CN-122022039-A - Polynomial chaotic expansion wind power plant generating capacity prediction method and system integrating active learning algorithm

CN122022039ACN 122022039 ACN122022039 ACN 122022039ACN-122022039-A

Abstract

The invention discloses a polynomial chaos expansion wind power plant generating capacity prediction method and system integrating an active learning algorithm, which are characterized in that a Latin hypercube sampling method is utilized to obtain a small amount of parameter samples of a wind power plant, wind rose data representing wind climate statistical characteristics are generated through wind direction sector division and wind speed interval combined statistics, by means of the method, a probability model of wind speed and wind direction is built, a key wind condition area with the largest contribution to prediction accuracy is automatically identified and calculated, a polynomial chaotic expansion agent model building module is formed, and an active learning module integrating exploration and utilization criteria is introduced, so that rapid and accurate prediction of wind power generation capacity of a wind power plant is achieved. The method can remarkably reduce the simulation cost, greatly improve the calculation efficiency on the basis of ensuring the prediction precision, effectively quantify the output uncertainty, and provide high-efficiency and reliable technical support for wind power plant layout optimization, operation evaluation and global sensitivity analysis.

Inventors

SHAO YIXIAO
CAO YONG
Zhan Lingyu
Thermodynamics Jade Mountain
ZHANG KAI
ZHOU DAI
WANG ZHENFAN
ZHANG RUI
CHEN YAORAN
HAN ZHAOLONG
ZHU HONGBO
TU YU

Assignees

上海交通大学

Dates

Publication Date: 20260512
Application Date: 20260128

Claims (8)

1. The polynomial chaotic expansion wind power plant generating capacity prediction method integrating the active learning algorithm is characterized by comprising the following steps of: Step S1, acquiring a small amount of wind speed-wind direction parameter samples of a wind power plant by using a Latin hypercube sampling method, carrying out wind direction sector division and wind speed interval joint statistics on the wind speed-wind direction parameter samples to generate wind rose data representing wind climate statistical characteristics, constructing a joint probability density function of wind speed-wind direction according to wind direction frequency distribution contained in the wind rose data and wind speed condition distribution information in each wind direction sector, deriving an adaptive multivariate orthogonal polynomial basis function by using a Lahm-Schmidt orthogonalization process or a corresponding recurrence relation based on the joint probability density function, and ensuring probability measure orthogonalization of polynomials about input random vectors so as to provide an optimal mathematical basis for expansion, and finally, carrying out high-precision fitting on input-output response relation of a complex physical model by using the group of orthogonal polynomial basis, and determining expansion correlation coefficients by a regression or projection method so as to establish a polynomial chaotic expansion proxy model; Step S2, introducing an active learning algorithm to guide the self-adaptive collection process of wind condition sample points, so as to define a theta criterion for fusion exploration and utilization balance, and dynamically selecting the fresh condition sample point with the maximum information value from an input random space for simulation or experiment by maximizing the theta criterion value in each iteration round based on the theta criterion, thereby updating a polynomial chaotic expansion agent model and improving the global prediction precision; Step S3, gradually decomposing an input space into a series of non-overlapping subdomains according to an extension theta criterion, independently constructing a local polynomial chaotic expansion proxy model in each subdomain to accurately capture the input-output response relation in the local area, selecting a father domain based on the extension theta criterion for carrying out refinement division in each iteration, then checking the number of existing sample points in a newly generated subdomain, if the number is insufficient, directly supplementing samples in the subdomain to support modeling, if the number is sufficient, independently constructing a local polynomial chaotic expansion proxy model in the subdomain, and determining whether to continuously participate in the next round of iteration division as a new father domain according to the subsequent precision requirement, wherein the process is continuously carried out until the global precision requirement is met or a termination condition is reached; and S4, weighting and integrating the local polynomial chaotic expansion proxy models by using an indication function to form a global segmented polynomial chaotic expansion proxy model, so that quick prediction is realized through segmented approximation, and then the operations of space decomposition, sample enhancement and model update are repeatedly executed through an iterative processing process until the model precision meets the preset requirement or reaches the maximum iterative times, after the model convergence is ensured, the obtained global segmented polynomial chaotic expansion proxy model is applied to quick prediction and uncertainty quantification of the power of the wind power plant, and high-efficiency statistical calculation of the generated energy of the wind power plant is supported.
2. A polynomial chaotic expansion wind farm generating capacity prediction system fused with an active learning algorithm, which is characterized by being used for realizing the polynomial chaotic expansion wind farm generating capacity prediction method fused with the active learning algorithm as claimed in claim 1, comprising: 1) The data input and preprocessing module is used for inputting wind rose data and wind farm parameters; 2) The polynomial chaotic expansion agent model construction module is in communication connection with the data input and preprocessing module and is used for automatically constructing an agent model in a polynomial chaotic expansion form of an input-output relation according to input parameter samples and output response data so as to realize rapid analysis and uncertainty quantification of engineering system response; 3) The active learning module is in communication connection with the polynomial chaotic expansion agent model construction module and is used for defining a fusion exploration and utilization balance theta criterion, wherein the theta criterion not only guides the dynamic selection of a sample point with the maximum information value in an input random space to carry out self-adaptive sampling, but also divides the space into non-overlapping subdomains through iterative execution sequence space decomposition so as to independently construct a local polynomial chaotic expansion agent model in each subdomain; 4) The result output and analysis module is in communication connection with the active learning module and is used for outputting annual energy generation predicted values, uncertainty quantization indexes and global sensitivity analysis results.
3. The system of claim 2, wherein the polynomial chaotic expansion proxy model construction module specifically comprises an input probability modeling unit, an orthogonal polynomial basis generation unit, an initial training sample set construction unit and a PCE coefficient solving unit, wherein the input probability modeling unit is used for fitting a joint probability density function of wind direction and wind speed according to wind rose data of a target wind power plant, the orthogonal polynomial basis generation unit is used for generating a corresponding orthogonal polynomial basis function through the input joint probability density function so as to form a polynomial chaotic expansion basis function, the initial training sample set construction unit is used for acquiring a small amount of initial sample point sets in an input variable space by using a Latin hypercube sampling method and calling a high-fidelity model to calculate total output power of the wind power plant under corresponding wind conditions so as to form an initial training data set, and the PCE coefficient solving unit is used for solving expansion coefficients of the polynomial chaotic expansion proxy model through a least square method based on the initial training data set.
4. The system of claim 3, wherein in the PCE coefficient solving unit, a polynomial chaotic expansion proxy model is expressed as: The zero-order coefficient alpha 0 is directly used for giving an expected value of output power and is used for estimating annual energy production of the wind power plant.
5. The system according to claim 2, wherein the active learning module specifically comprises an adaptive sampling criterion defining unit, a sequential spatial decomposition and sampling unit, a local PCE construction and global integration unit and an iterative convergence judging unit, the adaptive sampling criterion defining unit is used for defining a fusion exploration and utilization balance theta criterion, the theta criterion is used for evaluating the information value of candidate sample points, the sequential spatial decomposition and sampling unit is used for iteratively dividing an input space into non-overlapping subdomains, the local PCE construction and global integration unit is used for independently constructing a local polynomial chaotic expansion proxy model in each subdomain based on dedicated training data, and the iterative convergence judging unit is used for monitoring the prediction error or the change of the maximum local uncertainty of the proxy model.
6. The system of claim 5, wherein in the adaptive sampling criteria definition unit, a Θ criterion adaptively balances global exploration and local refinement according to an iterative process.
7. The system according to claim 5, wherein in the iterative convergence judging unit, the active learning cycle is terminated when the polynomial chaotic expansion agent model precision satisfies a preset threshold or reaches a maximum number of iterations.
8. The system of claim 2, wherein the system is coupled to an external training controlled high fidelity wind farm simulation software for acquiring training samples, wherein the high fidelity wind farm simulation software includes, but is not limited to, computational fluid dynamics based models or engineering wake models.

Description

Polynomial chaotic expansion wind power plant generating capacity prediction method and system integrating active learning algorithm Technical Field The invention belongs to the technical field of intersection of fluid mechanics, wind energy engineering and machine learning, and particularly relates to a polynomial chaotic expansion wind power plant generating capacity prediction method and system integrating an active learning algorithm, which are particularly suitable for a random wind power plant annual generating capacity prediction scene with complex wind conditions, and can be used for efficiently predicting the output power of a wind power plant under the complex wind conditions by using a small number of samples. Background In wind farm design and optimization research, the electricity generation amount is usually taken as a calculation period in one year, and the prediction accuracy of annual electricity generation amount (Annual Energy Production, AEP) is a key precondition and core constraint condition for ensuring efficient promotion of layout optimization and obtaining a global optimal solution. As a core index for evaluating the full life cycle power generation performance of the wind power plant, the accuracy of AEP prediction directly determines the convergence efficiency of the optimization algorithm and the credibility of the final result. When the prediction model has systematic errors or fails to fully reflect the uncertainty of wind resources, the optimization process of the prediction model is easy to fall into a local optimal solution, so that the maximum utilization of the wind resources is difficult to realize, and a series of actual engineering risks such as power generation loss, investment income reduction and the like are extremely easy to cause. In particular, in wind farms with complex terrains, high turbulence intensity or dense unit layout, the traditional AEP prediction method is excessively dependent on a simplified wake model, and adopts an idealized assumption on the joint probability distribution of wind speed and wind direction, so that it is often difficult to accurately describe the space-time variation characteristics of the flow fields and the aerodynamic mutual interference among units, and the prediction result has larger uncertainty. Therefore, developing an AEP prediction method which can ensure prediction accuracy and efficiently process input randomness has become a key requirement for pushing wind farm optimization design to project practicality. Aiming at the defects in the prior art, the invention provides a wind farm annual energy production prediction framework combining polynomial chaotic expansion and an active learning algorithm. The framework is used for enhancing the robustness of the model under the random input condition by organically combining the two methods, greatly reducing the calculation dependence on high-fidelity numerical simulation and synchronously improving the prediction precision and the calculation efficiency. Disclosure of Invention Aiming at the problems of limited prediction precision and high calculation cost caused by the fact that the existing prediction method depends on a simplified wake model and is insufficient in wind resource uncertainty, the invention provides a polynomial chaotic expansion wind power plant generating capacity prediction method and system which are integrated with an active learning algorithm. In one aspect, the invention provides a polynomial chaotic expansion wind power plant generating capacity prediction method fused with an active learning algorithm, which comprises the following steps (the main process can be referred to as fig. 1): the Latin hypercube sampling method (LHS) is adopted in the method, and has the core advantages that the Latin hypercube sampling method can conduct layered random sampling in a multidimensional parameter space (here, two dimensions of wind speed and wind direction), ensure that the whole distribution range of each input variable can be covered with equal probability and without repetition, and can better capture the integral statistical characteristics and the dependence relationship of the wind speed and the wind direction while remarkably reducing the number of required samples through systematic and layered uniform coverage of the combined distribution space of the wind speed and the wind direction. A Latin hypercube sampling method is utilized to obtain a small amount of wind speed-wind direction parameter samples with both representation and space coverage of a wind power plant, wind rose data representing wind climate statistical characteristics are generated by carrying out wind direction sector division and wind speed interval joint statistics on the wind speed-wind direction parameter samples to replace traditional statistical results needing massive data support, a wind speed-wind direction joint probability density function is constructed according to wind