the supremum; ||f|| L ≤1 indicates that the function f should satisfy 1-Lipschitz continuity.

[0100] Step 3. Update the marine environmental scenario reflecting the typhoon forecast time error characteristics.

[0101] The error distribution between the true typhoon data and the forecast data may show skewness and multimodal distribution. If directly used as the input of the monthly marine environmental scenario generator, it will increase the loss and the number of iterations of the generator, and it is difficult to converge. Therefore, the kernel density estimation method (Kernel Density Estimation, KDE) is used to estimate the distribution characteristics of the typhoon forecast wind speed error. Compared with the traditional parametric model of typhoon wind speed, the KDE method does not require the error data to follow a certain known distribution form, is suitable for the complex and irregular data distribution of typhoon wind speed, and for the differences in the distribution characteristics of wind speed errors caused by different forecast time periods, KDE can capture the possible multimodal distribution characteristics in the data and find a balance between smoothness and retaining data characteristics.

[0102] When the typhoon is outside the 24-hour warning line, the meteorological department monitors the position and intensity of the typhoon (the maximum average wind speed in the nearest 2 minutes and the lowest sea level pressure at the center) every 12 hours, and the forecast error is relatively large; after entering the 24-hour warning line, the monitoring interval is shortened to every 6 hours, and the forecast error is smaller. In order to reflect the error distribution characteristics under different forecast lead times, the wind speed error ε of the m-th typhoon at the forecast lead time t t,m is shown in the following formula:

[0103]

[0104] In the formula: and are the true wind speed and the forecast wind speed of the m-th typhoon at different forecast lead times t (outside or inside the 24-hour warning line), respectively.

[0105] According to formula (14), the sample data set of wind speed forecast error Then the probability density of the wind speed error estimated by KDE at the forecast lead time t is shown in formula (15):

[0106]

[0107] In the formula: h g is the bandwidth of the Gaussian kernel function, which determines the smoothness of the estimation result; N is the total number of wind speed samples; is the n-th wind speed error in the data set, is the root mean square error of the wind speed error.

[0108] Since the distribution of typhoon wind speed forecast error is asymmetric and affected by different forecast lead times, the distribution shows multimodal characteristics. Here, the Gaussian kernel function K(·) is selected to naturally superimpose the sample data points to form a multi-peak characteristic, and the symmetry assumption of the data distribution does not need to be forced, as shown in the following formula:

[0109]

[0110] In kernel density estimation, the selection of the bandwidth has an important influence on the estimation result of the probability density function. If the bandwidth is too small, the estimation result will be too fluctuating, capturing noise rather than the true distribution; if the bandwidth is too large, the estimation result will be too smooth, losing the detailed characteristics of the distribution. Therefore, the reasonable selection of the bandwidth is the key to kernel density estimation. In view of the multimodal distribution characteristics of typhoon wind speed forecast error, the bandwidth of the probability density of typhoon forecast wind speed error is updated in two stages: outside the 24-hour warning line and inside the 24-hour warning line.

[0111] (1) Outside the 24-hour warning line

[0112] For the typhoon error data outside the 24-hour warning line, with a time interval of 12h for the data, the traditional optimal theory window width theory is used to minimize the integrated mean square error MISE(·) between the estimated density of wind speed error and the true density, and the optimal bandwidth is calculated. Specifically as follows:

[0113]

[0114] In the formula: N out is the number of wind speed error samples outside the 24-hour warning line, and σ N represents the variance of the wind speed error sample data set; is the density function of the wind speed prediction error; f(ε t ) is the kernel density estimation result of the density function f(ε t ).

[0115] (2) Inside the 24-hour warning line

[0116] When the typhoon enters within the 24-hour warning line, since the time interval of the typhoon wind speed prediction error data is shortened to every 6 hours and the data volume increases significantly, therefore, the bandwidth selects to be more dependent on local data density information. By introducing the local density weight d n of the data point to quantify the local density characteristics around the nth data point, and thus update the bandwidth value. The specific calculation method is shown in formulas (20)-(21):

[0117]

[0118] In the formula: N in is the number of error samples within the 24-hour warning line, represents the nth wind speed error data of the mth typhoon at this forecast time, is the mean value of the error data sample; ζ g is the local density scale parameter, which controls the attenuation speed of the local density weight and can be optimized by cross-validation.

[0119] The present invention uses a typhoon similarity model to select historical typhoon forecast data to analyze the typhoon wind speed error distribution, combines the short-term ocean environment scenario under the influence of the current forecast typhoon, and updates the ocean environment scenario according to the wind speed error distribution at different typhoon forecast times, which helps to improve the quality of scenario generation.

[0120] Example 2

[0121] As one of the specific implementation examples of the present invention, the scenario constructed in this embodiment is a marine environment scenario including typhoon processes. A total of 318 sets of historical typhoon forecast data and real-time process data in a certain sea area from 2008 to 2020 are selected, and the atmospheric reanalysis dataset EAR5 (released by the European Centre for Medium-Range Weather Forecasts) in this area is collected, including significant wave height and wave peak period. The resolution of this dataset is 1h, and the single-month sample is a 30×24×3 matrix.

[0122] The scenario generation of the present invention is selected to send computational operations to the GPU for execution under the tensorflow deep learning framework, and the GPU is accelerated by CUDA. The network optimization algorithm uses the Adam optimizer, with an initial learning rate of 0.0002. The maximum number of iterations for training is set to 630 times, and the batch size is set to 64 according to the GPU memory. The dimension of the noise vector is 128, and noise elements are independently sampled from the standard normal distribution.

[0123] As Figure 3 shown is the curve of the generated loss varying with the number of training times. It can be seen that both the generator loss and the discriminator loss tend to be stable after about 200 iterations, and the marine environment scenario can be stably generated. The marine environment generated by the present invention includes the wind speed range, significant wave height, and wave peak period, as specifically shown in Figure 4 shown.

[0124] To verify the effectiveness of the proposed method for generating marine environment scenarios considering typhoon forecast uncertainty of the present invention, comparative analyses are respectively carried out on the following three scenario construction methods:

[0125] 1) Method 1: Without considering typhoon forecast errors, a marine scenario is constructed only based on the C-DCGAN model.

[0126] 2) Method 2: Considering typhoon forecast errors, but without updating the kernel density estimation results of the error distribution, a marine scenario is generated based on the TFE-C-DCGAN model.

[0127] 3) Method 3: Considering typhoon forecast errors and updating the kernel density estimation results of the error distribution, a marine scenario considering typhoon forecast uncertainty is generated based on the TFE-C-DCGAN model.

[0128] The wind speed results generated by the three scenario construction methods are as shown in Figure 4 shown. Since the historical wind speed data is input into the generator in Method 1, while the input data of Method 2 and Method 3 includes the historical typhoon wind speed error distribution, the wind speed generated by Method 1 is a wind speed curve, while the wind speed intervals are generated by Method 2 and Method 3, where the red curve is the actual wind speed.

[0129] Analysis Figure 5It can be seen that the wind speed results generated by Method 1 are relatively close to the true values in the initial stage. However, with the arrival of the typhoon, the prediction accuracy of Method 1 gradually decreases. In the middle and late stages of the typhoon, there is a large difference between the generated wind speed and the actual value, and the maximum error reaches 8.5 m / s. This indicates that Method 1 cannot effectively capture the change characteristics of the typhoon intensity, resulting in an excessive error in the generated results.

[0130] Since the wind speed error distribution of Method 2 is the same as that of Method 3 before 120 hours, the generated wind speed interval is the same as that of Method 3 and can accurately cover the actual wind speed curve. However, after 120 hours, due to the lack of update of the wind speed error distribution, the wind speed interval generated by Method 2 gradually loses accuracy and cannot effectively include the changes in the actual wind speed. This result shows the importance of wind speed error update in the generation of typhoon-containing scenarios.

[0131] As Figure 6 and Figure 7 shown, there is a difference of 5.5 m / s in the wind speed error distribution within and outside 24 hours at the maximum probability density, and the wind speed error value outside 24 hours increases significantly, reaching 14.3 m / s. The results show that the update of the wind speed error distribution plays an important role in improving the prediction accuracy. Compared with Method 2, Method 3 can better adapt to the changes in the actual wind speed by dynamically updating the kernel density estimation of the wind speed error. As Figure 5 shown, it can provide more stable and accurate wind speed prediction results, especially when the typhoon intensity fluctuates greatly, making the generated results more robust.

[0132] The preferred specific embodiments of the present invention have been described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and variations based on the concept of the present invention without creative work. Therefore, all technical solutions that can be obtained by those skilled in the art in the technical field of the present invention based on the concept of the present invention through logical analysis, reasoning, or limited experiments on the basis of the prior art should be within the protection scope determined by the claims.

Claims

1. A method for constructing an operating environment scenario of an offshore wind farm considering typhoon prediction errors, characterized in that the steps Including: Select similar historical typhoon forecast data according to the current typhoon forecast information; Based on the wind speeds of the similar historical typhoon forecasts and the current typhoon forecast, combined with the marine environmental element information, construct a marine environmental scenario information space; Construct a kernel density estimation model for the historical typhoon wind speed forecast error. The estimation model calculates and updates the corresponding kernel density bandwidth according to different typhoon forecast time periods, and obtains the historical forecast wind speed error distribution; Input the marine environmental scenario information space and the historical forecast wind speed error distribution into an improved generative adversarial network. The improved generative adversarial network generates a marine environmental scenario whose wind speed interval satisfies the historical typhoon wind speed forecast error distribution through a wind speed interval constraint loss.

2. A method for constructing an offshore wind power operation environment scenario considering typhoon forecast errors according to claim 1, characterized in that The similar historical typhoon forecast data is selected according to the similarity of typhoon intensity and path information, and the total similarity S total is composed of the similarity S pres of the central lowest pressure of the typhoon and the path similarity S path obtained by weighted average: S total = w pres · S pres + w path · S path where: w pres and w path are the weight coefficients of the similarity of the central lowest pressure and the similarity of the track, respectively; The similarity between the predicted intensity of the current typhoon and the central lowest air pressure of the i-th historical prediction is as follows: Where: P cur , P i his are the predicted minimum central pressures of the current typhoon and the i-th historical typhoon respectively; The path similarity is the minimum total distance of all possible path alignment methods calculated by the dynamic time warping algorithm between the typhoon path of the current forecast and the historical typhoon paths. The path similarity between the set of typhoon path sequence points at the current moment and the set of typhoon path sequence points of the q-th historical typhoon is as follows: As follows: In the formula: is the set of typhoon path sequence points T at the current moment cur and the set of typhoon path sequence points of the q-th historical typhoon The DTW distance; c(m,n) is the element in the lower right corner of the cumulative distance matrix C(m,n), and the cumulative distance matrix C(m,n) is the optimal alignment path between each pair of coordinate points of T cur and indicating the minimum alignment cost from the first point to the i-th point of T cur and from the first point to the j-th point of the path d(i,j) is the Euclidean distance between any point in the current typhoon sequence point set and any point in the historical typhoon sequence point set.

3. The method for constructing an offshore wind power operation environment scenario considering typhoon forecast errors according to claim 2, wherein The path similarity weight coefficient w path is obtained from the angle change between adjacent points on the path, i.e., the path curvature: Where: D vit is the path curvature difference between the typhoon forecast for this time and the historical typhoon forecasts, and respectively represent the path curvatures of the sequence points of the typhoon forecast for this time and the historical typhoon forecasts; d vit is the threshold of the curvature difference.

4. The method for constructing an offshore wind power operation environment scenario considering typhoon forecast errors according to claim 1, wherein The construction of the marine environmental scenario information space is specifically as follows: Arrange the similarity S between the historical forecast typhoon and the current forecast typhoon in descending order, and select the historical forecast typhoon wind speed dataset v total higher than the similarity threshold S thr and the corresponding actual typhoon wind speed dataset v for to form a typhoon wind speed matrix act ​ Combined with the wave height H related to the scene accessibility during the typhoon sp and peak period T sp , and obtain the ocean environment scene information space R sp =[V sp ,H sp ,T sp [.

5. A method for constructing an offshore wind power operation environment scenario considering typhoon forecast errors according to claim 1, characterized in that The kernel density estimation model of the historical typhoon wind speed forecast error, the wind speed error probability density estimated by KDE at the t forecast time limit is shown as follows: where: h g is the bandwidth of the Gaussian kernel function, which is updated in two stages: outside the 24-hour warning line and within the 24-hour warning line; N is the total number of wind speed samples; is the nth wind speed error in the dataset at different forecast lead times t, is the root mean square error of the wind speed error.

6. A method for constructing an offshore wind power operation environment scenario considering typhoon forecast errors according to claim 5, characterized in that, For the typhoon error data outside the 24-hour warning line, the optimal theoretical window width theory is used to minimize the mean integrated squared error MISE(·) between the estimated density of wind speed error and the true density, and the optimal bandwidth is calculated. The specific content is as follows: Where: N out is the number of wind speed error samples outside the 24-hour warning line, and σ N represents the variance of the wind speed error sample data set; f(ε t ) is the probability density function of the wind speed forecast error; is the kernel density estimation result of the probability density function f(ε t ).

7. A method for constructing an offshore wind power operation environment scenario considering typhoon forecast errors according to claim 5, characterized in that, For typhoon error data within the 24-hour warning line, the local density weight d of the data point is introduced n To quantify the local density features around the nth data point, the bandwidth value is updated as shown in the following formula: Where: N in is the number of error samples within the 24-hour warning line, represents the nth wind speed error data at this forecast time for the mth typhoon, is the mean value of the error data samples; ζ g is the local density scale parameter, which is used to control the attenuation rate of the local density weight.

8. A method for constructing an offshore wind power operation environment scenario considering typhoon forecast errors according to claim 1, characterized in that The improved generative adversarial network includes a generator and a discriminator. Specifically: Connect the historical typhoon wind speed error distribution, significant wave height and wave period data as the conditional vector c to the low-dimensional latent variable noise z to form the joint input G(z, c) of the generator G; The generator G projects the joint input into a high-dimensional space through a fully connected layer for shape transformation; adopts a multi-scale convolutional network to capture the local details and global patterns of marine environmental scenario variables through convolutional kernels of different sizes; adopts multi-layer transposed convolution to gradually upsample and generate a high-resolution marine scenario containing an uncertain wind speed interval, significant wave height and wave period; The discriminator D(s) receives the original sample data and the generated marine environmental scenario data respectively, and discriminates the authenticity of the generated data.

9. A method for constructing an offshore wind power operation environment scenario considering typhoon forecast errors according to claim 8, characterized in that The improved generative adversarial network sets a wind speed interval constraint loss L W , so that the wind speed interval in the generated ocean environmental scene satisfies the estimated probability density distribution f(ε t ) of the historical typhoon forecast wind speed error. The loss functions L G and L D are shown as follows: Where: E(·) represents the mathematical expectation; p s (s) is the generated distribution of ocean environmental scenario data of wind speed, wave height, and wave peak period; f w [G(s)] is the evaluation output of the discriminator on the scenario G(s) generated by the generator; λ w is the weight coefficient for balancing the adversarial loss of the generator and the wind speed interval constraint loss; p data (x) is the historical distribution of ocean environmental scenario data of wind speed, wave height, and wave peak period, f w (x) is the evaluation output of the discriminator on the historical data samples; The generator and the discriminator jointly optimize the objective function through an adversarial game, as shown in the following formula:

10. A method for constructing an offshore wind power operation environment scenario considering typhoon forecast errors according to claim 8, characterized in that, The improved generative adversarial network introduces the Wasserstein distance to replace the JS divergence of the generative adversarial network, which is defined as follows: where: f(x) and f(s) are the probability distributions of historical data and generated data respectively; sup is the supremum; ||f|| L ≤ 1 indicates that the function f should satisfy 1-Lipschitz continuity.

Images