A Mountain Wind Field Prediction Method Based on the Fusion of POD and Data Timeliness Neural Network

Abstract

This invention discloses a mountain wind field prediction method based on the fusion of POD and a data timeliness neural network. The method first acquires the original dataset, then uses the POD algorithm to decompose the data in the original dataset into POD coherent modes and POD mode coefficients from 0 to t seconds. Next, a time-series prediction module is constructed to predict the POD basis coefficients from t to Δt seconds. Finally, the mountain wind field prediction result is obtained through POD reconstruction. This invention improves the accuracy of mountain wind field prediction and solves the technical problem of low prediction accuracy caused by time fluctuations in POD mode coefficients due to the nonlinearity of the flow field in the prior art.

Description

Technical Field

[0001] This invention belongs to the field of meteorological forecasting technology, and in particular relates to a mountain wind field forecasting method based on the fusion of POD and data timeliness neural network. Background Technology

[0002] Wind energy is an abundant, renewable, and sustainable energy source. In recent years, due to the abundant wind resources and high land use efficiency in mountainous areas, the wind power industry has gradually expanded into these regions. However, the development of wind farms in mountainous areas faces some unique challenges. In mountainous areas, the wake of upstream mountains significantly affects the wind speed distribution downstream; simultaneously, the interaction between mountains leads to complex turbulent vortex structures, making the turbulence conditions near mountainous wind farms more complex. These characteristics can cause energy output losses or reduced efficiency in mountainous wind farms. Therefore, modeling and predicting wind speed fields in mountainous areas is 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 was applied after considering factors such as ground roughness, it was difficult to fully account for the characteristics of incoming flow and the influence of local complex environments under complex mountain conditions, resulting in significant differences between the predicted wind field and the measured data. Subsequently, researchers proposed the computational fluid dynamics (CFD) method. Although the CFD method can more accurately predict the wind field distribution in complex terrain, its computational cost is high. Subsequently, researchers proposed the intrinsic orthogonal decomposition (POD) method. As a reduced-order model, the POD method significantly reduces the computational resource requirements by approximating a high-dimensional system with a low-dimensional system, thus compensating for the high computational cost of the CFD method. However, POD can only reconstruct existing data and cannot directly predict the flow field. In recent years, hybrid methods combining POD dimensionality reduction and machine learning (ML) prediction have received widespread attention in the field of fluid dynamics. This hybrid method first extracts the main characteristic modes of the flow field through POD to reduce data dimensionality, and then uses machine learning (ML) methods to predict modal coefficients to replace the complete flow field, thereby achieving rapid reconstruction of the flow field at low cost. However, due to the highly nonlinear flow field, the time fluctuation characteristics of POD modal coefficients increase the prediction difficulty and reduce the prediction performance of ML models. Furthermore, accurately capturing this dynamic feature often requires a deeper network structure, which is prone to overfitting. Therefore, it is urgent to propose a wind field prediction optimization method that fully considers the complex characteristics of mountain wind fields and the dynamic changes of the flow field, so as to more accurately predict the wind speed distribution and wake characteristics of mountain wind fields, and thus formulate more effective strategies for the optimized layout and efficient operation of mountain wind farms. Summary of the Invention

[0004] This invention aims to provide a mountain wind field prediction method based on the fusion of POD (Programmable Observation Device) and a data-time-sensitive neural network. This addresses the technical problems of low wind field prediction accuracy and high computational resource consumption in existing prediction techniques, caused by the highly nonlinear characteristics of the flow field leading to fluctuations in POD modal coefficients over time.

[0005] This invention provides a method for predicting mountain wind fields based on the fusion of POD (Programmable Observation Depth) and a data timeliness neural network, comprising the following steps:

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

[0007] S2. The original wind field data is decomposed using the POD algorithm to obtain the POD basis modes that are spatially dependent only and their corresponding basis coefficients that are time-dependent only;

[0008] Step S2 uses the POD algorithm to decompose and reduce the dimensionality of the original wind field dataset, including the following sub-steps:

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

[0010] Step S202: Calculate the correlation matrix R of the snapshot set V. The formula for calculating the correlation matrix R is as follows:

[0011]

[0012] Where: N is the total number of snapshots, and V is the set of snapshots formed by the flow field at different times. T Let V be the transpose of V;

[0013] The eigenvalue decomposition of the correlation matrix R in step S202 is performed, 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 eigenvalues ​​λ corresponding to the correlation matrix R and the eigenvector Y corresponding to each eigenvalue λ;

[0016] Step S204: Calculate the eigenvalues ​​λ and the corresponding eigenvectors Y from step S203 to obtain the POD basis modes. and its corresponding basic coefficient a k (t).

[0017] POD fundamental mode in step S204 and its corresponding basic coefficient ak The formula for calculating (t) is as follows:

[0018]

[0019] Where: k is the order, λ k Y is the eigenvalue. k Feature vector;

[0020]

[0021] in: For POD fundamental mode, for The transpose of .

[0022] S3. Construct a time series forecasting module that takes into account the timeliness of data to obtain the predicted value of POD base coefficient;

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

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

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

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

[0027]

[0028] Where: t i t represents the actual measurement time. z t is the last time of the measured data, and e is a natural constant.

[0029] The loss function containing the freshness function in step S302 is expressed as follows:

[0030]

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

[0032] Step S303: Construct a wind field prediction neural network with 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 that considers data timeliness, train the time series prediction module that considers data timeliness using a training set, and adjust the weight of the training set data in the loss function during the training process until the training termination condition is met to 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 the mountain wind field prediction results by using the spatially dependent POD basis modes and POD basis coefficients.

[0035] Step S4 Mountain Wind Field Prediction Results The function is expressed as follows:

[0036]

[0037] Where: a k [t, Δt] represents the predicted POD base coefficient values ​​for seconds t to Δt. It is the POD fundamental mode;

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

[0039]

[0040] The beneficial effects of this invention are as follows: This invention proposes a mountain wind field prediction method based on the fusion of the POD method and neural network. First, the main characteristic modes of the flow field are extracted using the POD method and the flow field is reconstructed. Then, the concept of data timeliness is introduced, and a new loss function is constructed by combining the freshness function. Finally, time weights are introduced. The mountain wind field prediction method proposed in this invention improves the prediction accuracy of mountain wind fields and solves the technical problem of low prediction accuracy caused by the time fluctuation of POD mode coefficients due to the high nonlinearity of the flow field in the prior art.

[0041] This invention effectively solves the problem of high computational resource consumption in wind field prediction by integrating the POD method, machine learning technology, and freshness and loss functions that consider the timeliness of data. Among them, the POD method can significantly reduce the computational resource requirements and alleviate the computational pressure caused by the large amount of data processing. Attached Figure Description

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0043] Figure 1 This is a 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 This is a graphical representation of the freshness function, power function, and trigonometric function of this invention;

[0045] Figure 3 This is a 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. (a) is the prediction result of the long short-term memory network in 1 step, (b) is the prediction result of the long short-term memory network in 3 steps, and (c) is the prediction result of the long short-term memory network in 6 steps.

[0046] Figure 4 These are the prediction results of DT-LSTM, DT-GRU, and DT-BP optimized for 1st, 15th, and 25th modalities, respectively, based on data timeliness optimization methods. (a) DT-LSTM prediction results for 1st modality POD basis coefficients, (b) DT-LSTM prediction results for 15th modality POD basis coefficients, and (c) DT-LSTM prediction results for 25th modality POD basis coefficients. The prediction results are as follows: (d) DT-GRU prediction results for the POD basis coefficients in the first mode; (e) DT-GRU prediction results for the POD basis coefficients in the 15th mode; (f) DT-GRU prediction results for the POD basis coefficients in the 15th mode; (g) DT-BP prediction results for the POD basis coefficients in the first mode; (h) DT-BP prediction results for the POD basis coefficients in the 15th mode; (i) DT-BP prediction results for the POD basis coefficients in the 25th mode.

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

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

[0049] Figure 7 These are the prediction error contour maps of this method. Among them, (a) is the prediction error contour map of the short-term BP model; (b) is the prediction error contour map of the long-term BP model; (c) is the prediction error contour map of the short-term DT-BP model; (d) is the prediction error contour map of the long-term DT-BP model; (e) is the prediction error contour map of the short-term GRU model; (f) is the prediction error contour map of the long-term GRU model; (g) is the prediction error contour map of the short-term DT-GRU model; (h) is the prediction error contour map of the long-term DT-GRU model; (i) is the prediction error contour map of the short-term LSTM model; (j) is the prediction error contour map of the long-term LSTM model; (k) is the prediction error contour map of the short-term DT-LSTM model; and (l) is the prediction error contour map of the long-term DT-LSTM model. Detailed Implementation

[0050] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

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

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

[0053] Specifically, when studying the wind field characteristics near mountains, large eddy simulation (LES) is used to obtain accurate transient wind field data. LES is a numerical simulation technique based on the filtered Navier-Stokes equations, which can accurately analyze the large-scale eddy structure in the flow field and is suitable for wind field simulation under complex terrain.

[0054] During the simulation, a fine grid was divided for the area near the mountain. Transient data snapshots of the wind speed field near the mountain were continuously collected at 0.005s intervals, for a total of 3,600 snapshots. These data snapshots cover the spatial distribution of wind speed near the mountain at different times and fully record the changes in wind speed magnitude and direction over time and space.

[0055] S2. The original wind field data is decomposed using the POD algorithm to obtain the POD basis modes that are spatially dependent only and their corresponding basis coefficients that are time-dependent only;

[0056] In some specific implementations, after acquiring the raw wind field data, the data needs to be reduced in dimensionality and decomposed to obtain the spatially relevant POD fundamental modes and their corresponding time-relevant basic coefficients. Subsequently, the basic coefficients over a short period are systematically organized. This operation aims to provide data support for subsequent time-series-based model training and prediction, reflecting a focus on the short-term variation patterns of POD basic coefficients.

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

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

[0059]

[0060] Where: N is the total number of snapshots, and V is the set of snapshots formed by the flow field at different times. T Let V be the transpose of V;

[0061] The eigenvalue decomposition of the correlation matrix R in step S202 is performed 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 eigenvalues ​​λ corresponding to the correlation matrix R and the eigenvectors Y corresponding to each eigenvalue λ;

[0065] Step S204: Calculate the eigenvalues ​​λ and the corresponding eigenvectors Y from step S203 to obtain the POD basis modes. and its corresponding basic coefficient a k (t).

[0066] POD fundamental mode in step S204 and its corresponding basic coefficient a k The formula for calculating (t) is as follows:

[0067]

[0068] Where: k is the order, λ k Y is the eigenvalue. k Eigenvectors.

[0069]

[0070] in: For POD fundamental mode, for The transpose of .

[0071] S3. Construct a time-series forecasting module that takes into account the timeliness of data;

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

[0073] Step S301: The fundamental mode of POD obtained from step S2 and its corresponding basic coefficient a k In (t), the basis coefficients corresponding to the POD basis modes within the range of 0 to t seconds are selected;

[0074] Step S302: Construct a freshness function using the basis coefficients corresponding to the POD basis modes within 0 to t seconds, and construct a loss function containing the freshness function using the freshness function;

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

[0076] like Figure 2 As shown, this application considers three commonly used monotonically increasing function forms when designing the freshness function, and analyzes them:

[0077] 1) The definition of a power function is f(t) = t q According to the aforementioned principle, the power function must have a value greater than 0. As defined, the power function increases infinitely over time. However, for the POD coefficient, the timeliness of the data has certain limits; its importance should not be infinitely exaggerated but should gradually tend towards a stable value. Correspondingly, the rate of change of the weights should gradually decrease over time. This is because if the power function value tends to be large, the influence of previous historical data on the prediction model becomes extremely small, or even negligible, which is clearly unreasonable.

[0078] 2) The trigonometric function is defined as f(t) = 2πarctant. For the POD coefficient, the freshness function in trigonometric form presents a new problem: compared with other time series, the POD number changes at a faster rate, which means its timeliness changes more drastically. Data closer to the prediction point has a greater impact on the prediction effect, so 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, a freshness function suitable for this type of data characteristic should have the following characteristics:

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

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

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

[0083] Therefore, this application intends to adopt the exponential function form. However, when the traditional exponential function is applied to the prediction of POD coefficients, it does not meet the requirements of point (2). Therefore, this application constructs a freshness function form suitable for POD coefficients. This function can fully adapt to the data characteristics of POD coefficients, and its function expression is as follows:

[0084]

[0085] Among them, t i t represents the actual measurement time. z This represents the last time of the measured data.

[0086] Step S303: Construct a wind field prediction neural network with 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 that considers data timeliness, train the time series prediction module that considers data timeliness using a training set, dynamically adjust the weight of the training set data in the loss function during the training process until the training termination condition is met, and obtain the trained time series prediction module.

[0087] Traditional loss functions are only used to evaluate fitting performance and guide iterative training, assuming that all sample errors contribute equally to the overall objective. However, the POD coefficient fluctuates greatly over time and has strong eigenvalues, making it unreasonable to assign equal importance to the prediction effect of data at all times. As the POD coefficient changes over time, its information importance should also change; the closer the measured data is to the prediction time, the greater its impact on the prediction effect—this is "timeliness." This application introduces the freshness function from data mining to explore traditional time-series prediction methods. In POD coefficient prediction, data closer to the prediction time contains more useful information and is more valuable for improving prediction accuracy, thus it should be given greater weight. Combining the freshness function, this study proposes a new loss function Q that quantifies data timeliness, and its function expression is as follows:

[0088]

[0089] Wherein: S i S(t) represents the predicted value of neural network prediction method i at time t; S(t) represents the measured value of neural network prediction method i at time t.

[0090] Step S304: Input the test set into the trained time series prediction module to obtain the predicted POD basis coefficients a from t to Δt seconds. 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 configuration is as follows: the computing device is an NVIDIA GeForce GTX 1660 SUPER GPU with 6GB of video memory, 64GB of host memory, an AMD Ryzen 9 3900X 12-Core Processor, a 64-bit Windows operating system, and the deep learning framework is TensorFlow.

[0092] During training, the time series prediction module was trained using the following hyperparameter settings: learning rate was set to 0.001, batch size was set to 32, the Adam optimizer was selected as the optimization algorithm, the number of training epochs was set to 200, a custom loss function was used, which dynamically adjusted the weights at different times based on the freshness function, and the training set and test set were divided in a ratio of 70% and 30%, respectively.

[0093] S4. The POD algorithm is used to reconstruct the POD basis mode and the predicted POD basis coefficients from t to Δt seconds to obtain the mountain wind field prediction results.

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

[0095]

[0096] in: For basis functions that are only spatially dependent, a k (t) represents the basis coefficients at time t, and M represents the order when the energy percentage reaches 95%.

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

[0098]

[0099] As can be seen from the data in Table 1, the R-value of the LSTM model incorporating the freshness function is higher on both the training and test sets. 2 The R value is higher than that of the Long Short-Term Memory (LSTM) network, indicating that the introduction of the freshness function has a positive effect on improving the model's fitting effect and generalization ability. It allows the model to better capture data features and patterns, thereby improving prediction accuracy. The R value of the exponential function-LSTM network model proposed in this application on the test set is... 2 The values ​​were 0.972, 0.953, and 0.904, respectively, indicating that the model achieved R-values ​​on the test set. 2 The highest value indicates that the model has a good fit on the training set after the exponential function is combined with LSTM as a freshness function.

[0100] Figure 3 (a) to (c) are comparisons of the optimization effects of the freshness function, power function, and trigonometric function of this invention on long short-term memory networks; 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, closely following the main trend of the original data. In the initial stable fluctuation phase and the relatively regular fluctuation range in the later stages, it accurately captures the data change trend, showing a high degree of consistency with the original data. Even in the 400 to 800 time steps where the original data fluctuates frequently, the curve can still reflect the general fluctuations of the data to a certain extent. Compared with other model curves, its following of the original data trend is more stable, demonstrating that when processing 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 a significant advantage in reflecting the overall trend of the original data. From the initial fluctuations of the data to the subsequent large increases and decreases and complex fluctuation stages, it can effectively outline the changes in the data. At the key points of rapid rise and fall in the original data, although there are slight differences in the magnitude of change compared to the original data, it can respond to data changes more promptly than some model curves, and it always maintains a close following of 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) to (i) represent the prediction results of the data timeliness method optimized neural network (DT-ML), namely data timeliness method optimized LSTM (DT-LSTM), data timeliness method optimized GRU (DT-GRU), and data timeliness method optimized BP (DT-BP) for the 1st, 15th, and 25th order modes in this invention. The results show that the DT-ML model significantly improves the shortcomings of the traditional ML model in peak prediction and time delay, among which the DT-BP model has the most significant performance improvement.

[0102] DT stands for Data Timeliness.

[0103] Figure 5 (a) to (d) are box plots of the one-step prediction error of the DT-BP, DT-GRU, DT-LSTM, BackPropagation (BP), GatedRecurrentUnit (GRU), and Long Short-Term Memory (LSTM) algorithms. Figure 6 (a) to (d) are box plots of the prediction errors of the 6-step DT-BP, DT-GRU, DT-LSTM, BP, GRU, and LSTM algorithms. Figure 5 and Figure 6 The error box plots show that the improved model has significantly improved across all metrics, with reduced prediction error and a narrower error fluctuation range. Compared to BP, GRU, and LSTM, DT-BP, DT-GRU, and DT-LSTM have higher R-values. 2The values ​​are closer to 1, and the Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE) (%) are closer to 0. Furthermore, the reduced fluctuation range of these parameters indicates that this method has varying degrees of optimization effects on all three machine learning methods; among them, DT-BP's R... 2 The highest value and the lowest values ​​for MAE, RMSE, and MAPE (%) indicate that this method has the best optimization effect on BP. Figure 5 and Figure 6 The comparison shows that the error of 1-step prediction is much smaller than that of 6-step prediction. The improved model has significant optimization in both single-step and multi-step prediction, especially in single-step prediction.

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

[0105] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection 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 raw wind field data to obtain the raw wind field dataset, and divide it into training set and test set; S2. The original wind field data is decomposed using the POD algorithm to obtain the POD basis modes that are spatially dependent only and their corresponding basis coefficients that are time-dependent only; S3. Construct a time series forecasting module that takes into account the timeliness of data to obtain the predicted value of POD base coefficient; The specific steps for constructing the time series prediction module that considers data timeliness in step S3 are as follows: Step S301: The fundamental mode of POD obtained from step S2 and their corresponding base coefficients In the process, the basis coefficients corresponding to the POD basis modes within the range of 0 to t seconds are selected; Step S302: Construct a freshness function using the basis coefficients corresponding to the POD basis modes within 0~t seconds, and construct a loss function containing the freshness function using the freshness function; Step S303: Construct a wind field prediction neural network with 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 that considers data timeliness, train the time series prediction module that considers data timeliness using a training set, and adjust the weight of the training set data in the loss function during the training process until the training termination condition is met to obtain a trained time series prediction module. Step S304: Input the test set into the trained time series prediction module to obtain the POD basis coefficients for seconds t to Δt. ; S4. Reconstruct the mountain wind field prediction results using the spatially dependent POD basis modes and POD basis coefficients.

2. The mountain wind field prediction method based on the fusion of POD and data timeliness neural network according to claim 1, characterized in that, Step S2 uses the POD algorithm to decompose and reduce the dimensionality of the original wind field dataset, including the following sub-steps: Step S201: Extract flow field data from the original wind field dataset and sort the flow field data according to the extraction time to obtain snapshot set V; Step S202: Calculate the correlation matrix R of the snapshot set V. The formula for calculating the correlation matrix R is as follows: （1） in: N For the total number of snapshots, A set of snapshots of the flow field at different times. for The transpose of the matrix; Step S203: Perform eigenvalue decomposition on the correlation matrix R from step S202 to obtain the eigenvalues ​​λ corresponding to the correlation matrix R and the eigenvectors corresponding to each eigenvalue λ. Y ; Step S204: For the eigenvalues ​​λ and the eigenvectors corresponding to each eigenvalue described in step S203... Y Calculations were performed to obtain the POD fundamental mode. and their corresponding base coefficients .

3. 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 eigenvalue decomposition of the correlation matrix R in step S202 is performed, and the specific calculation formula is as follows: （2） Where: λ is the eigenvalue. This is the corresponding feature vector.

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, POD fundamental mode in step S204 and their corresponding base coefficients The calculation formula is as follows: （3） Where: k is the order. For eigenvalues, eigenvectors; （4） in: For POD fundamental mode, for The transpose 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 mathematical expression of the freshness function in step S302 is: （5） in: The actual measurement time. t is the last time of the measured data, and e is a natural constant.

6. The mountain wind field prediction method based on the fusion of POD and data timeliness neural network according to claim 1 or 5, characterized in that, The loss function containing the freshness function in step S302 is expressed as follows: （6） in: This represents the predicted value of neural network prediction method i at time t; This represents the measured value of neural network prediction method i at time t. N This represents the total number of snapshots.

7. The mountain wind field prediction method based on the fusion of POD and data timeliness neural network according to claim 1, characterized in that, Step S4 Mountain Wind Field Prediction Results The function is expressed as follows: （7） in: The predicted values ​​of the POD base coefficients for seconds t to Δt. It is the POD fundamental mode; In equation (7) The order is the one at which the energy percentage reaches 95%, and its functional relationship is as follows: （8）。

Images