CN-122026323-A - New energy power prediction method based on heavy tail distribution modeling

CN122026323ACN 122026323 ACN122026323 ACN 122026323ACN-122026323-A

Abstract

The application provides a new energy power prediction method based on heavy tail distribution modeling, which relates to the field of power generation power prediction, and comprises the steps of flattening an acquired characteristic sequence and a tag sequence into a one-dimensional structure, and dividing a first data set and a second data set; analyzing the high-order moment characteristics and the distribution distance of the samples, calculating the relative difficulty index RDI of each sample, dynamically mapping to obtain freedom degree parameters of Student-t distribution, constructing a self-adaptive Student-t loss function, integrating the self-adaptive Student-t loss function with an energy prediction model, and reversely propagating and updating model parameters to obtain an optimized target prediction model so as to perform multi-step prediction on the generated power. According to the application, the self-adaptive Student-t loss function is integrated with the energy prediction model, and the model optimization is guided by richer information quantity through systematic utilization of error distribution characteristics, so that the capturing capacity of complex modes and atypical events is enhanced, and the accuracy of new energy power prediction is improved.

Inventors

LIU LEI
ZHAO HONGWEI
WU GUOPING
LI BIN

Assignees

中国科学技术大学

Dates

Publication Date: 20260512
Application Date: 20260126

Claims (10)

1. The new energy power prediction method based on heavy tail distribution modeling is characterized by comprising the following steps of: flattening the acquired characteristic sequence and tag sequence into a one-dimensional structure based on time sequence, and dividing the characteristic sequence into a first data set and a second data set under two continuous time periods, wherein the characteristic sequence is a historical meteorological data sequence of renewable energy, and the tag sequence is a historical power generation sequence corresponding to the characteristic sequence; Analyzing the high-order moment characteristics and the distribution distance of samples between the first data set and the second data set, and calculating a relative difficulty index RDI of each sample for representing error distribution characteristics; Dynamically mapping based on a relative difficulty index RDI to obtain a freedom degree parameter of Student-t distribution so as to represent the degree of heavy tail error, and constructing a self-adaptive Student-t loss function based on the freedom degree parameter so as to model heavy tail error distribution; Integrating the self-adaptive Student-t loss function with an energy prediction model on the premise of not changing the forward structure of the preset energy prediction model, and driving a constructed heavy tail mechanism, wherein the heavy tail mechanism is used for processing heavy tail error distribution and comprises the steps of determining the loss value of the self-adaptive Student-t loss function as an optimization target and replacing MSE loss of the model, and self-adaptively adjusting the degree of freedom parameter of the Student-t loss through a relative difficulty index; and based on the loss value back propagation updating model parameters of the self-adaptive Student-t loss function, obtaining an optimized target prediction model to perform multi-step prediction on the generated power of the renewable energy.
2. The new energy power prediction method based on heavy tail distribution modeling according to claim 1, wherein the adaptive Student-t loss function is determined based on a degree of freedom parameter and a prediction error, and the adaptive Student-t loss function is in local Lipoz continuity with respect to the prediction error; in driving the heavy tail mechanism, step size is used The gradient descent of (a) optimizes the adaptive Student-t loss to converge the algorithm to a plateau, eta is the step size, and L is the local liplitz continuous constant of the adaptive Student-t loss function with respect to the prediction error.
3. The new energy power prediction method based on heavy tail distribution modeling according to claim 2, wherein the adaptive Student-t loss function is configured as a plug-and-play loss plug-in for call, and wherein the adaptive Student-t loss function is configured as a plug-and-play loss plug-in for call The expression is satisfied: ; The local liphatz continuous constant satisfies the expression: ; Wherein, the In order to predict the error of the signal, Is a degree of freedom parameter, prediction error Is bounded and is present So that M is positive number, and the degree of freedom parameter Bounded and of , wherein, , And Respectively is Lower and upper bounds of (2); When (when) At this time, the Student-t loss converges consistently to the MSE loss to achieve a smooth transition from the heavy tail mechanism to the Gaussian mechanism.
4. The new energy power prediction method based on heavy tail distribution modeling according to claim 3, wherein the adaptive mapping between the relative difficulty index RDI and the freedom degree parameters of the Student-t distribution is a negative correlation mapping; The self-adaptive mapping is realized by adopting a Sigmoid function, the input of the Sigmoid function is a relative difficulty index RDI, the output of the Sigmoid function is a numerical value of the normalized degree of freedom parameter, and the process of the self-adaptive mapping meets the expression: ; Wherein v (RDI) represents a degree of freedom parameter mapped and output based on the relative difficulty index RDI, g # ) Represents the sigmoid function and supports end-to-end training of the model by functional scalability, Positive and used to control the transition sensitivity.
5. The new energy power prediction method based on heavy tail distribution modeling according to claim 1, wherein the analyzing the high order moment characteristics and the distribution distance of the samples between the first data set and the second data set, calculating the relative difficulty index RDI of each sample for characterizing the error distribution characteristics, includes: analyzing a normalized wasperstein distance between the first data set and the second data set to capture macroscopic pattern differences between data sequences; analyzing the skewness difference and kurtosis difference of the first data set and the second data set to quantify the non-Gaussian characteristic change of the data sequence, wherein the skewness difference is used for representing the asymmetric mode change, and the kurtosis difference is used for representing the tail behavior change; And fusing the skewness difference and the kurtosis difference through a modulation function, and carrying out weighting treatment by combining the normalized Wasserstein distance to obtain a relative difficulty index RDI.
6. The new energy power prediction method based on heavy tail distribution modeling according to claim 5, wherein the modulation function is The expression is satisfied: Wherein, the And Is a weight parameter that can be learned and is, The difference in the degree of deviation is indicated, Representing kurtosis differences; ; ; Wherein, the Representing the skewness of the first data set X, Representing the skewness of the second data set Y, Representing the kurtosis of the first data set X, Representing kurtosis of the second dataset Y; , ; where Z represents a sample in the first data set or the second data set, Representing the mean of the samples in the first data set or the second data set, Representing the standard deviation of each sample in the first data set or the second data set, E represents the mathematical expectation.
7. The new energy power prediction method based on heavy tail distribution modeling according to any one of claims 1-6, wherein the renewable energy source comprises wind energy or solar energy, the historical meteorological data sequence is derived from a supervisory control and data acquisition SCADA system and meteorological monitoring equipment and comprises barometric pressure, relative humidity, temperature, dew point, wind direction at target level, and wind speed at target level; the energy prediction model is a transform variant model and comprises one of PatchTST model, iTransformer model and TimeXer model, and is trained by adopting an Adam optimizer and combining an early-stop strategy and a weight attenuation mechanism to perform end-to-end optimization.
8. The utility model provides a new forms of energy power prediction system based on heavy tail distribution modeling which characterized in that includes: The data flattening and dividing module is used for flattening the acquired characteristic sequence and the tag sequence into a one-dimensional structure based on time sequence and dividing the one-dimensional structure into a first data set and a second data set under two continuous time periods, wherein the characteristic sequence is a historical meteorological data sequence of renewable energy, and the tag sequence is a historical power generation sequence corresponding to the characteristic sequence; The relative difficulty index calculation module is used for analyzing high-order moment characteristics and distribution distances of samples between the first data set and the second data set and calculating a relative difficulty index RDI of each sample for representing error distribution characteristics; The loss function construction module is used for dynamically mapping to obtain freedom degree parameters of the Student-t distribution based on the relative difficulty index RDI so as to represent the degree of heavy tail error, and constructing a self-adaptive Student-t loss function based on the freedom degree parameters so as to model the heavy tail error distribution; The integrated driving module is used for integrating the self-adaptive Student-t loss function with the energy prediction model on the premise of not changing the forward structure of the preset energy prediction model and driving a constructed heavy tail mechanism, wherein the heavy tail mechanism is used for processing heavy tail error distribution and comprises the steps of determining the loss value of the self-adaptive Student-t loss function as an optimization target and replacing MSE loss of the model, and self-adaptively adjusting the freedom degree parameter of the Student-t loss through a relative difficulty index; And the training module is used for reversely propagating and updating model parameters based on the loss value of the self-adaptive Student-t loss function to obtain an optimized target prediction model so as to carry out multi-step prediction on the generated power of the renewable energy.
9. An electronic device comprising a processor, a memory, and a program stored on the memory and executable on the processor, the program when executed by the processor implementing the new energy power prediction method based on heavy tail distribution modeling of any one of claims 1 to 7.
10. A computer readable storage medium, wherein a program or instructions is stored on the computer readable storage medium, which when executed by a processor, implements the new energy power prediction method based on heavy tail distribution modeling according to any one of claims 1 to 7.

Description

New energy power prediction method based on heavy tail distribution modeling Technical Field The application relates to the technical field of power generation power prediction, in particular to a new energy power prediction method based on heavy tail distribution modeling. Background Accurate renewable energy prediction is critical to the operation of modern power systems, which directly affects grid stability, economic dispatch efficiency, and renewable energy grid-tie rate. Renewable energy sources such as wind energy and solar energy are influenced by meteorological factors such as wind speed fluctuation, solar irradiance change and the like, and have inherent variability and uncertainty, so that the power generation output presents intermittent and nonlinear characteristics, and the method brings fundamental challenges for reliable power generation prediction. Especially, extreme events such as sudden power rise/fall caused by extreme weather are liable to cause great hidden trouble to the safety of the power grid, so that high-precision renewable energy prediction has become a core requirement for guaranteeing the safe and economic operation of the power system. In recent years, a deep learning technique based on a transducer variant represented by PatchTST, iTransformer, timeXer has been widely used and has been significantly advanced in the field of renewable energy short-term and ultra-short-term power prediction. A typical implementation flow of the method is that a multivariate history sequence is obtained from a data acquisition and monitoring control (Supervisory Control And Data Acquisition, SCADA) system and a meteorological data source, and after time alignment, missing value processing and standardized preprocessing, a model is constructed through a sliding window to input data and labels. At the model architecture level, patchTST captures a long period mode through patch word segmentation and time dimension self-attention, iTransformer inverts attention modeling cross-variable dependency relationship in variable dimension, timeXer combines multi-scale decomposition and exogenous variable to enhance the feature extraction capability, and finally outputs multi-step prediction results through a decoding layer. The model training stage is usually based on Gaussian error assumption, takes mean square error MSE as a main loss function, can be assisted by mean absolute error MAE, quantile loss and the like, is matched with an Adam optimizer, an early stop strategy and a weight attenuation technology to realize end-to-end training optimization, adopts indexes such as MSE, MAE, normalized mean absolute error NMAE, normalized mean square error NMSE and the like to evaluate model performance, and rolls and outputs a prediction result according to a fixed prediction length at an electric field side or a cloud side during deployment. Although the prediction method based on deep learning is continuously optimized in the design of a model architecture, a plurality of defects still exist in practical application, and further improvement of prediction precision and robustness is limited. Firstly, the assumption that the mean square error MSE loss function implicit error commonly adopted by the existing method obeys Gaussian distribution is seriously inconsistent with the heavy tail distribution characteristic actually presented by the renewable energy prediction error, so that the prediction performance of the model is obviously reduced under key scenes such as extreme weather, high variability and the like. Secondly, the existing method adopts a unified loss function to process all samples, adopts a 'one-cut' optimization strategy, fails to distinguish prediction difficulty differences among different samples, cannot adapt to heterogeneous characteristics existing in renewable energy data, and is difficult to cope with complex scenes with distinct prediction difficulty under stable weather and extreme events. Based on the above, the current situation that the error distribution characteristics lack systematic analysis and utilization in the prior art becomes a key bottleneck for restricting the model performance breakthrough. Disclosure of Invention Aiming at the defects of the prior art, the application provides a new energy power prediction method based on heavy tail distribution modeling, which solves the problems that the existing renewable energy power prediction method based on deep learning is insufficient in prediction precision and robustness under critical scenes such as extreme weather and the like due to mismatching of loss function assumptions and actual error distribution and incapability of adapting to data heterogeneity by adopting a unified optimization strategy. In order to achieve the above purpose, the application is realized by the following technical scheme: According to the embodiment of the application, a new energy power prediction method based on heavy tail distribution modeling