A Wind Power Prediction Method and System Based on PatchTST

Abstract

The present invention discloses a wind power prediction method and system based on PatchTST, which relates to the technical field of wind power prediction. A wind power prediction system based on PatchTST includes: a data acquisition module, a data preprocessing module, a PatchTST prediction module, a prediction result output module, a wind power value evaluation module, and an edge computing optimization module. The present invention realizes adaptive time window adjustment through a dynamic Patch generation mechanism, improving the flexibility and computational efficiency of prediction; existing wind power prediction methods usually adopt a fixed time window, which is difficult to adapt to the drastic fluctuations or long-term stability of wind speed; the present invention dynamically adjusts the Patch length according to the wind speed change rate, making short-time Patches applicable to wind speed mutation scenarios and long-time Patches applicable to wind speed stable scenarios, thereby improving the prediction accuracy, reducing the consumption of computing resources, and enhancing the adaptability of the model under different wind conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power prediction, and particularly to a wind power prediction method and system based on PatchTST. Background Art

[0002] Wind power generation, as an important renewable energy source, plays a key role in the global energy structure transformation. However, due to the randomness, intermittency, and uncertainty of wind speed, the accuracy of wind power prediction directly affects the operation scheduling of wind farms and the stability of the power grid. Wind power prediction mainly relies on modeling wind speed, wind direction, meteorological factors, and historical wind turbine data to predict future power generation.

[0003] Currently, common wind power prediction methods include: physical model-based prediction methods, statistics-based prediction methods, and deep learning-based prediction methods. Although existing methods have improved the accuracy of wind power prediction, there are still problems such as fixed time windows, difficulty in adapting to wind speed changes, lack of multi-scale feature extraction capabilities, ignoring the spatial dependence relationship between wind turbines, low computational efficiency, difficulty in real-time deployment at the wind turbine end, lack of uncertainty estimation, and affecting wind power scheduling. Summary of the Invention

[0004] The present invention proposes a wind power prediction method based on PatchTST, which combines dynamic Patch generation, multi-scale feature extraction, spatio-temporal calculation, and probability confidence interval calculation to optimize the accuracy and stability of wind power prediction, is applicable to different wind farm scenarios, and improves the reliability of wind power grid connection scheduling.

[0005] A wind power prediction method based on PatchTST includes:

[0006] Data preprocessing: Obtain the wind speed, wind direction, meteorological data, and historical wind power data of the wind farm, and perform normalization, interpolation filling, and denoising processing on the data;

[0007] Dynamic Patch generation: According to the wind speed change rate, adaptively adjust the Patch length so that short-time Patches are applicable to wind speed mutation scenarios, and long-time Patches are applicable to wind speed stable scenarios;

[0008] Multi-scale feature extraction: Construct short-time scale Patches, medium-time scale Patches, and long-term time scale Patches, and optimize the information interaction of different scales through weighted fusion;

[0009] Spatio-temporal calculation: Introduce the spatial dependence calculation between wind turbines in the PatchTST prediction module, calculate the dependence relationship between wind turbines in combination with the wind speed propagation path, and establish a spatial dependence matrix between wind turbines;

[0010] Forecast result output: The PatchTST forecast module is used to calculate future wind power, and the probability model is combined to provide a forecast confidence interval to improve the credibility of the forecast results;

[0011] Wind power value evaluation: Based on the long-term time scale PatchTST prediction results and combined with the wind energy impact relationship between wind turbines calculated in space-time, a long-term value evaluation system for wind farms is constructed; the system comprehensively considers the stability of wind energy resources, wind farm power utilization rate and wind abandonment rate evaluation.

[0012] As a preferred technical solution of the present invention, dynamic patch generation adopts an adaptive window mechanism to adjust the patch length according to the change rate of wind speed, avoiding the limitation of fixed window length, so that the patch length can be automatically adjusted with the change of wind speed. The calculation method is as follows: Where L t is the patch length of the current time step, σ t is the wind speed change rate, α and β are adjustment parameters, and ∈ is a constant to prevent division by zero.

[0013] As a preferred technical solution of the present invention, multi-scale feature extraction adopts a weight allocation strategy based on the attention mechanism, which dynamically calculates the importance of patches at different time scales to optimize the fusion of wind speed features, so that short-term wind speed changes and long-term wind speed trends can be effectively modeled at the same time. The weight calculation formula is as follows: Among them, S i is the feature importance of patches at different time scales, λ is a hyperparameter, and N is the number of patches to ensure that both high-frequency and low-frequency features can play a role in the prediction and improve the model's perception of wind speed signals at different time scales.

[0014] As a preferred technical solution of the present invention, the space-time calculation adopts an adaptive attention mechanism based on the topological structure of the wind turbine, optimizes the spatial synergy between the wind turbines by dynamically calculating the wind speed correlation weights between the wind turbines, and calculates the wind speed propagation relationship using the following adjacency matrix: Among them A ij represents the wind speed correlation weight between wind turbines i and j, d ij is the geographical distance between the two wind turbines, and τ is a hyperparameter that controls the effect of distance on the correlation.

[0015] As a preferred technical solution of the present invention, the PatchTST prediction module adopts a low-rank decomposition attention mechanism to reduce computational complexity, improve the computational efficiency of wind power prediction, and optimize the deployment of the PatchTST prediction module on the wind turbine edge computing device to make it suitable for low-power computing environment, while reducing the computational burden of the prediction model and enhancing the real-time prediction capability of the wind farm.

[0016] As a preferred technical solution of the present invention, the prediction result output combines a probability model to calculate the prediction confidence interval, so as to improve the robustness of wind power prediction, and adaptively adjusts the confidence interval according to the historical error of the prediction data, so that the prediction on different time scales can have corresponding confidence evaluations, and improves the reliability of wind farm scheduling decisions.

[0017] As a preferred technical solution of the present invention, in combination with short-term meteorological forecast data, the PatchTST prediction value is corrected through a wind speed-wind power coupling correction mechanism to improve the credibility of wind farm power prediction, and the following correction formula is adopted:

[0018] P corr =P raw +γ(W forecast -W real ), where P corr is the corrected wind power prediction value, P raw is the original prediction value, W forecast and W real are the predicted and true wind speeds respectively, and γ is an adjustment coefficient, so that the prediction deviation can be adaptively adjusted according to the actual wind speed change, and the real-time performance of power prediction is improved.

[0019] As a preferred technical solution of the present invention, the knowledge distillation technology is adopted to compress the large-scale PatchTST prediction module into a lightweight model to optimize the calculation performance, improve the real-time prediction ability of the fan control system, so that the prediction model can adapt to different computing environments and reduce the hardware cost, and improve the feasibility of deployment at the fan end.

[0020] As a preferred technical solution of the present invention, the long-term value evaluation system of the wind farm includes:

[0021] Calculation of wind energy resource stability: Based on the long-term time-scale PatchTST prediction results, analyze the inter-annual wind power volatility of the wind farm, and construct a wind energy resource stability index to evaluate the long-term stability of the wind energy supply of the wind farm, and optimize the wind farm site selection and expansion plan;

[0022] Evaluation of wind farm power utilization rate: Combine the wind energy influence relationship between fans calculated in space-time, calculate the overall wind energy conversion efficiency of the wind farm, analyze the influence of different fan layouts on the power utilization rate, and provide suggestions for optimizing the fan layout;

[0023] Evaluation of curtailment rate: Use long-term prediction error statistics and probability confidence interval to analyze the uncertainty of wind power, calculate the curtailment risk of the wind farm, provide an optimization strategy for wind power energy storage, and adjust the wind power grid connection strategy.

[0024] A wind power prediction system based on PatchTST, comprising:

[0025] A data acquisition module, which is used to obtain the wind speed, wind direction, meteorological data and historical wind power data of a wind farm and connect to an external weather forecasting system;

[0026] A data preprocessing module, which is used to perform data normalization, missing value filling, outlier detection and denoising to optimize the data quality and improve the reliability of the data;

[0027] A PatchTST prediction module, which is used to perform dynamic Patch generation, multi-scale Patch calculation and spatio-temporal calculation;

[0028] A prediction result output module, which is used to calculate the future wind power and provide a prediction confidence interval in combination with a probability model;

[0029] A wind power value evaluation module, which is used to evaluate the long-term value of a wind farm based on the long-term time-scale PatchTST prediction results and combine the spatio-temporal calculation of the wind energy influence relationship between wind turbines, so as to optimize the construction and operation strategies of the wind farm;

[0030] An edge computing optimization module, which is used to optimize the inference efficiency of the PatchTST prediction module on a wind turbine control system or a wind farm server to improve the real-time prediction ability.

[0031] The present invention has the following advantages:

[0032] Through a dynamic Patch generation mechanism, the present invention realizes adaptive time window adjustment, improves the flexibility and computational efficiency of prediction; existing wind power prediction methods usually adopt a fixed time window, which is difficult to adapt to the drastic fluctuations or long-term stability of wind speed; the present invention dynamically adjusts the Patch length according to the wind speed change rate, making short-time Patches applicable to wind speed mutation scenarios and long-time Patches applicable to wind speed stable scenarios, thereby improving the prediction accuracy, reducing the consumption of computing resources, and enhancing the adaptability of the model under different wind conditions.

[0033] Through a multi-scale feature extraction method, the present invention realizes the comprehensive learning of short, medium and long-term wind speed patterns, improving the comprehensiveness of prediction; traditional methods mostly use a single time scale for prediction, which is difficult to learn short-term wind speed fluctuations and long-term trends at the same time. The present invention constructs short-time scale Patches, medium-time scale Patches and long-time scale Patches, and optimizes the information interaction of different time scales through weighted fusion, enabling the model to take into account short-time drastic changes and long-term trends, and improving the prediction stability and generalization ability.

[0034] The present invention uses a spatio-temporal calculation method to introduce the spatial dependence calculation between wind turbines, improving the overall prediction accuracy of the wind farm. Existing wind power prediction methods mainly rely on the time series of individual wind turbines, ignoring the spatial cooperation between wind turbines, resulting in limited prediction accuracy. The present invention introduces the spatial dependence calculation between wind turbines in the PatchTST prediction module, constructs a spatial dependence matrix between wind turbines by combining the wind speed propagation path, thereby optimizing the wind speed propagation effect, improving the accuracy of the overall wind power prediction, and reducing the impact of local wind speed changes on the prediction results.

[0035] The present invention constructs a long-term value evaluation system for the wind farm through the wind power value evaluation module, based on the long-term time-scale PatchTST prediction results and combining the spatio-temporal calculation method between wind turbines. Existing wind power planning methods usually rely on historical wind speed statistics and lack a dynamic assessment of the long-term stability of wind energy, resulting in limited scientificity of the wind farm site selection and expansion plans. The present invention uses long-term wind speed trend prediction and wind energy impact analysis between wind turbines to calculate the stability of wind energy resources, power utilization rate, and curtailment rate, optimize the wind farm construction site selection, improve the reliability of wind power investment decisions, and enhance the adaptability of the wind farm under different climate conditions.

[0036] The present invention optimizes the Transformer calculation structure through the PatchTST prediction module to improve the calculation efficiency and is applicable to real-time prediction at the wind turbine end. The traditional Transformer has a high calculation complexity and consumes a large amount of computing resources during long-time series prediction, making it difficult to run in real time at the wind turbine end. The present invention optimizes the PatchTST calculation structure, reduces redundant calculations, and improves the calculation efficiency, enabling wind power prediction to be applicable to the low-power computing environment of the wind turbine end or the wind farm server, enhancing the real-time prediction ability of the wind farm.

[0037] The present invention calculates the prediction confidence interval through a probability model to improve the reliability and scheduling feasibility of wind power prediction. Traditional methods mostly use single-value prediction and fail to provide an assessment of the uncertainty of the prediction, affecting the safety of wind power scheduling. The present invention combines a probability model to calculate the prediction confidence interval, enabling the prediction result to not only provide a single-point prediction value but also output a prediction range with different confidence levels, enhancing the risk control ability of wind farm scheduling and reducing the instability of wind power grid connection. Description of the Drawings

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only schematic diagrams of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on the provided drawings.

[0039] Figure 1It is a schematic structural diagram of a wind power prediction system based on PatchTST adopted in an embodiment of the present invention. Specific embodiments

[0040] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all 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 scope of protection of the present invention.

[0041] Embodiment 1, a wind power prediction method based on PatchTST:

[0042] This embodiment provides a wind power prediction method based on PatchTST, which improves the accuracy and adaptability of wind power prediction by introducing dynamic Patch generation, multi-scale feature extraction, spatio-temporal calculation and probability model to calculate the confidence interval. This method is mainly applicable to short-term and medium- and long-term wind power prediction in wind farms, and supports prediction requirements under different wind speed change patterns, including the following steps:

[0043] Step S1, data preprocessing: Obtain the wind speed, wind direction, meteorological data and historical wind power data of the wind farm, and perform normalization, interpolation filling and denoising processing on the data to ensure the integrity and consistency of the data;

[0044] S101, data collection:

[0045] Data collection is the basis of wind power prediction. By integrating multiple data sources, a comprehensive data input is constructed to improve the prediction ability of the model. The data collection module of the present invention includes the following contents:

[0046] Wind turbine sensor data collection: Collect key parameters such as wind speed, wind direction, blade angle, wind turbine speed, and power generation of wind turbines in the wind farm to record the operating state information of the wind turbines.

[0047] Meteorological monitoring data collection: Collect environmental variables such as temperature, air pressure, humidity, and precipitation through meteorological stations or remote sensing devices to improve the prediction accuracy of wind speed changes.

[0048] External meteorological forecast data access: Connect to the short-term weather forecasting system to collect information such as wind speed, wind direction, and temperature for the next 1h, 3h, 6h, and 24h to provide reference for wind power prediction.

[0049] Historical wind power data storage: Record the historical data of wind speed, wind direction, and power generation in the wind farm to support model training, error correction, and long-term trend analysis.

[0050] Distance and geographical distribution data of wind turbines: Record the geographical location information of wind turbines to construct the spatial dependence relationship between wind turbines and optimize the spatial feature extraction ability of the PatchTST prediction module.

[0051] S102, Data cleaning: Since the wind farm sensors may be affected by environmental interference and there are outliers, missing values or noises in data collection, it is necessary to clean the data to improve the reliability of the data.

[0052] Outlier detection, detection based on statistical method (Z-score method): X is the current data value, μ represents the mean value, σ is the standard deviation. If |Z| > 3, then this data point is considered an outlier.

[0053] The following method is used to correct abnormal data: For short-term anomalies (less than 10 minutes), linear interpolation is used to fill; for long-term anomalies (more than 1 hour), the mean value of historical data is used for replacement to reduce the impact of abnormal data on prediction.

[0054] S103, Data normalization is performed using the Z-score method to ensure that data with different features have the same numerical range.

[0055] S104, Interpolation filling: Since the wind farm data may have missing values due to sensor failures, data loss, etc., the present invention uses KNN interpolation to fill the missing values to improve data integrity.

[0056] KNN interpolation method: Where X i is the nearest k data points, is the weight, and d i represents the distance between data points.

[0057] Missing data filling strategy: If the missing time is short (5 - 10 minutes), linear interpolation is used to fill; if the missing time is long (1 - 3 hours), KNN interpolation is used to fill; if a certain variable is missing for a long time (more than 6 hours), then this variable is abandoned as an input to avoid misleading the model prediction.

[0058] S105, Data denoising: Wind speed data is usually affected by short-term disturbances (eddy currents, wind shear), resulting in high-frequency noise interfering with the prediction model. The present invention uses moving average filtering to remove high-frequency noise and improve data smoothness: Where N is the sliding window size (4 time steps), and X if is the original wind speed data.

[0059] S106, Data Storage and Management: Historical data storage, using a distributed database to store long time series data, supporting long-term prediction and analysis of wind farms. Caching mechanism, caching data within the recent period (within 1 hour) to improve the data reading speed during model inference. Data version management, storing the original data, cleaned data, and normalized data in different versions for easy model debugging and retrospective analysis.

[0060] Step S2, Dynamic Patch Generation: According to the wind speed change rate, adaptively adjust the Patch length, making short-time Patches applicable to scenarios of sudden wind speed changes and long-time Patches applicable to scenarios of stable wind speed, so as to optimize the calculation efficiency and improve the prediction accuracy, enhancing the prediction adaptability under different wind conditions;

[0061] S201, Dynamic Patches Adapt to Different Wind Conditions: In traditional time series modeling, a fixed time window is usually used for data slicing, while this method dynamically adjusts the Patch length according to the wind speed change rate to improve the adaptability of the model.

[0062] Short-time Patches (5min - 30min): Applicable to scenarios of sudden wind speed changes (severe fluctuations in wind speed under the influence of thunderstorms, gusts, typhoons). Use short-window Patches for local feature extraction to improve short-term prediction accuracy. Long-time Patches (1h - 6h): Applicable to scenarios of stable wind speed (stable wind speed stage at night). Use long-time Patches for feature extraction to improve the ability to capture the periodic trend of wind speed. Patch adaptability calculation, the Patch length is dynamically adjusted according to the fluctuation degree of the wind speed, enabling the model to flexibly respond to different wind conditions and improving the prediction robustness.

[0063] S202, Adaptive Window Mechanism, this method adopts a dynamic adjustment strategy of Patches based on the wind speed change rate to optimize the time window, enabling the Patch length to automatically adjust with the change of the wind speed;

[0064] Patch length calculation formula: where L t is the Patch length at the current time step, σ t is the wind speed change rate at the current time step, which is the absolute value of the difference between the wind speeds at the current and the previous time steps, α and β are adjustment parameters, controlling the change range of the Patch length, and ∈ is a constant to prevent division by zero (taking 10^-6).

[0065] Dynamic adjustment strategy of Patch length:

[0066] When the wind speed changes violently (σ t is greater than the threshold), the Patch length is shortened, enabling the model to capture rapidly changing features and enhancing the short-term prediction ability.

[0067] For example, when the wind speed fluctuates significantly in a short time (gust or thunderstorm weather), the Patch length is reduced to 5 - 10 minutes. When the wind speed changes smoothly (σ t is less than the threshold), the Patch length is extended, enabling the model to focus on long-term trends and improving the modeling ability of wind speed periodicity. For example, at night when the wind speed is relatively stable, the Patch length is extended to 3 - 6 hours to improve the robustness of the prediction.

[0068] S203, Patch generation process: Calculate the change in wind speed between the current time step and the previous time step to determine the change trend of the wind speed; calculate the Patch length through a formula to ensure that the Patch becomes shorter during rapid short-term wind speed changes and longer during long-term stability; slice the wind speed, wind direction, and meteorological data according to the calculated Patch length to form input samples; the data after Patch cutting is input into the PatchTST prediction module for wind power prediction.

[0069] S204, Optimization strategy for dynamic Patch generation:

[0070] 1. Set the upper and lower limits of the Patch length. To avoid computational instability caused by overly long or short Patches, set thresholds.

[0071] 2. Smooth the change in Patch length. To avoid drastic fluctuations in the Patch length, use a moving average method to smooth the change in the Patch length.

[0072] 3. Combine with the historical trend of wind speed. During the calculation of the Patch length, adjust it in combination with the historical wind speed trend. If the recent wind speed has fluctuated greatly, reduce the Patch length; if the recent wind speed is stable, appropriately increase the Patch length.

[0073] Compared with the traditional fixed-window Patch division method, the dynamic Patch generation method of the present invention has the following advantages: improving short-term prediction accuracy, being able to capture changes faster during rapid wind speed changes by shortening the Patch length, and improving short-term prediction ability. Enhancing the learning ability of long-term trends, enabling the model to learn the periodic changes of wind speed by extending the Patch length during wind speed stability, and improving long-term prediction accuracy. Optimizing computational efficiency, reducing the number of Patch divisions during wind speed stability, reducing computational overhead, and improving the inference speed of the PatchTST prediction module. Improving the adaptability of the model, adopting an adaptive adjustment strategy to enable the Patch length to be dynamically adjusted according to the wind speed, and improving the adaptability of the prediction model under different wind conditions.

[0074] Step S3, Multi-scale Feature Extraction: Construct short-term scale Patches, medium-term scale Patches, and long-term scale Patches, and optimize the information interaction of different scales through weighted fusion to improve the model's ability to capture the periodic changes of wind speed;

[0075] S301, Construct multi-scale Patches: Traditional time series prediction methods usually adopt a fixed time window, while this method constructs Patches of multiple time scales and fuses their information to improve the model's time feature perception ability.

[0076] S302, Multi-scale Feature Fusion: Since the time variation of wind speed has multi-scale features, this method adopts a weighted fusion strategy based on the attention mechanism to dynamically calculate the importance of Patches of different time scales, enabling the model to adaptively allocate attention weights at different time scales.

[0077] Multi-scale weight calculation, the feature importance of Patches of different time scales is calculated by the attention mechanism: where S i is the feature importance of Patches of different time scales, λ is a hyperparameter that controls the sensitivity of attention allocation; N is the number of Patches to ensure that both high-frequency and low-frequency features can play a role in prediction and improve the model's perception ability of wind speed signals of different time scales.

[0078] Weighted fusion strategy, calculate the feature importance of Patches of different time scales: According to the change pattern of historical data, calculate the relative contribution values of Patches of different scales. For example, in the case of large short-term fluctuations in wind speed, the weight of short-term scale Patches is higher. Dynamically weighted fusion of multi-scale features: According to the calculated weights, fuse the information of short, medium, and long-term time scales: where F fusion is the fused feature, W i is the weight of the i-th time scale Patch, and F i is the Patch feature of the corresponding time scale.

[0079] S303, Multi-scale Patch calculation process: Construct Patches of different time scales, and divide the input data into short, medium, and long-term Patches according to the set time window. Calculate the feature importance of Patches, adopt weighted calculation based on the attention mechanism, and assign appropriate weights to each Patch. Fuse multi-scale features, and combine the information of different scales through the weighted fusion strategy to improve the model's adaptability. Input the PatchTST prediction module, and the processed multi-scale features enter the PatchTST prediction module for wind power prediction.

[0080] Compared with traditional time series prediction methods, the multi-scale feature extraction method of the present invention has the following advantages: enhancing the prediction ability for short-term mutations of wind speed. Through short-time scale patches, it can quickly respond to wind speed mutations and improve short-term prediction accuracy. Optimizing the long-term trend modeling ability of wind speed. Through long-time scale patches, it can capture the periodic changes of wind speed and improve long-term prediction accuracy. Improving the stability of prediction results. By using the attention mechanism to weight and fuse patches of different time scales, it ensures that short-, medium-, and long-term features can all play a role in prediction, improving the stability and robustness of prediction results. Optimizing computational efficiency. Through adaptive time scale selection and low-rank decomposition calculation strategies, it improves the computational efficiency of the model, enabling it to process ultra-long time series data more efficiently.

[0081] Step S4, spatio-temporal calculation: Introduce the calculation of spatial dependence between wind turbines in the PatchTST prediction module, combine the wind speed propagation path to calculate the dependence relationship between wind turbines, and establish a spatial dependence matrix between wind turbines to optimize the wind speed propagation effect and improve the accuracy of overall wind power prediction;

[0082] S401, modeling the spatial dependence relationship between wind turbines: The wind turbines in a wind farm do not operate in isolation, and there is a wind speed influence effect between adjacent wind turbines. For example, (the upstream wind turbine in the wake effect will affect the wind speed of the downstream wind turbine, resulting in wind power attenuation; the local wind speed propagation is affected by factors such as terrain, wind turbine arrangement, and meteorological conditions within the wind farm, and has a certain spatial correlation). In order to accurately model the spatial cooperation between wind turbines, this method constructs a spatial dependence matrix between wind turbines for spatial calculation in the PatchTST prediction module.

[0083] S402, calculation of the spatial dependence matrix: The spatio-temporal calculation adopts an adaptive attention mechanism based on the wind turbine topological structure. By dynamically calculating the wind speed correlation weights between wind turbines, it optimizes the spatial cooperation between wind turbines, and uses the following adjacency matrix to calculate the wind speed propagation relationship: where A ij represents the wind speed correlation weight between wind turbines i and j, d ij is the geographical distance between the two wind turbines, and τ is a hyperparameter that controls the influence of distance on correlation. If d ij is greater than the set maximum influence distance, then set A ij to 0 to avoid the interference of irrelevant wind turbine features on the prediction results. Through normalization, the sum of the spatial dependence weights of each wind turbine is 1, ensuring the rationality of the calculation.

[0084] S403, Wind speed propagation path calculation: The propagation of wind speed within a wind farm is affected by multiple factors, including: wind speed and direction information. The wind speed propagation paths are different under different wind directions, which affect the dependence relationship between wind turbines. Wind turbine layout. Wind turbines with a compact arrangement have stronger spatial interaction than those with a dispersed arrangement. Topographical factors. Undulating terrain may affect the transmission mode of wind speed.

[0085] This method calculates the wind speed propagation path based on a dynamic wind speed propagation model and optimizes the dependence relationship between wind turbines: Calculate the influence of upstream wind turbines on downstream wind turbines, W j adj = W j + ∑ i A ij W ia where W j adj is the wind speed considering spatial influence, W j is the wind speed of the original wind turbine j, W ia is the wind speed of wind turbine i; Wake effect modeling, where P j is the corrected power of wind turbine i, is the wind power without considering the wake effect, and η is the wake loss coefficient.

[0086] S404, Space-time fusion strategy: In order to simultaneously consider the time dependence and spatial synergy of wind speed, this method adopts a space-time fusion strategy. Short-time scale Patch calculates short-term wind speed changes; medium-time scale Patch calculates wind speed trends; long-time scale Patch calculates wind speed periodicity; the spatial dependence matrix calculates the spatial propagation characteristics of wind speed. The final wind power prediction is calculated by the following formula: where W i，t is the wind speed weight of wind turbine i at time t, P i，t is the wind power of wind turbine i at time t, and T is the total duration.

[0087] Compared with traditional methods, the space-time calculation method of the present invention has the following advantages: improving the prediction accuracy of wind turbine clusters, enhancing the spatial adaptability of the prediction model by calculating the spatial dependence between wind turbines, making the prediction results more accurate. Enhancing the understanding of the wind speed propagation path, combining the wind turbine topological structure and the wind speed propagation model to improve the modeling ability of the wind speed propagation path. Optimizing the calculation efficiency, adopting a sparse adjacency matrix and a low-rank decomposition attention mechanism to reduce the calculation amount and improve the inference speed. Enhancing the applicability to wind farms, being applicable to the wind turbine layouts of different wind farms and having strong generalization ability.

[0088] Step S5, Prediction Result Output: The PatchTST prediction module is used to calculate the future wind power, and a probability model is combined to provide a prediction confidence interval, improving the credibility of the prediction results and providing a reliable basis for the power dispatch of the wind farm.

[0089] S501, Future Wind Power Prediction: This method uses the PatchTST prediction module to calculate the future wind power based on the prediction results of steps such as dynamic Patch generation, multi-scale feature extraction, and spatio-temporal calculation: P t = PatchTST(X t ), where P t is the predicted value of wind power at time t, and X t is the input of the PatchTST prediction model, including historical wind speed, wind direction, meteorological data, and the spatial relationship between wind turbines.

[0090] S502, Low-Rank Decomposition Attention Mechanism: Due to the high computational complexity of the Transformer structure, this method introduces a low-rank decomposition attention mechanism in the PatchTST prediction module to reduce the computational amount and improve the computational efficiency of wind power prediction: A = UV T , where A is the original attention matrix, and U and V are the matrices after low-rank decomposition, reducing the computational complexity.

[0091] S503, Confidence Interval Calculation: Wind power prediction has a certain degree of uncertainty. To improve the credibility of the prediction results, this method uses a probability model to calculate the prediction confidence interval and dynamically adjusts it in combination with historical errors. The probability model is Bayesian deep learning.

[0092] S504, Since wind power prediction depends on wind speed input, short-term wind speed errors may lead to wind power deviations. To improve the prediction accuracy, this method combines short-term meteorological forecast data for wind speed-wind power coupling correction, using the following correction formula: P corr = P raw + γ(W forecast - W real ), where P corr is the corrected predicted value of wind power, P raw is the original predicted value, W forecast and W real are the predicted and actual wind speeds respectively, and γ is an adjustment coefficient, enabling the prediction deviation to be adaptively adjusted according to the actual wind speed change and improving the real-time performance of power prediction.

[0093] S505, Lightweight Model Optimization: To improve the deployment feasibility of the PatchTST prediction module at the wind turbine end, this method uses knowledge distillation technology to compress the large-scale PatchTST prediction model into a lightweight model.

[0094] Knowledge distillation technology:

[0095] Use a large model (Teacher-Model) to guide the training of a small model (Student-Model):

[0096] L = α1L hard +(1 - α1)L soft , where L hard is the original supervised loss, and L soft is the knowledge distillation loss, and α1 is the adjustment parameter.

[0097] S506, Cross-regional model generalization: Since the wind speed distributions of different wind farms are different, traditional models have poor generalization ability in cross-regional prediction. This method uses self-supervised learning methods (through time series contrast learning and pre-training with unlabeled data) for pre-training the wind power prediction model to improve the cross-regional adaptability of the PatchTST prediction module.

[0098] S507, Prediction result output process: Use PatchTST to predict future wind power; use a probability model to estimate the confidence interval to improve prediction credibility; combine meteorological data to adjust the wind power prediction value; use knowledge distillation to optimize the calculation efficiency and improve the adaptability for deployment at the fan end, and use self-supervised learning to improve the cross-regional generalization ability of the model.

[0099] Step S6, Based on the long-term time-scale PatchTST prediction results, combined with the wind energy influence relationship between wind turbines calculated by space-time, construct a long-term value evaluation system for wind farms. The system comprehensively considers the evaluation of wind energy resource stability, wind farm power utilization rate, and curtailment rate to optimize the planning, operation, and wind energy utilization efficiency of wind farms.

[0100] S601, Calculation of wind energy resource stability: Predict the wind speed trend through long-term time-scale PatchTST, analyze the inter-annual volatility of wind energy resources in the wind farm, and calculate the wind energy resource stability index (WRSI): where σ W is the standard deviation of wind power prediction, and P W is the average wind power. If the WRSI is low (small wind energy fluctuations), then this area is suitable for large-scale wind power development; if the WRSI is high (large wind energy fluctuations), then an energy storage system is needed to assist in smoothing the power output.

[0101] S602, Evaluation of wind farm power utilization rate: Combine the wind energy influence relationship between wind turbines calculated by space-time to analyze the wind energy transmission and shielding effects (such as wake effect) between wind turbines.

[0102] Calculate the ratio of the actual power output of the wind turbine to the theoretical maximum power output, and define the wind farm power utilization rate (WPU): Where P actual is the actual power output of the wind farm, and P theoreical is the maximum wind power output theoretically achievable by the wind speed. If the WPU is low, it indicates that there is room for optimization in the wind turbine layout. For example: adjust the wind turbine arrangement method to reduce the mutual interference between wind turbines and improve the wind energy conversion efficiency; adjust the wind turbine operating parameters (such as blade angle, rotation speed, etc.) to optimize the performance of the wind turbine.

[0103] S603, curtailment rate assessment: Combine the long-term wind power prediction error, calculate the curtailment rate (WCR) of the wind farm, and optimize the energy storage and grid dispatching strategies: Where P curtailed is the wind power that fails to be grid-connected, and P total is the total power generation of the wind farm.

[0104] Calculate the future curtailment rate using the probability confidence interval, and combine the energy storage system to provide an optimization plan: If the WCR is higher than the threshold, it is recommended to increase the energy storage capacity to store the excess wind energy and improve the wind power consumption rate; if the WCR is lower than the threshold, optimize the matching strategy between wind power and the grid load to reduce the impact of wind power fluctuations on the grid stability.

[0105] S604, suggestions for wind farm site selection and operation optimization:

[0106] Combine the calculation results of S601 - S603 to provide wind farm site selection, energy storage optimization and operation strategies:

[0107] Priority development area: High wind energy stability (low WRSI), high power utilization rate (high WPU), suitable for large-scale wind farms.

[0108] Energy storage adaptation area: Large wind energy fluctuations (high WRSI), relatively high curtailment rate (high WCR), it is recommended to configure an energy storage system to improve the wind energy utilization efficiency.

[0109] Grid connection optimization area: Rich wind energy resources, but great challenges in wind power grid connection, it is recommended to adopt intelligent power dispatching to improve the matching between the wind farm and the grid.

[0110] Example 2, a wind power prediction system based on PatchTST, as shown in Figure 1 shown, including the following modules:

[0111] 1. Data acquisition module, used to obtain the real-time and historical data of the wind farm, and provide high-quality input data for the PatchTST prediction module. It includes the following units:

[0112] Wind turbine sensing unit: Collects parameters such as wind speed, wind direction, blade angle, and rotational speed of wind turbines in a wind farm, and provides information on the operating status of the wind turbines.

[0113] Meteorological monitoring unit: Connects to the environmental monitoring station of the wind farm, collects meteorological data, including temperature, air pressure, humidity, precipitation, etc., to improve prediction accuracy.

[0114] External meteorological data access unit: Obtains short-term (1h, 3h, 6h) and long-term (24h) meteorological forecast data from meteorological agencies to improve the modeling ability of wind speed trends.

[0115] Historical data storage unit: Stores the historical wind power data of the wind farm, supporting model training and long-term trend analysis.

[0116] 2. Data preprocessing module, which is used to clean and optimize the collected data to ensure that the PatchTST prediction module can perform calculations with high-quality data input. It includes the following units:

[0117] Normalization unit: Performs normalization on data with different dimensions to ensure that the numerical ranges of the input data are consistent, avoiding numerical deviations during model training.

[0118] Missing value filling unit: Adopts methods such as linear interpolation, spline interpolation, or KNN interpolation to fill in missing data, reducing the impact of data gaps on the model.

[0119] Outlier detection unit: Uses statistical methods to detect outliers and correct them.

[0120] Denoising unit: Adopts methods such as wavelet transform or moving average to remove high-frequency noise, improving the stability of prediction data to reduce the interference of short-term wind speed fluctuations on the prediction model.

[0121] 3. PatchTST prediction module, which is the core calculation module and performs wind power prediction based on the PatchTST structure. It includes the following units:

[0122] Dynamic Patch generation unit: Adaptively adjusts the Patch length according to the wind speed change rate, making short-time Patches suitable for wind speed mutation scenarios and long-time Patches suitable for wind speed stable scenarios, optimizing the calculation efficiency and improving the prediction accuracy. Adopts an adaptive window mechanism to adjust the Patch length according to the wind speed change rate, avoiding the limitations of a fixed window length, so that the Patch length can be automatically adjusted with the wind speed change.

[0123] Multi-scale Patch calculation unit: construct short-time scale patches (5min-30min), medium-time scale patches (1h-6h) and long-time scale patches (1d-7d), and optimize the information interaction of different scales through weighted fusion to improve the model's ability to capture periodic changes in wind speed. Adopt a weight allocation strategy based on the attention mechanism, dynamically calculate the importance of patches at different time scales to optimize the fusion of wind speed features, so that short-term wind speed changes and long-term wind speed trends can be effectively modeled at the same time.

[0124] Space-time calculation unit: By calculating the spatial dependency between wind turbines and combining the wind speed propagation path to calculate the dependency between wind turbines, a spatial dependency matrix between wind turbines is established to optimize the wind speed propagation effect and improve the accuracy of the overall wind power prediction. An adaptive attention mechanism based on the wind turbine topology is adopted to dynamically calculate the wind speed correlation weights between wind turbines to optimize the spatial synergy between wind turbines, and the adjacency matrix is used to calculate the wind speed propagation relationship. A low-rank decomposition attention mechanism is used to reduce the computational complexity, improve the computational efficiency of wind power prediction, and optimize the deployment of the PatchTST prediction module on the wind turbine edge computing device.

[0125] 4. The prediction result output module is used to output the wind power prediction results and provide the prediction confidence interval in combination with the probability model to improve the credibility of the prediction. It includes the following units:

[0126] Wind power prediction unit: predicts future wind power output values based on PatchTST and provides short-term (1h), medium-term (6h) and long-term (24h) prediction results.

[0127] Probability confidence interval calculation unit: Combined with the probability model, the prediction confidence interval is calculated to improve the robustness of wind power prediction, and the confidence interval is adaptively adjusted according to the historical error of the prediction data, so that predictions at different time scales can have corresponding confidence assessments, thereby improving the reliability of wind farm scheduling decisions.

[0128] Wind speed-wind power coupling correction unit: Combined with short-term meteorological forecast data, the PatchTST prediction value is corrected through the wind speed-wind power coupling correction mechanism to improve the credibility of wind farm power prediction.

[0129] 5. Wind power value evaluation module: It is used to evaluate the long-term value of wind farms and optimize the construction and operation strategies of wind farms based on the long-term time scale PatchTST prediction results and the wind energy impact relationship between wind turbines calculated in space-time. It includes the following units:

[0130] Wind Resource Stability Calculation Unit: Analyze interannual wind power fluctuations based on long-term wind speed forecasts, calculate the Wind Resource Stability Index (WRSI), and optimize wind farm site selection and expansion plans;

[0131] Wind farm power utilization rate evaluation unit: Combining the wind energy interaction among wind turbines, analyze the influence of the wind turbine layout on the power utilization rate, and provide optimization suggestions for the wind turbine arrangement to improve the wind energy conversion efficiency;

[0132] Abandoned wind rate evaluation unit: Based on the long-term prediction error statistics and probability confidence intervals, calculate the possible abandoned wind risks of the wind farm, and combine the energy storage system status to optimize the wind power grid connection and energy storage scheduling strategies to improve the wind power consumption capacity.

[0133] 6. Edge computing optimization module, used to improve the computing efficiency of the prediction system, make it suitable for the deployment of the wind turbine side or the wind farm server, and improve the real-time performance of wind power prediction. It includes the following units:

[0134] Model compression unit: Adopt the knowledge distillation technology to compress the large-scale PatchTST prediction model into a lightweight model to optimize the computing performance, improve the real-time prediction ability of the wind turbine control system, enable the prediction model to adapt to different computing environments and reduce the hardware cost, and improve the feasibility of deployment on the wind turbine side.

[0135] Self-supervised learning unit: Use the self-supervised learning method to pre-train the wind power prediction model to enhance the generalization ability of the PatchTST prediction module in different wind farm scenarios, make it applicable to wind farms with different wind speed distributions, improve the accuracy of cross-regional wind power prediction, and reduce the dependence on labeled data.

[0136] Edge inference optimization unit: Optimize the inference efficiency of the PatchTST prediction module on the wind turbine side or the wind farm server, improve the computing speed, reduce the energy consumption, and enable the model to run stably on low-power devices.

[0137] The specific implementation manners described above further elaborate on the purpose, technical solutions, and beneficial effects of the present invention. It should be understood that the above is only the specific implementation manner of the present invention and is not used 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 shall be included in the protection scope of the present invention.

Claims

1. A wind power prediction method based on PatchTST, characterized in that Including: Data preprocessing: Obtain the wind speed, wind direction, meteorological data and historical wind power data of the wind farm, and perform normalization, interpolation filling and denoising processing on the data; Dynamic Patch generation: According to the wind speed change rate, adaptively adjust the Patch length so that short-time Patches are applicable to the wind speed mutation scenario, and long-time Patches are applicable to the wind speed stable scenario; Multi-scale feature extraction: Construct short-time scale Patches, medium-time scale Patches and long-time scale Patches, and optimize the information interaction of different scales through weighted fusion; Space-time calculation: Introduce the spatial dependence calculation between wind turbines in the PatchTST prediction module, calculate the dependence relationship between wind turbines in combination with the wind speed propagation path, and establish the spatial dependence matrix between wind turbines; Prediction result output: Use the PatchTST prediction module to calculate the future wind power, and combine the probability model to provide the prediction confidence interval to improve the credibility of the prediction result; Wind power value evaluation: Based on the long-time scale PatchTST prediction result, combine the wind energy influence relationship between wind turbines calculated by space-time calculation, and construct a long-term value evaluation system for the wind farm; The system comprehensively considers the stability of wind energy resources, the power utilization rate of the wind farm and the curtailment rate evaluation.

2. The method for predicting wind power based on PatchTST according to claim 1, wherein, Dynamic Patch generation adopts an adaptive window mechanism, adjusts the Patch length according to the change rate of wind speed, avoids the limitations of a fixed window length, enables the Patch length to be automatically adjusted with the change of wind speed, and its calculation method is as follows: Where L t is the Patch length at the current time step, σ t is the wind speed change rate, α and β are adjustment parameters, and ∈ is a constant to prevent division by zero.

3. A wind power prediction method based on PatchTST according to claim 1, characterized in that Multi-scale feature extraction adopts a weight allocation strategy based on the attention mechanism. By dynamically calculating the importance of patches at different time scales, the wind speed feature fusion is optimized so that short-term wind speed changes and long-term wind speed trends can be effectively modeled at the same time. The weight calculation formula is as follows: Among them, S i is the feature importance of patches at different time scales, λ is a hyperparameter, and N is the number of patches to ensure that both high-frequency and low-frequency features can play a role in the prediction and improve the model's perception of wind speed signals at different time scales.

4. A wind power prediction method based on PatchTST according to claim 1, characterized in that Space-time calculation adopts an adaptive attention mechanism based on the fan topology structure, optimizes the spatial cooperation between fans by dynamically calculating the wind speed correlation weights between fans, and calculates the wind speed propagation relationship using the following adjacency matrix: where A ij represents the wind speed correlation weight between fans i and j, d ij is the geographical distance between the two fans, and τ is a hyperparameter that controls the influence of distance on the correlation.

5. A wind power prediction method based on PatchTST according to claim 1, characterized in that, The PatchTST prediction module adopts a low-rank decomposition attention mechanism to reduce the computational complexity, improve the computational efficiency of wind power prediction, and optimize the deployment of the PatchTST prediction module on the edge computing device of the wind turbine, making it applicable to the low-power computing environment, while reducing the computational burden of the prediction model and enhancing the real-time prediction ability of the wind farm.

6. A wind power prediction method based on PatchTST according to claim 1, characterized in that The prediction result output combines the probability model to calculate the prediction confidence interval to improve the robustness of the wind power prediction, and adaptively adjusts the confidence interval according to the historical error of the prediction data, so that the prediction of different time scales can have corresponding confidence evaluations, and improve the reliability of the wind farm scheduling decision.

7. A wind power prediction method based on PatchTST according to claim 1, characterized in that, Combined with short-term meteorological forecast data, the PatchTST prediction value is corrected through a wind speed-wind power coupling correction mechanism to improve the credibility of wind farm power prediction, and the following correction formula is adopted: P corr = P raw + γ(W forecast - W real ), where P corr is the corrected wind power prediction value, P raw is the original prediction value, W forecast and W real are the predicted and actual wind speeds respectively, and γ is an adjustment coefficient, enabling the prediction deviation to be adaptively adjusted according to the actual wind speed change, thereby improving the real-time performance of power prediction.

8. A wind power prediction method based on PatchTST according to claim 1, characterized in that, Adopt the knowledge distillation technology to compress the large-scale PatchTST prediction module into a lightweight model to optimize the computational performance, improve the real-time prediction ability of the wind turbine control system, make the prediction model adaptable to different computational environments and reduce the hardware cost, and improve the feasibility of deployment at the wind turbine end.

9. A wind power prediction method based on PatchTST according to claim 1, characterized in that, The long-term value evaluation system of the wind farm includes: Calculation of wind energy resource stability: Based on the long-time scale PatchTST prediction result, analyze the inter-annual wind power volatility of the wind farm, and construct a wind energy resource stability index to evaluate the long-term stability of the wind energy supply of the wind farm and optimize the wind farm site selection and expansion plan; Evaluation of the power utilization rate of the wind farm: Combine the wind energy influence relationship between wind turbines calculated by space-time calculation, calculate the overall wind energy conversion efficiency of the wind farm, analyze the influence of different wind turbine layouts on the power utilization rate, and provide optimization suggestions for wind turbine layout; Curtailment rate evaluation: Use long-term prediction error statistics and probability confidence intervals to analyze the uncertainty of wind power, calculate the curtailment risk of the wind farm, provide optimization strategies for wind power energy storage, and adjust the wind power grid connection strategy.

10. A wind power prediction system based on PatchTST, characterized in that, The system applies any one of the methods for predicting wind power based on PatchTST described in claims 1 to 9 above, including: The data acquisition module is used to obtain the wind speed, wind direction, meteorological data and historical wind power data of the wind farm and connect to the external weather forecasting system; The data preprocessing module is used to perform data normalization, missing value filling, outlier detection and denoising to optimize the data quality and improve the reliability of the data; The PatchTST prediction module is used to perform dynamic Patch generation, multi-scale Patch calculation and spatio-temporal calculation; The prediction result output module is used to calculate the future wind power and provide a prediction confidence interval in combination with the probability model; The wind power value evaluation module is used to evaluate the long-term value of the wind farm based on the long-term time-scale PatchTST prediction results, combined with the spatio-temporal calculation of the wind energy influence relationship between wind turbines, and optimize the construction and operation strategies of the wind farm; The edge computing optimization module is used to optimize the inference efficiency of the PatchTST prediction module on the wind turbine control system or the wind farm server and improve the real-time prediction ability.

Images