Daily wind power prediction method and system based on temporal and spatial dense wind field derivation

Abstract

The present invention discloses a method and system for predicting intraday wind power based on the deduction of spatiotemporal dense wind fields. The method comprises: integrating physical prior knowledge, modeling the spatiotemporal dynamic evolution process of dense wind fields as multi-order partial differential constraints, solving partial differential factors of each order by moment estimation method; using a spatiotemporal wind field prediction-correction method based on Kalman gain, using a time series generation structure to reduce error accumulation, and citing the wind field partial differential moment estimation loss as a regularization term, training to obtain multi-step prediction results of dense wind fields; introducing dense wind field prediction results to capture the distribution characteristics of future wind power, and obtaining multi-step wind power prediction results within the day. The present invention can be applied to wind power prediction occasions where spatiotemporal wind field data is accessed to wind farms, providing accurate power prediction results for 1 to 10 hours, with a much higher accuracy than traditional prediction technologies that rely solely on on-site measurements.

Description

Technical Field

[0001] The present invention belongs to the technical field of renewable energy development and utilization, and specifically relates to a method and system for predicting intraday wind power based on deduction of spatiotemporal dense wind fields. Background Art

[0002] With the access of large-scale wind turbines to the power grid, the intermittent and non-stationary nature of wind resources has led to output power fluctuations, affecting the safe and stable control of the power grid. Therefore, it is necessary to use wind power multi-time scale accurate prediction results, rolling optimization to support power grid dispatching decisions, reduce power grid backup capacity and auxiliary costs, and thus improve the safe and stable operation level of the power grid.

[0003] Considering the random fluctuation characteristics of wind power, grasping the unknown non-stationary fluctuations of wind power in the future is the core problem that needs to be solved in wind power prediction. The data source is an important factor affecting the data-driven wind power prediction model. The existing wind power prediction methods only consider the on-site measurement information or numerical weather forecast data of wind farms, which makes it difficult to capture the dynamic characteristics of spatiotemporal wind fields, and show poor performance in prediction situations with large wind power fluctuations and long time scales. Therefore, integrating the spatiotemporal wind field data around wind farms to provide reliable predicted wind speed support for intraday wind power prediction of wind farms, thereby improving the prediction efficiency of intraday wind power, is the focus and difficulty of current intraday wind power prediction research. Summary of the invention

[0004] Purpose of the invention: In view of the difficulties in wind power prediction mentioned above, the present invention proposes a method and system for daily wind power prediction based on the deduction of spatiotemporal dense wind fields, which effectively improves the accuracy of wind power prediction.

[0005] Technical solution: In order to achieve the above-mentioned invention object, the invention proposes a method for predicting intraday wind power based on the derivation of spatiotemporal dense wind field, which comprises the following steps:

[0006] (1) Input historical wind field images, integrate physical prior knowledge, model the convection, diffusion and other uncertain motions of dense wind fields as weighted multi-order partial differential constraints, use moment estimation method to construct partial differential factors of each order to solve the likelihood function, and solve the partial differential factors of each order;

[0007] (2) The dense wind field hidden state transfer method is constructed by fusing the partial differential constraint results of the dense wind field with the uncertain motion characteristics of the input wind field through the Kalman gain, and the wind field partial differential moment estimation loss is used as the regularization term to train and obtain the multi-step prediction results of the dense wind field;

[0008] (3) The historical wind power sequence is input, and the sparse attention mechanism is used to construct the mapping relationship between the historical wind power sequence and the wind power sequence to be predicted. The mean and variance of the wind power sequence to be predicted are obtained based on the multi-step prediction results of the dense wind field. The intraday wind power prediction model is trained to obtain the intraday multi-step wind power prediction results.

[0009] Furthermore, in step (1), physical prior knowledge is integrated to model the convection, diffusion and other uncertain motions of the dense wind field as weighted multi-order partial differential constraints, including:

[0010] The spatiotemporal wind field image is represented as u(x, t), where x = (x, y) represents the spatial position of the wind field image pixel, t represents the temporal position of the wind field image, and the convective motion of the wind field is modeled as the first-order partial differential The diffusion motion of the wind field is modeled as the second-order partial differential Other uncertain motions in the wind farm are modeled as higher-order partial differentials Based on the above physical prior modeling process, the spatiotemporal derivation characteristics of the wind field are modeled as weighted multi-order partial differentials:

[0011]

[0012]

[0013] Where n is the order of partial differential, C k is the weighting coefficient, represents the k-order partial differential factor, c i,j is the weight of each partial differential factor, i and j are the orders of the partial differential of the wind field image along the x-axis and y-axis respectively.

[0014] Furthermore, in the step (1), the likelihood function is solved by constructing partial differential factors of various orders using the moment estimation method, and solving the partial differential factors of various orders includes:

[0015] The calculation of partial differential factors is equivalent to calculating the coefficient matrix:

[0016]

[0017] in, represents an equivalence relation, represents the coefficient matrix corresponding to the k-order partial differential, and * represents the convolution operation;

[0018] Expand the right side of the above equation using the Taylor expansion formula and ignore the expansion terms higher than k order to obtain:

[0019]

[0020] The numerically computable items are summarized as the state matrix Its composition can be solved directly;

[0021]

[0022] Apply the Frobenius norm to construct the likelihood function for gradient descent solution:

[0023]

[0024] Among them, Δ k represents the target matrix of k-order partial differential factors, ||·|| F represents the Frobenius norm.

[0025] Furthermore, in step (2), the dense wind field partial differential constraint result and the uncertain motion characteristics of the input wind field are fused through the Kalman gain to construct a dense wind field hidden state transfer method, which includes:

[0026] The wind field image is encoded through word embedding operation to map the explicit wind field data into the latent space It is expressed as:

[0027]

[0028] Where E represents the latent space mapping method, f(·) represents the activation function, Represents the explicit wind field image u t The feature representation mapped to the latent space, represents the weight of the i-th channel in the l-th convolution layer, b l represents the bias term of the lth layer convolution, n c represents the number of layers implementing convolution, express The number of channels.

[0029] In latent space, the wind field dynamic information derived from physical prior knowledge and the wind field latent space mapping result are fused by learnable weighted fusion, and the output is used as the wind field information hidden state at the next moment. The calculation method of wind field time series state transfer in latent space is obtained. PDE :

[0030] Φ PDE :h t ,E(u t )→h t+1

[0031]

[0032] in, is the transition matrix, which represents the wind field derivation result driven entirely by partial differential prior knowledge, K t is a learnable fusion weight, if K t= 0, indicating that the wind field time series derivation is only driven by partial differential prior knowledge, K t =1 means that the wind field derivation is obtained from the autoregressive learning of historical wind field images, and ⊙ means bitwise multiplication;

[0033] The fusion weight is determined by the Kalman gain method, which is expressed as:

[0034]

[0035] Among them, W h and W u Denote the transition matrix and wind field latent space representation E(u t ), b k represents the bias term;

[0036] The predicted wind field latent space is represented as h t+1 The wind field single-step prediction result is obtained by back-mapping to the explicit space, and the dense wind field prediction result is obtained by iteratively generating the structure through time series, which can be expressed as:

[0037]

[0038] Among them, E -1 represents the inverse operation of latent space mapping, that is, mapping the latent representation of the wind field image to the explicit space, and T represents the step size of the wind power sequence to be predicted.

[0039] Furthermore, in step (2), the dense wind field multi-step prediction loss function incorporating partial differential moment estimation loss is: It can be expressed as:

[0040]

[0041] Among them, ||·||2 represents the L2 norm, and λ1 represents the penalty coefficient of the partial differential factor calculation error regularization term.

[0042] Furthermore, in step (3), using the sparse attention mechanism to construct a mapping relationship between the historical wind power sequence and the wind power sequence to be predicted includes:

[0043] Modeling the wind power measurement series as a Gaussian distribution in and b are the mean and variance of the historical wind power probability distribution, respectively;

[0044] Apply sparse attention mechanism to construct historical wind power sequence x t-d:t And the wind power series to be predicted The mapping relationship between them is expressed as:

[0045]

[0046] in, represents the sparse attention mechanism, and They represent the mean and variance of the predicted wind power series, and d represents the step size of the historical wind power series.

[0047] Furthermore, in step (3), the mean and variance of the wind power sequence to be predicted based on the multi-step prediction results of the dense wind field are expressed as:

[0048]

[0049] in, It is a nonlinear cross-modal feature fusion operator.

[0050] The present invention also provides a daily wind power prediction system based on the derivation of a spatiotemporal dense wind field, comprising:

[0051] The dense wind field dynamic modeling module is used to integrate physical prior knowledge based on the input historical wind field images, model the convection, diffusion and other uncertain motions of the dense wind field as weighted multi-order partial differential constraints, and use the moment estimation method to construct partial differential factors of each order to solve the likelihood function and solve the partial differential factors of each order;

[0052] The wind field multi-step prediction module is used to fuse the dense wind field partial differential constraint results and the uncertain motion characteristics of the input wind field through the Kalman gain, construct a dense wind field hidden state transfer method, and use the wind field partial differential moment estimation loss as a regularization term to train and obtain the dense wind field multi-step prediction results;

[0053] The intraday wind power prediction module is used to construct the mapping relationship between the historical wind power sequence and the wind power sequence to be predicted based on the input historical wind power sequence using the sparse attention mechanism, obtain the mean and variance of the wind power sequence to be predicted based on the multi-step prediction results of the dense wind field, train the intraday wind power prediction model, and obtain the intraday multi-step wind power prediction results.

[0054] The present invention also provides a computer device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and when the programs are executed by the processors, the steps of the intraday wind power prediction method based on the derivation of spatiotemporal dense wind fields as described above are implemented.

[0055] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the intraday wind power prediction method based on derivation of a spatiotemporal dense wind field as described above are implemented.

[0056] Beneficial effects: (1) The present invention models the dynamic characteristics of dense wind farms through partial differential physics prior knowledge, and models the time series derivation characteristics of dense wind farms through time series generation structure, thereby improving the pertinence of dense wind farms in time and space and significantly reducing the computational burden, and focusing on making up for the problems of time series memory forgetting and large computational complexity in existing wind farm multi-step prediction models. (2) The present invention captures the "distribution drift" characteristics of future wind power fluctuations through dense wind farm prediction results, and outputs reliable multi-step prediction results of intraday wind power. (3) The two-stage wind power prediction model proposed in the present invention can effectively improve the wind power prediction accuracy 1 to 10 hours in advance. The model is highly efficient and can be extended to all wind farms within the coverage area of ​​dense wind farms, providing highly reliable intraday wind power prediction results, which has strong practical significance. (4) The present invention is in line with the actual economic and safe dispatching of power grids and can be applied to centralized new energy prediction centers or wind farms to improve the utilization level of wind resources, alleviate the impact of wind power non-stability and uncertainty on the power grid, meet the needs of local wind power consumption and economic and safe regulation of power grids, and can efficiently utilize computing resources, with good social and economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 The invention provides a method for modeling the dynamic derivation characteristics of a wind field based on partial differential priors;

[0058] Figure 2 It is a multi-step wind field prediction framework based on a time series generation structure of the present invention;

[0059] Figure 3 The intraday wind power prediction model proposed by the present invention;

[0060] Figure 4 The wind farm target site location prediction result in the embodiment of the present invention;

[0061] Figure 5 is a multi-step wind field image prediction result in an embodiment of the present invention;

[0062] Figure 6 The wind field image prediction result in the embodiment of the present invention;

[0063] Figure 7 The wind power prediction result in the embodiment of the present invention;

[0064] Figure 8 This is the mitigation effect of the “distribution drift” of wind power prediction in the embodiment of the present invention. DETAILED DESCRIPTION

[0065] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0066] The present invention proposes a method for predicting intraday wind power based on the derivation of a spatiotemporal dense wind field, comprising the following steps:

[0067] Step S1, modeling the dynamic derivation characteristics of the wind field based on partial differential priors: integrating physical prior knowledge, modeling the convection, diffusion and other uncertain motions of the dense wind field as weighted multi-order partial differential constraints, using the moment estimation method to construct partial differential factors of each order to solve the likelihood function, and solving the partial differential factors of each order;

[0068] Step S2, establish a dense wind field spatiotemporal inference prediction model driven by physical-data fusion: through the Kalman gain, the dense wind field partial differential constraint results and the uncertain motion characteristics of the input wind field are fused to construct a dense wind field hidden state transfer method, that is, a spatiotemporal wind field prediction-correction process based on the Kalman gain, using the time series generation structure to reduce error accumulation, and referencing the wind field partial differential moment estimation loss as a regularization term, and training to obtain the dense wind field multi-step prediction results;

[0069] Step S3, considering the common problem of "distribution drift" in wind power prediction, the multi-step prediction results of dense wind fields are introduced to capture the distribution characteristics of future wind power and obtain the multi-step wind power prediction results within the day.

[0070] like Figure 1 As shown, in step S1, the wind field dynamic derivation characteristic modeling method based on partial differential prior includes:

[0071] The spatiotemporal wind field image is represented as u(x, t), where x = (x, y) represents the spatial position of the wind field image pixel, and t represents the temporal position of the wind field image. Therefore, the convective motion of the wind field follows the constant brightness assumption, so it can be modeled as a first-order partial differential The wind field also diffuses from the high pressure zone to the low pressure zone, so it can be modeled as a second-order partial differential Circulation and other uncertain motions are represented as higher-order partial differentials Based on the above physical prior modeling process, the spatiotemporal derivation characteristics of the wind field can be modeled as weighted multi-order partial differentials:

[0072]

[0073]

[0074] Where n is the order of partial differential, C k is the weighting coefficient, represents the k-order partial differential factor, c i,j is the weight of each partial differential factor, i and j are the orders of the partial differential of the wind field image along the x-axis and y-axis respectively. The solution of each order of partial differential factors is obtained by the moment estimation method. Considering the unanalyzability of the spatiotemporal wind field u(x,t), any order of partial differential factors All of them cannot be solved directly. The indirect solution methods of partial differential factors include:

[0075] First, the calculation of the partial differential factor is equivalent to calculating the coefficient matrix:

[0076]

[0077] in, represents an equivalence relation, Represents the coefficient matrix corresponding to the k-order partial differential, and * represents the convolution operation.

[0078] Then, according to the Taylor expansion formula, the right-hand side of the above formula can be expanded and the expansion terms higher than k order can be ignored, which can be expressed as:

[0079]

[0080] The numerically computable items are summarized as the state matrix Its composition can be solved directly.

[0081]

[0082] Finally, compared with the above formula, we can know that solving the partial differential factors of each order, that is, the coefficient matrix Make an accurate estimate so that the coefficient equivalence is as close as possible. Therefore, in order to measure the state matrix Similarity with the target matrix, using the Frobenius norm to construct the likelihood function for gradient descent solution:

[0083]

[0084] Among them, Δ k represents the target matrix of k-order partial differential factors, ||·|| F represents the Frobenius norm.

[0085] Reference Figure 2 In step S2, the dense wind field spatiotemporal inference prediction model driven by physics-data fusion includes:

[0086] First, in order to enrich the feature representation, the wind field image is encoded through word embedding operation to map the explicit wind field data into the latent space It can be expressed as:

[0087]

[0088] Where E represents the latent space mapping method, f(·) represents the activation function, Represents the explicit wind field image u t Feature representation mapped to latent space. represents the weight of the i-th channel in the l-th convolution layer, b l represents the bias term of the lth layer convolution, n c represents the number of layers implementing convolution, express The number of channels.

[0089] Secondly, in the latent space, the wind field dynamic information derived from physical prior knowledge and the wind field latent space mapping result are learned and fused, and the output is used as the wind field information hidden state at the next moment, and the wind field time series state transfer calculation method Φ in the latent space is obtained. PDE :

[0090] Φ PDE :h t ,E(u t )→h t+1

[0091]

[0092] in, is the transition matrix, which represents the wind field derivation result driven entirely by partial differential prior knowledge, K t is a learnable fusion weight, if K t = 0, indicating that the wind field time series derivation is only driven by partial differential prior knowledge, K t =1 indicates that the wind field derivation is obtained from the autoregressive learning of historical wind field images. ⊙ indicates bitwise multiplication. C k Represents the weight of each order of partial differential factors.

[0093] Then, the fusion weight is determined by the Kalman gain method, which can be expressed as:

[0094]

[0095] Among them, W h and W u Denote the transition matrix and wind field latent space representation E(u t ), b k Represents the bias term.

[0096] Finally, the predicted wind field latent space is represented by h t+1 Back-mapping to the explicit space and obtaining the wind field single-step prediction result can be expressed as:

[0097]

[0098] Among them, E -1 It represents the inverse operation of latent space mapping, that is, mapping the latent representation of the wind field image to the explicit space. Thus, the single-step prediction result of the wind field can be obtained. Finally, the multi-step wind farm prediction results are obtained through iterative calculation. It can be expressed as:

[0099]

[0100] Among them, T represents the step size of the wind power sequence to be predicted. Finally, a loss function is established to train and optimize the model. Considering that there is an error in the process of calculating the partial differential factor moment estimation, the partial differential factor calculation error is introduced as a regular term to participate in the training optimization, and its loss function is It can be expressed as:

[0101]

[0102] Among them, ||·||2 represents the L2 norm, and λ1 represents the penalty coefficient of the partial differential factor calculation error regularization term.

[0103] Reference Figure 3 In step S3, the intraday wind power prediction model proposed by the present invention uses the predicted spatiotemporal wind field prediction results and historical wind power measurements as inputs, and outputs the intraday multi-step wind power prediction results.

[0104] First, the wind power measurement sequence is modeled as a Gaussian distribution in and b are the mean and variance of the historical wind power probability distribution, respectively. Since the probability distribution of historical and predicted wind power is quite different, it is necessary to normalize the historical wind power input into a smooth sequence that obeys the standard Gaussian distribution.

[0105] Next, the sparse attention mechanism is applied to construct the historical wind power sequence x t-d:t And the wind power series to be predicted The mapping relationship between them can be expressed as:

[0106]

[0107] in, Represents sparse attention mechanism. and They represent the mean and variance of the predicted wind power series. d represents the step size of the historical wind power series.

[0108] Since the dense wind field prediction results can reflect the uncertain fluctuation characteristics of wind power in the predicted time period, the dense wind field prediction results can be used to The mean and variance of the wind power sequence to be predicted are obtained by guidance, which can be expressed as:

[0109]

[0110] in, Indicates that the wind power sequence to be predicted follows a Gaussian distribution, is a nonlinear cross-modal feature fusion operator. In this embodiment, a cross-modal attention mechanism is adopted, which can be expressed as:

[0111]

[0112]

[0113] Among them, f softmax is the softmax activation function, d k Indicates the original wind power feature dimension.

[0114] Finally, the mean prediction result of the wind power sequence to be predicted is introduced as a regular term to establish the loss function Parallel training of the daily wind power prediction model:

[0115]

[0116] Among them, λ2 is the regularization term balancing factor. Thus, the intra-day multi-step wind power prediction results can be obtained.

[0117] In order to verify the performance of the method proposed in the present invention, the following experiments were carried out in the embodiment. The ERA-5 meteorological reanalysis data set developed by the European Meteorological Forecast Centre was selected. Provide high-precision onshore wind field data with a geographical resolution of 0.1° from 1950 to the present, and the minimum time resolution is 1 hour. The measured and simulated elevation of the wind farm is 10 meters. The on-site measurement data of the wind farm station used in this embodiment is located in a wind farm station in Jinhu County, Huai'an City, Jiangsu Province. The station has an installed capacity of 45MW. The on-site measurement includes wind power and wind speed data from July 2021 to the present, with a time resolution of 1 hour. Based on the above data set, three error indicators are used to evaluate the multi-step prediction results of the wind farm in this embodiment: mean deviation error (MAE), root mean square error (RMSE), and structural similarity (SSIM), which are defined as follows:

[0118]

[0119]

[0120]

[0121] in, and I x,y,i represents the predicted and measured wind field image pixel values ​​at the geographical location (x, y) and the temporal position i, respectively. h and w are the wind field spatial resolutions, and n s Indicates the number of samples in the test set. and μi are the predicted and measured wind field image means, and σ i are the standard deviations of the predicted and measured wind field images, respectively. C1 and C2 are equalization factors.

[0122] Secondly, the absolute reference error index is used to evaluate the intraday wind power forecast results:

[0123]

[0124] in, and P i The predicted value and true value of wind power at the i-th moment respectively.

[0125] The ERA-5 meteorological reanalysis data set and the on-site measurement data of a wind farm in Jinhu, Huai'an are time-series synchronized and matched, and the data from July 2021 to July 2022 are selected as the training set (8760 samples), and 4380 samples are randomly selected from July 2022 to July 2023 as test samples. This embodiment uses the convection extrapolation method and the convection-diffusion model as comparison methods, and the multi-step overall prediction error of the intraday wind field obtained by the present invention is shown in Table 1. Figure 4 The comparison between the predicted wind speed and the actual wind speed of the wind farm in Huai'an Jinhu Wind Farm is shown. It can be seen that the prediction model of the present invention fits the wind speed fluctuation well. In order to analyze the wind farm forecast results of the prediction model in detail, the detailed wind farm forecast results 1 to 10 hours in advance are as follows Figure 5 As shown in Table 1 and Figure 4 , Figure 5 It can be seen that the present invention can greatly improve the accuracy of wind farm forecasts, especially the accuracy of wind farm forecasts within 5 hours in advance, and effectively alleviate the problem of multi-step forecast error accumulation. Figure 6 As shown by Figure 6 It can be seen that the dense wind field prediction model based on partial differential knowledge prior proposed in the present invention can take into account the convection and diffusion characteristics of the spatiotemporal wind field, and also shows higher performance in terms of visual prediction effect.

[0126] Table 1 Errors of the test sample of wind field prediction one hour in advance based on ERA-5 meteorological reanalysis data

[0127] Error evaluation index MAE(m / s) RMSE(m / s) SSIM Convection Extrapolation Method 0.399 0.630 0.958 Convection-Diffusion Model 0.274 0.424 0.968 The present invention 0.248 0.389 0.973

[0128] Based on the comprehensive evaluation index of wind power, this embodiment uses the prediction model of Persistence and the prediction model removing the spatiotemporal dense wind field prediction results as the comparison model. The wind power forecast results of this embodiment one hour in advance are shown in Table 2, and the visual prediction results are shown in Figure 7 As shown in Table 2 and Figure 7It can be seen that the present invention shows superiority in wind power prediction, especially when the prediction time step increases, the intraday wind power prediction method based on dense spatiotemporal wind field prediction results proposed by the present invention shows more outstanding accuracy improvement performance. Figure 8 The contribution of the present invention in alleviating the “distribution drift” of wind power forecasting is demonstrated.

[0129] Table 2 Wind power prediction test sample error based on ERA-5 meteorological reanalysis data and data from a wind farm in Jinhu, Huai'an

[0130] Error evaluation index 1 hour in advance 3 hours in advance 5 hours in advance 10 hours in advance Persistence 5.742 7.612 9.068 11.075 Relying only on local measurements 4.477 7.335 8.430 10.289 The present invention 4.471 6.887 8.192 9.753

[0131] In summary, the intraday wind power prediction method designed by the present invention based on the time-space dense wind field derivation generation technology first models the diffusion, convection and other uncertain motions of the wind field based on multi-order partial differential priors; secondly, the moment estimation method is used to solve the partial differential factors of each order; then, the dense wind field partial differential constraint results and the wind field uncertain motion characteristics are fused through the Kalman gain to construct a dense wind field hidden state transfer method, and the wind field multi-step prediction results are obtained through the time series generation structure; finally, the wind field prediction results are used to alleviate the "distribution drift" problem in wind power prediction, and the multi-layer sparse attention mechanism is used to construct a multi-step prediction framework for intraday wind power. The present invention or a part of the present invention can be deployed in power grid departments or wind farms with high wind power prediction accuracy to guide them to adjust the power generation plan and scheduling plan based on the intraday wind power intraday prediction results to reduce the spare capacity.

[0132] Based on the same technical concept as the method embodiment, the present invention also provides a daily wind power prediction system based on time-space dense wind field derivation, including:

[0133] The dense wind field dynamic modeling module is used to integrate physical prior knowledge based on the input historical wind field images, model the convection, diffusion and other uncertain motions of the dense wind field as weighted multi-order partial differential constraints, and use the moment estimation method to construct partial differential factors of each order to solve the likelihood function and solve the partial differential factors of each order;

[0134] The wind field multi-step prediction module is used to fuse the dense wind field partial differential constraint results and the uncertain motion characteristics of the input wind field through the Kalman gain, construct a dense wind field hidden state transfer method, and use the wind field partial differential moment estimation loss as a regularization term to train and obtain the dense wind field multi-step prediction results;

[0135] The intraday wind power prediction module is used to construct the mapping relationship between the historical wind power sequence and the wind power sequence to be predicted based on the input historical wind power sequence using the sparse attention mechanism, obtain the mean and variance of the wind power sequence to be predicted based on the multi-step prediction results of the dense wind field, train the intraday wind power prediction model, and obtain the intraday multi-step wind power prediction results.

[0136] It should be understood that the intraday wind power prediction system based on the deduction of spatiotemporal dense wind fields in the embodiments of the present invention can implement all the technical solutions in the above-mentioned method embodiments, and the functions of its various functional modules can be specifically implemented according to the methods in the above-mentioned method embodiments. The specific implementation process can refer to the relevant description in the above-mentioned embodiments, which will not be repeated here.

[0137] The present invention also provides a computer device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and when the programs are executed by the processors, the steps of the intraday wind power prediction method based on the derivation of spatiotemporal dense wind fields as described above are implemented.

[0138] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the intraday wind power prediction method based on derivation of a spatiotemporal dense wind field as described above are implemented.

[0139] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, devices (systems), computer equipment or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0140] The present invention is described with reference to a flowchart of a method according to an embodiment of the present invention. It should be understood that each process in the flowchart and a combination of processes in the flowchart can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the process in the flowchart. Figure 1 A device that specifies functions in a process or multiple processes.

[0141] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A function specified in a process or multiple processes.

[0142] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 The steps of a specified function in a process or multiple processes.

Claims

1. A method for predicting intraday wind power based on the derivation of spatiotemporal dense wind fields, characterized in that: The following steps are involved: (1) Input historical wind field images, integrate physical prior knowledge, model the convection, diffusion and other uncertain motions of dense wind fields as weighted multi-order partial differential constraints, use moment estimation method to construct partial differential factors of each order to solve the likelihood function, and solve the partial differential factors of each order; (2) The dense wind field hidden state transfer method is constructed by fusing the partial differential constraint results of the dense wind field with the uncertain motion characteristics of the input wind field through the Kalman gain, and the wind field partial differential moment estimation loss is used as the regularization term to train and obtain the multi-step prediction results of the dense wind field; (3) The historical wind power sequence is input, and the sparse attention mechanism is used to construct the mapping relationship between the historical wind power sequence and the wind power sequence to be predicted. The mean and variance of the wind power sequence to be predicted are obtained based on the multi-step prediction results of the dense wind field. The intraday wind power prediction model is trained to obtain the intraday multi-step wind power prediction results.

2. The method according to claim 1, characterized in that In the step (1), physical prior knowledge is integrated to model the convection, diffusion and other uncertain motions of the dense wind field as weighted multi-order partial differential constraints, including: The spatiotemporal wind field image is represented as u(x, t), where x = (x, y) represents the spatial position of the wind field image pixel, t represents the temporal position of the wind field image, and the convective motion of the wind field is modeled as the first-order partial differential The diffusion motion of the wind field is modeled as the second-order partial differential Other uncertain motions in the wind farm are modeled as higher-order partial differentials Based on the above physical prior modeling process, the spatiotemporal derivation characteristics of the wind field are modeled as weighted multi-order partial differentials: Where n is the order of partial differential, C k is the weighting coefficient, represents the k-order partial differential factor, c i,j is the weight of each partial differential factor, i and j are the orders of the partial differential of the wind field image along the x-axis and y-axis respectively.

3. The method according to claim 2, characterized in that In the step (1), the likelihood function is solved by constructing partial differential factors of various orders using the moment estimation method. Solving the partial differential factors of various orders includes: The calculation of partial differential factors is equivalent to calculating the coefficient matrix: in, represents an equivalence relation, represents the coefficient matrix corresponding to the k-order partial differential, and * represents the convolution operation; Expand the right side of the above equation using the Taylor expansion formula and ignore the expansion terms higher than k order to obtain: The numerically computable items are summarized as the state matrix Its composition can be solved directly; Apply the Frobenius norm to construct the likelihood function for gradient descent solution: Among them, Δ k represents the target matrix of k-order partial differential factors, ||·|| F represents the Frobenius norm.

4. The method according to claim 1, characterized in that In the step (2), the dense wind field hidden state transfer method is constructed by fusing the partial differential constraint result of the dense wind field with the uncertain motion characteristics of the input wind field through the Kalman gain, including: The wind field image is encoded through word embedding operation to map the explicit wind field data into the latent space It is expressed as: Where E represents the latent space mapping method, f(·) represents the activation function, represents the explicit wind field image u t The feature representation mapped to the latent space, represents the weight of the i-th channel in the l-th convolution layer, b l represents the bias term of the lth layer convolution, n c represents the number of layers implementing convolution, express The number of channels; In latent space, the wind field dynamic information derived from physical prior knowledge and the wind field latent space mapping result are fused by learnable weighted fusion, and the output is used as the wind field information hidden state at the next moment. The calculation method of wind field time series state transfer in latent space is obtained. PDE : Φ PDE :h t ,E(in t )→h t+1 in, is the transition matrix, which represents the wind field derivation result driven entirely by partial differential prior knowledge, K t is a learnable fusion weight, if K t = 0, indicating that the wind field time series derivation is only driven by partial differential prior knowledge, K t =1 means that the wind field derivation is obtained from the autoregressive learning of historical wind field images, and ⊙ means bitwise multiplication; The fusion weight is determined by the Kalman gain method, which is expressed as: Among them, W h and W u Denote the transition matrix and wind field latent space representation E(u t ), b k represents the bias term; The predicted wind field latent space is represented as h t+1 It is back-mapped to the explicit space and the single-step wind field prediction result is obtained. The dense wind field prediction result is obtained by iteratively generating the structure through time series, which is expressed as: Among them, E -1 represents the inverse operation of latent space mapping, that is, mapping the latent representation of the wind field image to the explicit space, and T represents the step size of the wind power sequence to be predicted.

5. The method according to claim 3, characterized in that: In step (2), the dense wind field multi-step prediction loss function incorporating partial differential moment estimation loss is: It is expressed as: Among them, ||·||2 represents the L2 norm, and λ1 represents the penalty coefficient of the partial differential factor calculation error regularization term.

6. The method according to claim 1, characterized in that In the step (3), using the sparse attention mechanism to construct a mapping relationship between the historical wind power sequence and the wind power sequence to be predicted includes: Modeling the wind power history series as a Gaussian distribution in and b are the mean and variance of the historical wind power probability distribution, respectively; Apply sparse attention mechanism to construct historical wind power sequence x t-d:t And the wind power series to be predicted The mapping relationship between them is expressed as: in, represents the sparse attention mechanism, and They represent the mean and variance of the predicted wind power series, and d represents the step size of the historical wind power series.

7. The method according to claim 6, characterized in that In step (3), the mean and variance of the wind power sequence to be predicted based on the multi-step prediction results of the dense wind field are expressed as: in, It is a nonlinear cross-modal feature fusion operator.

8. A daily wind power prediction system based on the derivation of spatiotemporal dense wind fields, characterized in that: include: The dense wind field dynamic modeling module is used to integrate physical prior knowledge based on the input historical wind field images, model the convection, diffusion and other uncertain motions of the dense wind field as weighted multi-order partial differential constraints, and use the moment estimation method to construct partial differential factors of each order to solve the likelihood function and solve the partial differential factors of each order; The wind field multi-step prediction module is used to fuse the dense wind field partial differential constraint results and the uncertain motion characteristics of the input wind field through the Kalman gain, construct a dense wind field hidden state transfer method, and use the wind field partial differential moment estimation loss as a regularization term to train and obtain the dense wind field multi-step prediction results; The intraday wind power prediction module is used to construct the mapping relationship between the historical wind power sequence and the wind power sequence to be predicted based on the input historical wind power sequence using the sparse attention mechanism, obtain the mean and variance of the wind power sequence to be predicted based on the multi-step prediction results of the dense wind field, train the intraday wind power prediction model, and obtain the intraday multi-step wind power prediction results.

9. A computer device, characterized in that: include: one or more processors; Memory; as well as One or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and when the programs are executed by the processors, the steps of the intraday wind power prediction method based on spatiotemporal dense wind field derivation as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the intraday wind power prediction method based on spatiotemporal dense wind field deduction as described in any one of claims 1 to 7 are implemented.

Images