Mountain wind field prediction method based on POD and data timeliness neural network fusion

Abstract

The invention discloses a mountain wind field prediction method based on POD and data timeliness neural network fusion, and the method comprises the steps: firstly obtaining an original data set, and then carrying out the decomposition of the data in the original data set through employing a POD algorithm, and decomposing the data into a POD coherent mode and a 0-t second POD mode coefficient; then, constructing a time sequence prediction module, and predicting POD base coefficients of t-delta t seconds by using the time sequence prediction module; and finally, obtaining a mountain wind field prediction result through POD reconstruction. According to the method provided by the invention, the prediction precision of the mountain wind field is improved, and the technical problem of low prediction precision caused by time fluctuation of the POD modal coefficient due to high nonlinearity of the flow field in the prior art is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of meteorological prediction, and particularly relates to a mountain wind field prediction method based on the fusion of POD and data timeliness neural network. Background Art

[0002] Wind energy is an abundant renewable and sustainable energy source. In recent years, due to the rich wind energy resources and high land use efficiency in mountainous areas, the wind power industry has gradually developed towards mountainous areas. However, the development of mountain wind farms faces some unique challenges. In mountainous areas, the wake of the upstream mountain will significantly affect the wind speed distribution in the downstream mountainous area; at the same time, the interaction between mountains will lead to complex turbulent vortex structures, making the turbulent conditions near mountain wind farms more complex. These characteristics will cause energy output losses or reduced working efficiency in mountain wind farms. Therefore, the modeling and prediction of the mountain wind speed field are crucial for the efficient utilization of wind energy.

[0003] In the field of mountain wind field modeling and prediction, researchers Jackson and Hunt proposed a linear theoretical model based on two-dimensional hills. Although this model has been applied after considering factors such as ground roughness, it is difficult to fully consider the oncoming flow characteristics and the influence of local complex environments under complex mountain conditions, resulting in significant differences between the predicted results of the terrain wind field and the measured data. Subsequently, some researchers proposed the computational fluid dynamics (CFD) method. Although the CFD method can more accurately predict the wind field distribution of complex terrains, the computational cost is high. Then, researchers proposed the proper orthogonal decomposition (POD) method. As a reduced-order model, the POD method approximates the high-dimensional system with a low-dimensional system, significantly reducing the computational resource requirements and making up for the defect of the high computational cost of the CFD method. However, POD can only reconstruct existing data and cannot directly perform flow field prediction. In recent years, the hybrid method combining POD dimensionality reduction and machine learning (ML) prediction has received wide attention in the field of fluid dynamics. This hybrid method first extracts the main characteristic modes of the flow field through POD to reduce the data dimension, and then uses the machine learning (ML) method to predict the modal coefficients instead of the complete flow field, thereby realizing the rapid reconstruction of the flow field at low cost. However, due to the time-varying characteristics of the POD modal coefficients caused by the high nonlinearity of the flow field, the prediction difficulty is increased, the prediction performance of the ML model is reduced, and accurately capturing this dynamic characteristic often requires a deeper network structure, which is prone to overfitting. Therefore, there is an urgent need to propose an optimized wind field prediction method that fully considers the complex characteristics of the mountain wind field and the dynamic changes of the flow field, so as to more accurately predict the wind speed distribution and wake characteristics of the mountain wind field, and then formulate more effective strategies for the optimized layout and efficient operation of mountain wind farms. Summary of the Invention

[0004] The present invention aims to provide a mountain wind field prediction method based on the fusion of POD and data timeliness neural networks, so as to solve the technical problems in the existing prediction technology that due to the highly nonlinear characteristics of the flow field, the POD modal coefficients fluctuate over time, resulting in low wind field prediction accuracy and large consumption of computing resources.

[0005] The present invention provides a mountain wind field prediction method based on the fusion of POD and data timeliness neural networks, comprising the following steps:

[0006] S1. Obtain the original wind field data, obtain the original wind field data set, and divide it into a training set and a test set;

[0007] S2. Use the POD algorithm to decompose the original wind field data to obtain the POD basis modes related only to space and their corresponding basis coefficients related only to time;

[0008] In step S2, the original wind field data set is decomposed and dimensionally reduced using the POD algorithm, including the following sub-steps:

[0009] Step S201: Extract the flow field data from the original wind field data set, and sort the flow field data according to the order of extraction time to obtain the snapshot set V;

[0010] Step S202: Calculate the correlation matrix R of the snapshot set V, and the calculation formula of the correlation matrix R is as follows:

[0011]

[0012] where: N is the total number of snapshots, V is the snapshot set composed of flow fields at different times, V T is the transpose matrix of V;

[0013] Perform eigenvalue decomposition on the correlation matrix R in step S202, and the specific calculation formula is as follows:

[0014] RY = λY (2)

[0015] where: λ is the eigenvalue, and Y is the corresponding eigenvector. Step S203: Perform eigenvalue decomposition on the correlation matrix R in step S202 to obtain the eigenvalue λ corresponding to the correlation matrix R and the eigenvector Y corresponding to each eigenvalue λ;

[0016] Step S204: Calculate the eigenvalue λ in step S203 and the eigenvector Y corresponding to each eigenvalue to obtain the POD basis mode and its corresponding basis coefficient a k (t).

[0017] In step S204, the POD basis mode and its corresponding basis coefficient ak (t) The calculation formula is as follows:

[0018]

[0019] Where: k is the order, λ k is the eigenvalue, Y k is the eigenvector;

[0020]

[0021] Where: is the POD basis mode, is the transpose matrix of.

[0022] S3. Construct a time series prediction module considering data timeliness to obtain the predicted value of the POD basis coefficient;

[0023] The specific steps for constructing the time series prediction module considering data timeliness in step S3 are as follows:

[0024] Step S301: From the POD basis modes and their corresponding basis coefficients obtained in step S2, screen the basis coefficients corresponding to the POD basis modes within 0 to t seconds;

[0025] Step S302: Use the basis coefficients corresponding to the POD basis modes within 0 to t seconds to construct a freshness function, and use the freshness function to construct a loss function containing the freshness function;

[0026] The mathematical expression of the freshness function in step S302 is:

[0027]

[0028] Where: t i is the measured point time, t z is the end time of the measured data, and e is the natural constant.

[0029] The function expression of the loss function containing the freshness function in step S302 is:

[0030]

[0031] Where: S i (t) represents the predicted value of neural network prediction method i at the t-th moment; S(t) represents the measured value of neural network prediction method i at the t-th moment, and N is the total number of snapshots.

[0032] Step S303: Construct a wind field prediction neural network containing a loss function, replace the loss function in the wind field prediction neural network with a loss function having a freshness function, obtain a time series prediction module considering data timeliness, and use the training set to train the time series prediction module considering data timeliness. During the training process, adjust the weights of the training set data in the loss function until the training termination condition is met, and obtain a trained time series prediction module;

[0033] Step S304: Input the test set into the trained time series prediction module to obtain the POD basis coefficients from t to Δt seconds.

[0034] S4. Reconstruct using only the POD basis modes related to space and the predicted values of the POD basis coefficients to obtain the mountain wind field prediction result.

[0035] The mountain wind field prediction result in Step S4 has the following functional expression:

[0036]

[0037] where: a k [t, Δt] is the predicted value of the POD basis coefficients from t to Δt seconds, is the POD basis mode;

[0038] In Equation (7), M is the order when the energy proportion reaches 95%, and its functional relationship is shown in Equation (8):

[0039]

[0040] The beneficial effects of the present invention are as follows: The present invention proposes a mountain wind field prediction method based on the fusion of the POD method and neural network. First, use the POD method to extract the main characteristic modes of the flow field and perform flow field reconstruction. Then, introduce the concept of data timeliness, and construct a new loss function in combination with the freshness function. Finally, introduce the temporal weight. The mountain wind field prediction method proposed by the present invention improves the prediction accuracy of the mountain wind field and solves the technical problem of low prediction accuracy caused by the time fluctuation of the POD modal coefficients due to the high nonlinearity of the flow field in the prior art;

[0041] The present invention effectively solves the problem of large computational resource consumption in wind field prediction by integrating the POD method, machine learning technology, and the freshness function and loss function considering data timeliness. Among them, the POD method can significantly reduce the computational resource requirements and relieve the computational pressure brought by the huge amount of data processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the accompanying drawings required for the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0043] Figure 1 is the flowchart of the mountain wind field prediction method based on the fusion of POD and data timeliness neural network of the present invention;

[0044] Figure 2 is the schematic diagram of the images of the freshness function, power function, and trigonometric function of the present invention;

[0045] Figure 3 is the comparison of the optimization effects of the freshness function, power function, and trigonometric function on the long short-term memory network of the present invention. (a) is the 1-step prediction result of the long short-term memory network, (b) is the 3-step prediction result of the long short-term memory network, and (C) is the 6-step prediction result of the long short-term memory network;

[0046] Figure 4 are the prediction results of the POD basis coefficients for the 1st, 15th, and 25th modes by the LSTM optimized by the data timeliness method (DT-LSTM), the GRU optimized by the data timeliness method (DT-GRU), and the BP optimized by the data timeliness method (DT-BP); (a) is the prediction result of the POD basis coefficient for the 1st mode by DT-LSTM, (b) is the prediction result of the POD basis coefficient for the 15th mode by DT-LSTM, (c) is the prediction result of the POD basis coefficient for the 25th mode by DT-LSTM, (d) is the prediction result of the POD basis coefficient for the 1st mode by DT-GRU, (e) is the prediction result of the POD basis coefficient for the 15th mode by DT-GRU, (f) is the prediction result of the POD basis coefficient for the 15th mode by DT-GRU, (g) is the prediction result of the POD basis coefficient for the 1st mode by DT-BP; (h) is the prediction result of the POD basis coefficient for the 15th mode by DT-BP; (i) is the prediction result of the POD basis coefficient for the 25th mode by DT-BP;

[0047] Figure 5 is the box plot of the 1-step prediction errors of the DT-BP, DT-GRU, DT-LSTM, BP, GRU, and LSTM algorithms, (a) R 2 ; (b) MAE; (c) RMSE; (d) MAPE (%)

[0048] Figure 6 is the box plot of the 6-step prediction errors of the DT-BP, DT-GRU, DT-LSTM, BP, GRU, and LSTM algorithms, (a) R2 ; (b) MAE; (c) RMSE; (d) MAPE (%)

[0049] Figure 7 is the prediction error cloud chart of this method. Among them, (a) short-term BP model prediction error cloud chart; (b) long-term BP model prediction error cloud chart; (c) short-term DT-BP model prediction error cloud chart; (d) long-term DTBP model prediction error cloud chart; (e) short-term GRU model prediction error cloud chart; (f) long-term GRU model prediction error cloud chart; (g) short-term DT-GRU model prediction error cloud chart; (h) long-term DT-GRU model prediction error cloud chart; (i) short-term LSTM model prediction error cloud chart; (j) long-term LSTM model prediction error cloud chart; (k) short-term DT LSTM model prediction error cloud chart; (l) long-term DT-LSTM model prediction error cloud chart. Specific implementation manners

[0050] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0051] Such as Figure 1 , the mountain wind field prediction method based on the fusion of POD and data timeliness neural network of the present invention includes: 6 steps such as acquisition of original wind field data, decomposition of wind field data, construction of freshness function and new loss function, training and prediction of machine learning model, reconstruction and prediction of terrain wind field.

[0052] S1. Obtain the original wind field data, obtain the original wind field data set, and divide it into a training set and a test set;

[0053] Specifically: When studying the wind field characteristics near the mountain, in order to obtain accurate transient wind field data, the large eddy simulation method is used to carry out the simulation work. Among them, the large eddy simulation is a numerical simulation technology based on the filtered Navier-Stokes equation, which can more accurately analyze the large-scale vortex structure in the flow field and is suitable for wind field simulation under complex terrain.

[0054] During the simulation process, refined grid division is carried out for the area near the mountain, and the transient data snapshots of the wind speed field near the mountain are continuously collected at a time interval of 0.005 s, a total of 3600 are collected. These data snapshots cover the spatial distribution of the wind speed near the mountain at different times, and completely record the changes of the wind speed magnitude and direction over time and space.

[0055] S2. Decompose the original wind field data using the POD algorithm to obtain the POD basis modes that are only related to space and their corresponding basis coefficients that are only related to time;

[0056] In some specific embodiments, after collecting the original wind field data, it is necessary to reduce the dimension and decompose the wind field data to obtain the POD basis modes that are only related to space and the corresponding basis coefficients that are only related to time. Subsequently, the basis coefficients within a short period of time are sorted in an orderly manner. This operation aims to provide data support for subsequent model training and prediction based on time series, reflecting the attention to the variation law of short-time POD basis coefficients.

[0057] Step S201: Extract the flow field data from the original wind field dataset and sort the flow field data in the order of extraction time to obtain the snapshot set V;

[0058] Step S202: Calculate the correlation matrix R of the snapshot set V. The calculation formula of the correlation matrix R is as follows:

[0059]

[0060] where: N is the total number of snapshots, V is the snapshot set composed of flow fields at different times, and V T is the transpose matrix of V;

[0061] Perform eigenvalue decomposition on the correlation matrix R in step S202. The specific calculation method is as follows:

[0062] RY = λY (2)

[0063] where: λ is the eigenvalue and Y is the corresponding eigenvector.

[0064] Step S203: Perform eigenvalue decomposition on the correlation matrix R in step S202 to obtain the eigenvalue λ corresponding to the correlation matrix R and the eigenvector Y corresponding to each eigenvalue λ;

[0065] Step S204: Calculate the eigenvalue λ in step S203 and the eigenvector Y corresponding to each eigenvalue to obtain the POD basis mode and its corresponding basis coefficient a k (t).

[0066] The POD basis mode in step S204 and its corresponding basis coefficient a k (t) are calculated as follows:

[0067]

[0068] where: k is the order, λ k is the eigenvalue, and Y k is the eigenvector.

[0069]

[0070] Wherein: is the POD-based mode, is the transposed matrix of.

[0071] S3. Construct a time series prediction module considering data timeliness;

[0072] The specific steps for constructing the time series prediction module considering data timeliness in step S3 are as follows:

[0073] Step S301: From the POD-based mode obtained in step S2 and its corresponding basis coefficient a k (t), screen the basis coefficients corresponding to the POD-based mode within 0 to t seconds;

[0074] Step S302: Use the basis coefficients corresponding to the POD-based mode within 0 to t seconds to construct a freshness function, and use the freshness function to construct a loss function containing the freshness function;

[0075] The freshness function is a concept in data mining used to describe the degree of importance of data during the mining process. During the prediction process of POD coefficients, the data closer to the prediction time contains more useful information, has greater value for improving prediction accuracy, and should be given a greater weight, that is, it is given higher importance. In this application, the freshness function should be greater than 0 and monotonically increasing within the training data range.

[0076] Such as Figure 2 shown, when designing the freshness function, this application considered three commonly used monotonically increasing function forms and analyzed them:

[0077] 1) The definition of the power function is f(t) = t q . According to the foregoing principle, the value of the power function must be greater than 0. From the definition of the power function, the growth of the power function will increase infinitely with time. However, for POD coefficients, there is a certain limit to their data timeliness, and their importance should not be exaggerated infinitely but should gradually tend to a stable value. Correspondingly, as time goes by, the change rate of the weight should gradually become smaller. Because if the power function value tends to a larger value, the influence of historical data in the early stage on the prediction model will become extremely small and even negligible, and this situation is obviously unreasonable.

[0078] 2) The definition of the trigonometric function is f(t) = 2πarctant. For the POD coefficient, there are new problems with the freshness function in the form of a trigonometric function. Compared with other time series, the change rate of the POD number changes faster, which means that its timeliness changes more violently. The data closer to the prediction point has a greater impact on the prediction effect. Therefore, the corresponding weight should match this timeliness change, that is, the difference between f(t + 1) and f(t) should not be too small.

[0079] In summary, the freshness function applicable to this data characteristic should have the following characteristics:

[0080] (1) f(t) > 0 and is a single-point increasing function;

[0081] (2) That is, the change rate of the weight gradually decreases;

[0082] (3) f(t + 1) - f(t) should conform to the change rate of the POD coefficient change and should not be too large or too small.

[0083] Therefore, this application intends to adopt the functional form of an exponential function. However, when the traditional exponential function is applied to the prediction of the POD coefficient, it does not meet the requirements of point (2); therefore, this application constructs a freshness function form applicable to the POD coefficient. This function can fully adapt to the data characteristics of the POD coefficient, and its function is expressed as:

[0084]

[0085] where t i is the time of the measured point, and t z is the end time of the measured data.

[0086] Step S303: Construct a wind field prediction neural network containing a loss function. Replace the loss function in the wind field prediction neural network with a loss function with a freshness function to obtain a time series prediction module considering data timeliness. Use the training set to train the time series prediction module considering data timeliness. During the training process, dynamically adjust the weight of the training set data in the loss function until the training termination condition is met to obtain a trained time series prediction module;

[0087] Traditional loss functions are only used to evaluate the fitting performance and guide iterative training, assuming that the contributions of all sample errors to the overall objective are equivalent. However, the POD coefficients fluctuate greatly over time and have strong characteristics. It is not reasonable to assign the same importance to the discrimination of the prediction effects of data at all times. Since the POD coefficients change over time, the importance of their information should also change over time. The measured data closer to the prediction time has a greater impact on the prediction effect. This is the "timeliness". In this application, a freshness function in data mining is introduced to explore traditional time series prediction methods. In the prediction of POD coefficients, the closer the data is to the prediction time, the more useful information it has and the more valuable it is for improving the prediction accuracy, and a greater weight should be assigned. Combining with the freshness function, this study proposes a new loss function Q with the function of quantifying data timeliness, and its function expression is as follows:

[0088]

[0089] Where: S i (t) represents the predicted value of the neural network prediction method i at the t-th moment; S(t) represents the measured value of the neural network prediction method i at the t-th moment.

[0090] Step S304: Input the test set into the trained time series prediction module to obtain the predicted POD basis coefficients a k [t, Δt].

[0091] Specifically: In terms of hardware configuration, the training process is carried out in a hardware environment with high-performance computing capabilities. The specific hardware and software configurations are as follows: The computing device is an NVIDIA GeForce GTX 1660 SUPER GPU, the video memory is 6GB, the host memory is 64GB, the processor is an AMD Ryzen 9 3900X 12-Core Processor, the operating system is a 64-bit Windows system, and the deep learning framework is TensorFlow.

[0092] During the training process, the time series prediction module is trained with the following hyperparameter settings: The learning rate is set to 0.001, the batch size is set to 32, the optimization algorithm selects the Adam optimizer, the number of training epochs is set to 200, and the loss function uses a custom loss function, which dynamically adjusts the weights at different times based on the freshness function. The training set and the test set are divided according to a ratio of 70% and 30%.

[0093] S4. Use the POD algorithm to reconstruct the POD basis modes and the predicted POD basis coefficients for t to Δt seconds to obtain the mountain wind field prediction result.

[0094] Reconstruct the original flow field V'(x, t) at a certain t moment using the first M modes. The calculation formula of the original flow field is as follows:

[0095]

[0096] Wherein: is a basis function related only to space, and a k (t) is the basis coefficient at time t, and M is the order when the energy proportion reaches 95%.

[0097] Table 1. R of the optimization effect of three freshness functions on LSTM 2 Comparison results

[0098]

[0099] It can be seen from the data in Table 1 that whether on the training set or the test set, the R of the LSTM model combined with the freshness function 2 value is higher than that of the long short-term memory network, indicating that the introduction of the freshness function has a positive effect on improving the fitting effect and generalization ability of the model, enabling the model to better capture the data features and laws, and thus improving the prediction accuracy. The R of the exponential function-long short-term memory network model proposed in this application for the test set 2 values are 0.972, 0.953, and 0.904 respectively. The R of this model on the test set 2 value is the highest, indicating that after the exponential function is used as the freshness function and combined with LSTM, the model has a good fitting effect on the training set.

[0100] Figure 3 (a) to (c) are the comparison of the optimization effects of the freshness function, power function, and trigonometric function of the present invention on the long short-term memory network; from Figure 3 (a), it can be seen that the exponential function-long short-term memory network curve performs excellently in grasping the overall trend and closely fits the main trend of the original data. In the initial stable fluctuation stage and the relatively regular fluctuation interval in the later stage, it can accurately capture the data change trend and has a high degree of fit with the original data. Even in the time steps from 400 to 800 where the original data fluctuates frequently, this curve can to a certain extent reflect the general ups and downs of the data. Compared with the curves of other partial models, its follow-up of the original data trend is more stable, indicating that when dealing with complex fluctuating data, the combination of the exponential function and the long short-term memory network can effectively extract key trend information; from Figure 3(b) It can be seen that the exponential function-long short-term memory network curve has obvious advantages in reflecting the overall trend of the original data. From the initial fluctuations of the data to the subsequent large rises and falls and complex fluctuation stages, it can better outline the changing context of the data. At the key nodes where the original data rises and falls rapidly, although there are slight differences in the change amplitude compared with the original data, compared with some model curves, it can respond to data changes more timely and always closely follow the trend of the original data in the later fluctuations, demonstrating the good performance of this model in time series trend prediction;

[0101] Figure 4 (a)-(i) represent the prediction results of the Pod basis coefficients for the 1st, 15th, and 25th order modes of the neural network optimized by the data timeliness method (DT-ML) in the present invention, that is, the LSTM optimized by the data timeliness method (DT-LSTM), the GRU optimized by the data timeliness method (DT-GRU), and the BP optimized by the data timeliness method (DT-BP); the results show that the DT-ML model significantly improves the deficiencies of the traditional ML model in peak prediction and time delay, and the performance of the DT-BP model is improved most significantly.

[0102] The full English name of DT is Data Timeliness, and the Chinese is data timeliness.

[0103] Figure 5 (a)-(d) are the box plots of the one-step prediction errors of the DT-BP, DT-GRU, DT-LSTM, backpropagation (BP, BackPropagation), gated recurrent unit (GRU, GatedRecurrentUnit), and long short-term memory network (LSTM) algorithms; Figure 6 (a)-(d) are the box plots of the six-step prediction errors of the DT-BP, DT-GRU, DT-LSTM, BP, GRU, and LSTM algorithms. From Figure 5 and Figure 6 From the box plots of the errors, it can be seen that the improved models are significantly optimized in various indicators, the prediction errors are reduced, and the error fluctuation range is also narrowed; compared with BP, GRU, and LSTM, the R of DT-BP, DT-GRU, and DT-LSTM 2The closer the value is to 1, the closer the mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) (%) are to 0, and the narrowing of the fluctuation range of these parameters indicates that this method has varying degrees of optimization effects on the three machine learning methods. Among them, the R of DT-BP 2 is the largest, and the MAE, RMSE, and MAPE (%) are the smallest, indicating that this method has the best optimization effect on BP. Figure 5 and Figure 6 By comparison, it can be seen that the error of the 1-step prediction is much smaller than that of the 6-step prediction. The improved model has significant optimization in both single-step and multi-step predictions, especially in single-step predictions.

[0104] Figure 7 (a) to (l) are the prediction error nephograms of the mountain wind field prediction method proposed by the present invention based on the fusion of POD and data timeliness neural network, Figure 7 indicating that the error distribution near the hills is consistent with the time-varying characteristics of the vortices in these areas. As the vortices evolve downstream, the intensity of the vortex structure gradually decreases, and the prediction error also decreases accordingly. This indicates that the prediction error is closely related to the evolution of the vortices. By comparison, it is found that for all machine learning methods (BP, GRU, LSTM), the improved model effectively reduces the prediction error of the flow field near the mountains, and this improvement is significant for both short-term and medium-term predictions.

[0105] The above are only the preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A mountain wind field prediction method based on the fusion of POD and data timeliness neural network, characterized in that, The specific steps are as follows: S1. Obtain the original wind field data, get the original wind field data set, and divide it into a training set and a test set; S2. Decompose the original wind field data using the POD algorithm to obtain the POD basis modes related only to space and their corresponding basis coefficients related only to time; S3. Construct a time series prediction module considering data timeliness to obtain the predicted values of the POD basis coefficients; S4. Use the POD basis modes related only to space and the predicted values of the POD basis coefficients for reconstruction to obtain the predicted results of the mountain wind field.

2. The mountain wind field prediction method based on the fusion of POD and data timeliness neural network according to claim 1, wherein In step S2, the original wind field data set is decomposed and dimension-reduced using the POD algorithm, including the following sub-steps: Step S201: Extract the flow field data from the original wind field data set, and sort the flow field data in the order of extraction time to obtain the snapshot set V; Step S202: Calculate the correlation matrix R of the snapshot set V. The calculation formula of the correlation matrix R is as follows: Where: N is the total number of snapshots, V is the snapshot set composed of flow fields at different times, and V T is the transpose matrix of V; Step S203: Perform eigenvalue decomposition on the correlation matrix R in step S202 to obtain the eigenvalues λ corresponding to the correlation matrix R and the eigenvectors Y corresponding to each eigenvalue λ; Step S204: Calculate the eigenvalue λ and the eigenvector Y corresponding to each eigenvalue in step S203 to obtain the POD basis mode and its corresponding basis coefficient a k (t).

3. The mountain wind field prediction method based on the fusion of POD and data timeliness neural network according to claim 2, wherein, The specific calculation formula for performing eigenvalue decomposition on the correlation matrix R in step S202 is as follows: RY = λY(2) Where: λ is the eigenvalue, and Y is the corresponding eigenvector.

4. The mountain wind field prediction method based on the fusion of POD and data timeliness neural network according to claim 2, characterized in that, The POD-based mode in step S204 and its corresponding basis coefficient a k (t) The calculation formula is as follows: where: k is the order, λ k is the eigenvalue, and Y k is the eigenvector; Wherein: is the POD-based mode, is the transposed matrix of.

5. The mountain wind field prediction method based on the fusion of POD and data timeliness neural network according to claim 1, characterized in that The specific steps for constructing the time series prediction module considering data timeliness in step S3 are as follows: Step S301: From the base modes of the POD obtained in Step S2 and their corresponding base coefficients a k (t), screen the base coefficients corresponding to the POD base modes within 0 to t seconds; Step S302: Use the basis coefficients corresponding to the POD basis modes within 0 to t seconds to construct a freshness function, and use the freshness function to construct a loss function containing the freshness function; Step S303: Construct a wind field prediction neural network containing the loss function, replace the loss function in the wind field prediction neural network with the loss function with the freshness function to obtain a time series prediction module considering data timeliness, and use the training set to train the time series prediction module considering data timeliness. During the training process, adjust the weight of the training set data in the loss function until the training termination condition is met to obtain the trained time series prediction module; Step S304: Input the test set into the trained time series prediction module to obtain the POD basis coefficients a for the time period from t to Δt seconds k [t, Δt].

6. The mountain wind field prediction method based on the fusion of POD and data timeliness neural network according to claim 5, characterized in that, The mathematical expression of the freshness function in step S302 is: where: t i is the time of the measured point, t z is the end time of the measured data, and e is the natural constant.

7. The mountain wind field prediction method based on the fusion of POD and data timeliness neural network according to claim 5 or 6, characterized in that, The function expression of the loss function containing the freshness function in step S302 is: Where: S i (t) represents the predicted value of neural network prediction method i at the t-th moment; S(t) represents the measured value of neural network prediction method i at the t-th moment, and N is the total number of snapshots.

8. The mountain wind field prediction method based on the fusion of POD and data timeliness neural network according to claim 1, characterized in that The prediction result of the mountain wind field in step S4 is expressed by the following function: Where: a k [t, Δt] is the predicted value of the POD basis coefficient from t to Δt seconds, is the POD basis mode; In formula (7), M is the order when the energy ratio reaches 95%, and its functional relationship is as follows:

Images