CN-122021312-A - Wind farm power prediction method

CN122021312ACN 122021312 ACN122021312 ACN 122021312ACN-122021312-A

Abstract

The invention relates to the technical field of wind power prediction, in particular to a wind power plant power prediction method. The method comprises the steps of pre-training a basic prediction model based on multi-wind-field historical data, and carrying out rapid self-adaptive updating on the model by combining real-time data by utilizing a meta-learning framework and an online incremental learning algorithm. In order to improve robustness in extreme weather, a physical enhanced data synthesis engine is constructed to generate extreme event enhanced data, and an countermeasure training strategy optimization model is adopted. Meanwhile, a power curve online calibrator is constructed based on the depth twin network, the physical degradation state of the fan is identified in real time, and a physical constraint signal is generated. And finally, carrying out on-line correction on the model prediction output by using the signal, and feeding back the result to a model updating process to form a closed-loop optimization system. The invention integrates the rapid adaptability of meta learning, the distribution generalization capability of countermeasure training and the consistency constraint of the physical rule, and realizes the self-adaptive, high-robustness and high-credibility wind power plant power prediction.

Inventors

YANG XUSHUO
NING GANG
Bai Huageng
MAO WENMING
QI HONGBIN

Assignees

水发丰电能源(济南)有限公司

Dates

Publication Date: 20260512
Application Date: 20260130

Claims (10)

1. The wind farm power prediction method is characterized by comprising the following steps of: s100, pre-training based on multi-wind-field historical data to obtain a basic prediction model, and adaptively updating model parameters by utilizing real-time operation data of a target wind field; S200, constructing a physical enhanced data synthesis engine, generating enhanced training data containing extreme meteorological features, and combining an countermeasure training strategy to improve the prediction robustness of the model after self-adaptive updating in extreme weather; s300, constructing a power curve online calibrator based on a depth twin network, dynamically identifying the physical degradation state of the fan by comparing a theoretical power curve with the actual fan running state, and generating a physical constraint signal; S400, performing on-line correction on the prediction output of the self-adaptive updated model by using the physical constraint signal, and feeding back the corrected prediction result and corresponding input data to the model updating process to form a closed-loop optimization system.
2. The method according to claim 1, wherein when the basic prediction model is applied to a target wind farm, based on a meta-learning framework and an online incremental learning algorithm, the model parameters are adaptively updated by using real-time operation data of the target wind farm, specifically comprising: And in the online learning stage, combining with an elastic weight consolidation strategy, and applying constraint on important model parameters when gradient updating is carried out on small sample data flowing in real time so as to relieve catastrophic forgetting.
3. The method according to claim 1, wherein said constructing a physically enhanced data synthesis engine, in particular comprises: And inputting simulated flow field data and historical meteorological data serving as conditions into a condition generation countermeasure network to generate an enhanced training data sample which has high fidelity and contains extreme weather features.
4. The method according to claim 1, wherein the constructing the power curve online calibrator based on the depth twin network specifically comprises: The method comprises the steps of constructing a characteristic extraction network of shared parameters, respectively processing a real-time fan running state sequence and theoretical power curve data under standard working conditions, mapping the real-time fan running state sequence and the theoretical power curve data to the same characteristic space, and dynamically quantifying physical state deviation of the fan caused by efficiency attenuation and blade pollution by measuring the distance between two characteristic representations.
5. The method according to claim 2, wherein the combined challenge training strategy is in particular: And introducing an antagonism training module in the internal circulation or external circulation optimization process of the meta-learning framework, and taking the enhanced training data and the real-time abnormal data identified on line as antagonism samples for optimizing model parameters so as to improve the generalization capability of the model on the distributed external samples.
6. The method according to claim 4, wherein the generating a physical constraint signal is in particular: The physical state deviation of the depth twin network metric is converted into an adjustment quantity or adjustment distribution of the power predicted value through a tiny mapping network, and the adjustment quantity or adjustment distribution is used as a physical constraint signal.
7. The method according to claim 6, wherein the on-line correction of the predicted output is performed using a physical constraint signal, in particular: And carrying out weighted fusion on the preliminary power predicted value output by the self-adaptive updated model and the adjustment quantity indicated by the physical constraint signal, or carrying out sampling correction on the preliminary predicted value by using the adjustment distribution to obtain a final power predicted value with physical consistency.
8. A method according to claim 3, wherein in the training of the condition generating countermeasure network, a perceptual loss function is introduced, the perceptual loss being calculated based on features extracted from a pre-trained convolutional neural network capable of understanding the characteristics of the meteorological image, to ensure that the generated enhanced training data maintains a high level of consistency with the real data in the key meteorological features.
9. The method of claim 1, wherein the closed loop optimization system further comprises a model weight reduction step of taking the basic prediction model or the self-adaptive updated model as a teacher model after stabilizing, training a student model with a more simplified structure through a knowledge distillation technology, wherein the student model inherits the prediction and self-adaptation capability of the teacher model and is used for final deployment and reasoning.
10. The method of any one of claims 1-9, further comprising monitoring in real time a deviation of the physical state of the power curve output by the on-line calibrator, triggering an early warning signal when the deviation continuously exceeds a preset threshold, and recording corresponding fan identification, time and operating condition information for guiding operation and maintenance decisions.

Description

Wind farm power prediction method Technical Field The invention relates to the technical field of wind power prediction, in particular to a wind power plant power prediction method. Background With the continuous improvement of the duty ratio of wind power in an energy structure, high-precision wind power plant power prediction is crucial for safe, stable and economic operation of a power grid. Traditional power prediction methods rely mainly on numerical weather forecast (NWP) data and historical power data, modeled using time series analysis or machine learning models. In recent years, deep learning models, such as a Recurrent Neural Network (RNN), a long-short-term memory network (LSTM) and a Convolutional Neural Network (CNN), have become research hotspots in the field due to their strong timing and spatial feature extraction capabilities. However, existing wind farm power prediction methods based on deep learning still face a number of challenges that have not been thoroughly addressed. First, the generalization and adaptation ability of the model are insufficient. A model trained on specific wind farm data is often difficult to apply directly to new, geographically diverse wind farms, or to accommodate the natural decay of wind turbine performance over time, which typically requires the re-collection of large amounts of data and training of the model, which is costly and lagging. Second, the prediction of polar weather events is weak. Rare but highly-influencing events such as typhoons, extreme gusts, freezing and the like have rare samples in historical data, so that a data-driven model is extremely easy to fail in the 'out-of-distribution' scenes, prediction deviation is huge, and high risk is brought to a power grid. Moreover, most predictive models are purely data-driven "black boxes," the predicted results of which may be disjointed from the current actual physical state of the fan (e.g., reduced efficiency due to blade contamination), lacking physical consistency guarantees, and systematic errors may occur when the performance of the device is degraded. In addition, the existing method generally cuts links such as a prediction model, equipment state monitoring, data enhancement and the like, and a closed loop system capable of performing self-correction and continuous optimization by utilizing online feedback cannot be formed. Therefore, a new method for predicting the power of the wind power plant, which can adapt to different stations, robustly cope with extreme events, fuse physical laws and continuously evolve in operation, is needed. Thus, the prior art is still to be further developed. Disclosure of Invention The invention aims to overcome the technical defects and provide a wind power plant power prediction method to solve the problems in the prior art. In order to achieve the technical purpose, the invention provides a wind farm power prediction method, which comprises the following steps: s100, pre-training based on multi-wind-field historical data to obtain a basic prediction model, and adaptively updating model parameters by utilizing real-time operation data of a target wind field; S200, constructing a physical enhanced data synthesis engine, generating enhanced training data containing extreme meteorological features, and combining an countermeasure training strategy to improve the prediction robustness of the model after self-adaptive updating in extreme weather; s300, constructing a power curve online calibrator based on a depth twin network, dynamically identifying the physical degradation state of the fan by comparing a theoretical power curve with the actual fan running state, and generating a physical constraint signal; S400, performing on-line correction on the prediction output of the self-adaptive updated model by using the physical constraint signal, and feeding back the corrected prediction result and corresponding input data to the model updating process to form a closed-loop optimization system. Specifically, when the basic prediction model is applied to a target wind power plant, based on a meta-learning framework and an online incremental learning algorithm, the model parameters are adaptively updated by using real-time operation data of the target wind power plant, and the method specifically comprises the following steps: And in the online learning stage, combining with an elastic weight consolidation strategy, and applying constraint on important model parameters when gradient updating is carried out on small sample data flowing in real time so as to relieve catastrophic forgetting. Specifically, the construction of the physically enhanced data synthesis engine specifically includes: And inputting simulated flow field data and historical meteorological data serving as conditions into a condition generation countermeasure network to generate an enhanced training data sample which has high fidelity and contains extreme weather features. Specifically, the constructing