New energy power prediction method based on heavy tail distribution modeling

Abstract

The invention provides a new energy power prediction method based on heavy-tailed distribution modeling, and relates to the field of power generation power prediction, and the method comprises the steps: flattening an obtained feature sequence and a tag sequence into a one-dimensional structure, and dividing the one-dimensional structure into a first data set and a second data set; analyzing high-order moment characteristics and distribution distances of the samples, calculating a relative difficulty index (RDI) of each sample, dynamically mapping to obtain a degree of freedom parameter of Student-t distribution, and constructing a self-adaptive Student-t loss function; and integrating the adaptive Student-t loss function with the energy prediction model, and performing back propagation to update model parameters to obtain an optimized target prediction model so as to perform multi-step prediction on the generated power. According to the method, the self-adaptive Student-t loss function and the energy prediction model are integrated, and more information amount guidance is provided for model optimization through systematic utilization of error distribution characteristics, so that the capacity of capturing complex modes and atypical events is enhanced, and the precision of new energy power prediction is improved.

Description

Technical Field

[0001] This application relates to the field of power generation prediction technology, specifically to a new energy power prediction method based on heavy-tailed distribution modeling. Background Technology

[0002] Accurate renewable energy forecasting is crucial for the operation of modern power systems, directly impacting grid stability, economic dispatch efficiency, and renewable energy grid integration rates. Renewable energy sources such as wind and solar power are inherently variable and uncertain due to meteorological factors like wind speed fluctuations and changes in solar irradiance, resulting in intermittent and nonlinear power generation. This poses a fundamental challenge to reliable power generation forecasting. In particular, extreme events such as sudden power spikes / drops caused by extreme weather can easily pose significant threats to grid security. Therefore, high-precision renewable energy forecasting has become a core requirement for ensuring the safe and economical operation of power systems.

[0003] In recent years, deep learning techniques based on Transformer variants, such as PatchTST, iTransformer, and TimeXer, have been widely applied and have achieved significant progress in the field of short-term and ultra-short-term power forecasting for renewable energy. A typical implementation flow for these methods is as follows: Multivariate historical sequences are acquired from Supervisory Control and Data Acquisition (SCADA) systems and meteorological data sources. After time alignment, missing value handling, and standardization preprocessing, the model input data and labels are constructed using a sliding window. At the model architecture level, PatchTST captures long-cycle patterns through patch segmentation and time-dimensional self-attention; iTransformer models cross-variable dependencies using variable-dimensional inverted attention; and TimeXer combines multi-scale decomposition and exogenous variable enhancement to improve feature extraction capabilities. Finally, multi-step prediction results are output through a decoding layer. The model training phase is usually based on the Gaussian error assumption, with mean squared error (MSE) as the main loss function. It can be supplemented by mean absolute error (MAE), quantile loss, etc., and combined with Adam optimizer, early stopping strategy and weight decay technique to achieve end-to-end training optimization. The model performance is evaluated by indicators such as MSE, MAE, standardized mean absolute error (NMAE) and standardized mean squared error (NMSE). During deployment, the prediction results are output in a rolling manner at a fixed prediction length on the electric field side or in the cloud.

[0004] Despite continuous optimization in model architecture design, the aforementioned deep learning-based prediction methods still suffer from numerous shortcomings in practical applications, limiting further improvements in prediction accuracy and robustness. First, the assumption that the mean squared error (MSE) loss function, commonly used in existing methods, implies a Gaussian distribution of latent error is severely inconsistent with the heavy-tailed distribution characteristics of renewable energy prediction errors. This leads to a significant drop in prediction performance under critical scenarios such as extreme weather and high variability. Second, existing methods use a uniform loss function for all samples, employing a "one-size-fits-all" optimization strategy. This fails to differentiate the varying prediction difficulty among different samples, making it unsuitable for the heterogeneous characteristics of renewable energy data and ill-suited to complex scenarios where prediction difficulty differs drastically between stable weather and extreme events. Therefore, the lack of systematic analysis and utilization of error distribution characteristics in current technologies has become a key bottleneck restricting breakthroughs in model performance. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this application provides a new energy power prediction method based on heavy-tailed distribution modeling. This method solves the problems of insufficient prediction accuracy and robustness in key scenarios such as extreme weather caused by the mismatch between the loss function assumptions and the actual error distribution, the inability of the uniform optimization strategy to adapt to data heterogeneity, and the lack of accuracy and robustness in existing deep learning-based renewable energy power prediction methods.

[0006] To achieve the above objectives, this application provides the following technical solution: In a first aspect, embodiments of this application provide a new energy power prediction method based on heavy-tailed distribution modeling. This method includes: flattening the acquired feature sequence and label sequence into a one-dimensional structure based on time series, and dividing it into a first dataset and a second dataset under two consecutive time periods; the feature sequence is a historical meteorological data sequence of renewable energy, and the label sequence is a historical power generation sequence corresponding to the feature sequence; analyzing the higher-order moment features and distribution distance of samples between the first dataset and the second dataset, and calculating the relative difficulty index (RDI) representing the error distribution characteristics of each sample; dynamically mapping the degree of freedom parameters of the Student-t distribution based on the relative difficulty index (RDI) to characterize the degree of heavy-tailed error, and constructing a new energy power prediction method based on the degree of freedom parameters. An adaptive Student-t loss function is constructed to model the heavy-tailed error distribution. Without changing the forward structure of the preset energy prediction model, the adaptive Student-t loss function is integrated with the energy prediction model to drive the constructed heavy-tailed mechanism. The heavy-tailed mechanism is used to handle the heavy-tailed error distribution and includes: determining the loss value of the adaptive Student-t loss function as the optimization objective and replacing the model's MSE loss; adaptively adjusting the degree of freedom parameter of the Student-t loss through the relative difficulty index; updating the model parameters based on the loss value of the adaptive Student-t loss function through backpropagation to obtain the optimized target prediction model for multi-step prediction of renewable energy power generation.

[0007] Secondly, embodiments of this application provide a new energy power prediction system based on heavy-tailed distribution modeling. The system includes: a data flattening and partitioning module, a relative difficulty index calculation module, a loss function construction module, an integration driving module, and a training module.

[0008] Specifically, the data flattening and partitioning module flattens the acquired feature and label sequences into a one-dimensional structure based on time series, dividing them into a first dataset and a second dataset for two consecutive time periods; the feature sequences are historical meteorological data sequences of renewable energy, and the label sequences are historical power generation sequences corresponding to the feature sequences. The relative difficulty index calculation module analyzes the higher-order moment features and distribution distance of samples between the first and second datasets, calculating the relative difficulty index (RDI) for each sample to characterize the error distribution. The loss function construction module dynamically maps the degrees of freedom parameters of the Student-t distribution based on the relative difficulty index (RDI) to characterize the degree of heavy-tailed error, and constructs an adaptive Student-t loss function based on the degrees of freedom parameters to model the heavy-tailed error distribution. The integration-driven module integrates the adaptive Student-t loss function with the energy prediction model without changing the forward structure of the preset energy prediction model, driving the constructed heavy-tailed mechanism; the heavy-tailed mechanism handles the heavy-tailed error distribution and includes: determining the loss value of the adaptive Student-t loss function as the optimization objective and replacing the model's MSE loss, and adaptively adjusting the degrees of freedom parameters of the Student-t loss through the relative difficulty index. The training module is used to update the model parameters through backpropagation based on the loss value of the adaptive Student-t loss function, so as to obtain the optimized target prediction model for multi-step prediction of renewable energy power generation.

[0009] Thirdly, embodiments of this application provide an electronic device, which includes: a processor, a memory, and a program stored in the memory and executable on the processor. When the program is executed by the processor, it implements the new energy power prediction method based on heavy-tailed distribution modeling described in the first aspect.

[0010] Fourthly, embodiments of this application provide a computer-readable storage medium storing a program or instructions that, when executed by a processor, implement the new energy power prediction method based on heavy-tailed distribution modeling described in the first aspect.

[0011] This application provides a new energy power prediction method based on heavy-tailed distribution modeling. Compared with existing technologies, it has the following advantages: This application addresses the heavy-tailed error characteristics in the renewable energy power prediction process by proposing an adaptive Student-t loss function to model the heavy-tailed error distribution. It abandons the Gaussian distribution assumption relied upon by traditional MSE (Mean Sequencing Engineering) and introduces a relative difficulty index (RDI) to finely quantify the prediction difficulty of each sample in the adaptive Student-t loss function process. The shape of the loss function can be dynamically adjusted according to the difficulty of the sample itself, providing a differentiated strategy to better adapt the model to the inherent heterogeneity in renewable energy data. To accurately characterize the RDI, this application flattens the feature sequence and label sequence into a one-dimensional structure based on time series, dividing the historical periods corresponding to the feature sequence and label sequence into two consecutive periods, corresponding to early history and recent history, respectively. The distribution difference between these two consecutive periods can indirectly characterize the shift trend between historical and future patterns. Based on this shift trend, the deep statistical characteristics of the error distribution are considered to determine the RDI, and the degrees of freedom parameters of the Student-t distribution are determined through mapping to characterize the degree of heavy-tailed error. The model not only fits the time series pattern of power generation but also learns to adjust the loss sensitivity according to the sample difficulty. This application integrates the adaptive Student-t loss function with the energy prediction model. By systematically utilizing the characteristics of the error distribution, it provides more informative guidance for model optimization, thereby enhancing the ability to capture complex patterns and atypical events and improving the accuracy of new energy power prediction. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating a new energy power prediction method based on heavy-tailed distribution modeling provided in an embodiment of this application. Figure 2 This is an exemplary architecture diagram corresponding to a new energy power prediction method based on heavy-tailed distribution modeling provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a new energy power prediction system based on heavy-tailed distribution modeling provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0015] This application provides a new energy power prediction method based on heavy-tailed distribution modeling, which solves the problems of existing deep learning-based renewable energy power prediction methods, such as the mismatch between loss function assumptions and actual error distribution, the inability of the unified optimization strategy to adapt to data heterogeneity, and insufficient prediction accuracy and robustness in key scenarios such as extreme weather, thereby achieving accurate prediction of new energy power.

[0016] The technical solution in this application is to solve the above-mentioned technical problems, and the general idea is as follows: Accurate renewable energy forecasting is crucial for the operation of modern power systems, directly impacting grid stability, economic dispatch efficiency, and renewable energy grid integration rates. However, renewable energy sources (such as wind and solar power) inherently possess variability and uncertainty; meteorological factors such as wind speed fluctuations and changes in solar irradiance can cause power generation output to exhibit intermittent and nonlinear characteristics. This poses a fundamental challenge to reliable power generation forecasting, making high-precision forecasting a key requirement for ensuring the safe and economical operation of the power system.

[0017] In recent years, deep learning technology has made significant progress in the field of renewable energy forecasting. These methods have improved forecasting accuracy through architectural innovation, thus ensuring the safe and stable operation of the power system to a certain extent. However, the research and application of renewable energy forecasting currently face several challenges: (1) The “non-Gaussian” and “heavy-tailed” characteristics of prediction error: Traditional deep learning methods generally use mean squared error (MSE) as the loss function, and its core assumption is that the prediction error follows a Gaussian distribution (i.e., a bell curve). However, renewable energy (such as wind power and solar power) is affected by extreme weather and intermittent factors, and its prediction error often exhibits a “heavy-tailed distribution” in practice. The larger the tail thickness, the higher the probability of extreme events occurring. This means that the probability of extremely large prediction errors is much higher than that expected by a Gaussian distribution.

[0018] (2) The “heterogeneity” and “time-varying” nature of prediction difficulty: The prediction difficulty of renewable energy is not constant. When the weather is stable, prediction is relatively easy; but when extreme events such as typhoons or rapid cloud movement occur, prediction becomes extremely difficult. This huge difference in prediction difficulty between samples is called “heterogeneity”.

[0019] (3) Vulnerability of the model to "extreme events" and "out-of-distribution" samples: The grid operators are most concerned about extreme events (such as power surges / decreases) because they are the main risks leading to grid instability. However, these events account for a small proportion of the data, and traditional models often perform poorly at these critical points. The model performs well under normal conditions, but fails at critical moments (extreme weather), leading to huge economic and safety risks.

[0020] In recent years, a renewable energy power forecasting method based on Transformer variants (represented by PatchTST, iTransformer, and TimeXer) has been widely used for short-term and ultra-short-term forecasting of wind power / solar power. The system obtains multivariate historical sequences from SCADA and meteorological sources, and after time alignment, missing value processing, and standardization, it constructs input X and label Y using a sliding window. On the model side, PatchTST captures long-period patterns through patch tokenization and time-dimensional self-attention, iTransformer models cross-variable dependencies with variable-dimensional inverted attention, and TimeXer combines multi-scale decomposition and exogenous variable enhancement. Finally, the multi-step prediction is output through the decoding layer. During the training phase, the mean squared error (MSE) is usually used as the main loss under the Gaussian error assumption (MAE / quantile loss can be supplemented), and end-to-end optimization is performed with Adam / early stop / weight decay, and the results are evaluated using indicators such as MSE / MAE / NMAE / NMSE. During deployment, the results are output in a rolling manner at the electric field side or in the cloud according to a fixed prediction length.

[0021] Existing technologies for power forecasting for renewable energy sources have the following main drawbacks: (1) Although deep learning-based renewable energy prediction methods have made significant progress in model architecture, these methods generally use mean squared error (MSE) as the loss function. The implicit Gaussian distribution assumption is seriously inconsistent with the heavy-tailed characteristics of the actual renewable energy prediction error, which leads to a decline in the prediction performance of the model in extreme weather and high variability periods.

[0022] (2) Most methods use a uniform loss function to process all samples, failing to distinguish the differences in prediction difficulty between different samples. This "one-size-fits-all" optimization strategy ignores the heterogeneity in renewable energy data: the prediction difficulty is drastically different between stable periods and extreme event periods, making it difficult for the model to adapt to complex and ever-changing real-world application scenarios.

[0023] (3) Most models fail to fully utilize the distribution characteristics of prediction errors. However, the statistical characteristics of error distribution, such as tail characteristics and concentration, are actually of great value in guiding model optimization and improving prediction robustness. Existing methods lack systematic analysis and utilization of error distribution characteristics, which limits further improvement in model performance.

[0024] Therefore, there is an urgent need to propose a new energy power prediction method based on heavy-tailed distribution modeling to address the technical deficiencies in existing technologies for power prediction of renewable energy.

[0025] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0026] The following section first introduces a new energy power prediction method based on heavy-tailed distribution modeling provided in the embodiments of this application.

[0027] This application provides a flowchart illustrating a new energy power prediction method based on heavy-tailed distribution modeling, as shown in the embodiments below. Figure 1 As shown, the new energy power prediction method based on heavy-tailed distribution modeling may include the following steps S110-S150.

[0028] S110. Based on the time series, the acquired feature sequence and label sequence are flattened into a one-dimensional structure and divided into the first dataset and the second dataset under two consecutive time periods; the feature sequence is the historical meteorological data sequence of renewable energy, and the label sequence is the historical power generation sequence corresponding to the feature sequence.

[0029] Understandably, this application flattens the feature sequence and label sequence into a one-dimensional structure based on time series, and merges and flattens the feature sequence samples and label sequence samples at the same time node into data segments corresponding to the same time node, so as to comprehensively consider the data of the two sequences and perform subsequent analysis.

[0030] It should be noted that the feature sequences and label sequences used in this application are historical meteorological data sequences and historical power generation sequences that have undergone time alignment, missing value processing, and standardization preprocessing. The feature sequences and label sequences correspond to each other, therefore the historical time periods they correspond to are the same; the flattened one-dimensional structure corresponds to the same historical time period as the feature sequences and label sequences. This application performs time-series partitioning on the one-dimensional structure to obtain a first dataset and a second dataset. The first dataset and the second dataset correspond to two consecutive time periods. When performing time-series partitioning on the one-dimensional structure, the time periods corresponding to the two consecutive time periods can be equal or different.

[0031] It should be emphasized that the one-dimensional structure of history in this application is used to divide the historical period corresponding to the data into early history and recent history. The difference in data distribution within these two consecutive periods can indirectly characterize the trend of the shift between historical and future patterns. This application transforms the "difficulty of predicting the future using history" into "the distribution difference between two consecutive periods within history," indirectly characterizing the difficulty of future prediction through the dynamic changes in history.

[0032] S120. Analyze the higher-order moment features and distribution distance of samples between the first and second datasets, and calculate the relative difficulty index (RDI) for each sample to characterize the error distribution.

[0033] In one example, higher-order moment features may include skewness difference and kurtosis difference, with the distribution distance being the normalized Wasserstein distance.

[0034] Understandably, higher-order moment features can quantify the non-Gaussian differences between the first and second datasets, skewness differences reflect changes in distribution asymmetry, and kurtosis differences reflect changes in the density of extreme values. Distribution distance can quantify the degree of pattern shift between the first and second datasets; the greater the distribution distance, the more significant the difference between early and recent historical meteorological / power generation patterns.

[0035] It should be noted that the essence of the Relative Difficulty Index (RDI) is to transform pattern shift and non-Gaussian features into quantifiable sample prediction difficulty scores by fusing higher-order moment features and distribution distance. The RDI designed in this application considers both the change in weather / power generation patterns through distribution distance and analyzes non-Gaussianity through higher-order moment features to determine whether changes in weather / power generation patterns make prediction difficult, thus avoiding the limitations of a single dimension. Based on this, this application can output a sample-level RDI, providing an accurate difficulty label for each sample and solving the problem of indiscriminate sample treatment in traditional loss functions.

[0036] S130. The degree-of-freedom parameters of the Student-t distribution are obtained by dynamic mapping based on the relative difficulty index RDI to characterize the degree of heavy-tailed error. An adaptive Student-t loss function is constructed based on the degree-of-freedom parameters to model the heavy-tailed error distribution.

[0037] Understandably, this application fundamentally abandons the Gaussian distribution assumption relied upon by traditional MSE, instead employing the Student-t distribution, which can flexibly characterize tail behavior, to model prediction errors. This shift allows the loss function to more accurately match the heavy-tailed characteristics actually exhibited in renewable energy prediction errors, thereby significantly improving the model's prediction robustness and performance during extreme weather and highly variable periods.

[0038] It should be noted that the relative difficulty index (RDI) characterizes the prediction difficulty of the sample. The degree of freedom parameter is obtained by mapping based on the relative difficulty index (RDI). The smaller the degree of freedom parameter, the stronger the heavy-tailed distribution and the higher the probability of extreme values. Through the dynamic mapping between the relative difficulty index (RDI) and the degree of freedom parameter, the constructed adaptive Student-t loss function can be adapted to the sample difficulty.

[0039] The adaptive Student-t loss function in this application is a scenario-specific tool. It employs a robust loss for difficult samples corresponding to heavy-tailed errors, while using an efficient loss close to the Mean Squared Error (MSE) for simple samples corresponding to Gaussian errors. Based on this, the dynamic degrees of freedom parameter allows the loss function to have sample-level adaptive capabilities, eliminating the need for manual adjustment of loss parameters and adapting to the heterogeneity of different datasets. This application can overcome the Gaussian error assumption of traditional MSE, accurately model the heavy-tailed error distribution of renewable energy prediction, and improve the prediction robustness for extreme events.

[0040] S140. Without changing the forward structure of the preset energy prediction model, the adaptive Student-t loss function is integrated with the energy prediction model to drive the constructed heavy-tail mechanism. The heavy-tail mechanism is used to handle the heavy-tail error distribution and includes: determining the loss value of the adaptive Student-t loss function as the optimization objective and replacing the model's MSE loss, and adaptively adjusting the degree of freedom parameter of the Student-t loss through the relative difficulty index.

[0041] Understandably, this application does not require changing the original forward structure of the energy prediction model or reconstructing the model architecture when introducing the improved adaptive Student-t loss function. During model training, the predicted power output of the forward structure based on the data in the feature sequence is compared with the power data at the corresponding time node provided by the label sequence to calculate the loss. The loss calculation process is handled by the adaptive Student-t loss function integrated with the energy prediction model. The adaptive Student-t loss function is adjusted according to the value of the degree of freedom parameter to provide appropriate loss calculations for samples with different prediction difficulties.

[0042] S150. The model parameters are updated by backpropagation of the loss value based on the adaptive Student-t loss function to obtain the optimized target prediction model for multi-step prediction of renewable energy power generation.

[0043] Understandably, during a training round, after forward propagation and loss calculation, this application uses the loss value of the adaptive Student-t loss function as a signal to backpropagate and calculate the gradients corresponding to all trainable parameters. End-to-end training allows for collaborative parameter optimization; the model not only fits the temporal patterns of power generation but also learns to adjust loss sensitivity based on sample difficulty, improving prediction accuracy (especially for extreme events). The optimized target prediction model exhibits strong generalization ability, having learned to adapt to unseen dynamic scenarios, such as novel weather patterns and extreme weather combinations.

[0044] The above is a specific implementation method of a new energy power prediction method based on heavy-tailed distribution modeling provided in the embodiments of this application. Please refer to it as well. Figure 1 and Figure 2 This application integrates the adaptive Student-t loss function with the energy prediction model. By systematically utilizing the error distribution characteristics, it provides more informative guidance for model optimization, thereby enhancing the ability to capture complex patterns and atypical events and improving the accuracy of new energy power prediction.

[0045] In some embodiments, the analysis of the higher-order moment features and distribution distance of samples between the first and second datasets, and the calculation of the relative difficulty index (RDI) for representing the error distribution characteristics of each sample, i.e., the aforementioned S120 may specifically include the following steps: S210. Analyze the normalized Wasserstein distance between the first and second datasets to capture macroscopic pattern differences between data sequences; S220. Analyze the skewness and kurtosis differences between the first and second datasets to quantify the non-Gaussian variation of the data sequences; wherein, the skewness difference is used to characterize asymmetric pattern changes, and the kurtosis difference is used to characterize tail behavior changes. S230. The skewness difference and kurtosis difference are fused by the modulation function and weighted by the normalized Wasserstein distance to obtain the relative difficulty index RDI.

[0046] In the embodiments of this application, it is understood that, on the one hand, this application uses normalized Wasserstein distance to accurately capture the macroscopic pattern differences between two datasets, providing a basic quantitative basis for prediction difficulty and ensuring sensitive capture of overall changes in the data sequence. On the other hand, this application quantifies asymmetric pattern changes and tail behavior changes through skewness difference and kurtosis difference, respectively, accurately characterizing the non-Gaussian nature of the data and making up for the inadequacy of a single dimension in describing complex error distributions.

[0047] Furthermore, this application utilizes a modulation function to fuse skewness and kurtosis differences, and combines this with normalized Wasserstein distance for weighting. This achieves both the organic integration of microscopic non-Gaussian characteristics and macroscopic pattern differences, and allows the contributions of different dimensions to adapt to data characteristics through learnable weights. The resulting Relative Difficulty Index (RDI) comprehensively and accurately characterizes the error distribution characteristics and prediction difficulty of each sample, providing a reliable basis for subsequent adjustment of the degrees of freedom parameters of the adaptive Student-t loss function. This helps the model achieve sample-level differentiated optimization and improves the predictive robustness for extreme events and high-variability scenarios. This application proactively integrates the deep statistical characteristics of the error distribution into the design of the loss function through the Relative Difficulty Index (RDI). This systematic utilization of error distribution characteristics provides more informative guidance for model optimization, thereby enhancing the ability to capture complex patterns and atypical events (such as extreme weather).

[0048] In one example, the modulation function Satisfying the expression: in, and These are learnable weight parameters. Indicates skewness difference. Indicates kurtosis variation; ; ; in, Represents the first dataset X skewness, Represents the second dataset Y skewness, Represents the first dataset X peak, Represents the second dataset Y Peak value; , ; In the formula, Z represents a sample in the first dataset or the second dataset. This represents the mean of all samples in either the first or second dataset. represents the standard deviation of each sample in the first or second dataset, and E represents the expected value.

[0049] In the embodiments of this application, it can be understood that this application is achieved through... and Two learnable parameters adaptively adjust the contribution weights of skewness and kurtosis differences to the modulation result, allowing the model to automatically optimize weight allocation based on the characteristics of different datasets, avoiding the subjectivity and limitations of manual parameter tuning. Furthermore, both skewness and kurtosis differences are calculated using absolute values, ensuring the non-negativity of the difference values. The linear modulation function is continuously differentiable and can directly participate in the end-to-end training of the model, seamlessly integrating with the backpropagation process without adding extra computational complexity. This lays a reliable quantitative foundation for the subsequent dynamic mapping of the relative difficulty index (RDI) to the degree-of-freedom parameters.

[0050] In one example, renewable energy sources include wind or solar power. Historical meteorological data sequences are derived from SCADA (Supervisory and Data Acquisition) systems and meteorological monitoring equipment, and include air pressure, relative humidity, temperature, dew point, wind direction at the target altitude, and wind speed at the target altitude. Understandably, characteristics such as air pressure, relative humidity, temperature, dew point, wind direction at the target altitude, and wind speed are highly correlated with wind and solar power generation capacity, providing ample physical evidence for the model and improving the reliability of predictions from the outset.

[0051] In another example, the energy prediction model is a Transformer variant model, including one of the PatchTST model, iTransformer model, and TimeXer model; the energy prediction model is trained using the Adam optimizer, combined with an early stopping strategy and a weight decay mechanism for end-to-end optimization.

[0052] Understandably, this application selects Transformer variant models such as PatchTST, iTransformer, and TimeXer to fully leverage the advantages of the Transformer architecture in modeling long-term time-series dependencies. These models can effectively capture the complex temporal correlation between meteorological data and power generation, and are adapted to the core requirements of multi-step forecasting of renewable energy. At the same time, these variant models have been specifically optimized for time series tasks (such as block embedding in PatchTST and reverse attention in iTransformer). Compared with traditional time series models (such as LSTM), they have higher accuracy and smaller error accumulation in scenarios with long forecast durations.

[0053] In some embodiments, the adaptive Student-t loss function is determined based on the degrees of freedom parameters and the prediction error; the adaptive Student-t loss function exhibits local Lipshitz continuity with respect to the prediction error. A step size is used in driving the heavy-tail mechanism. Gradient descent is used to optimize the adaptive Student-t loss so that the algorithm converges to a stationary point, where η is the step size and L is the local Lipshitz continuity constant of the adaptive Student-t loss function with respect to the prediction error.

[0054] It should be noted that the adaptive adjustment of the degree-of-freedom parameters based on the relative difficulty index (RDI) preserves the convergence guarantee. The step size constraint is to ensure the stable convergence of gradient descent. The 'L' in the step size constraint quantifies the maximum possible rate of change of the loss function gradient. 'L' provides a safe upper limit for the gradient descent step size, ensuring the stability and convergence of model optimization under heavy-tailed error scenarios. It is a key parameter connecting the characteristics of the loss function and the design of the optimization algorithm. A step size less than 2 / L avoids parameter update overshoot; it prevents slow convergence due to a step size that is too small, and oscillations around the stationary point due to a step size that is too large. This prevents training oscillations caused by heavy-tailed error and ensures that the algorithm eventually converges to the stationary point.

[0055] In one example, the adaptive Student-t loss function is configured as a plug-and-play loss plugin for easy invocation. It is important to emphasize that this application does not change the core architecture of the energy prediction model; performance improvements can be achieved simply by replacing the original MSE loss with the adaptive Student-t loss function, demonstrating strong versatility and ease of use.

[0056] Adaptive Student-t Loss Function Satisfying the expression: ; The local Lipschitz continuity constant satisfies the expression: ; in, For prediction error, For degrees of freedom parameters; prediction error Bounded and exists , making M is a positive number; degrees of freedom parameter Bounded and ,in, , and They are respectively The lower and upper bounds; when At that time, the Student-t loss consistently converges to the MSE loss, thus achieving a smooth transition from the heavy-tailed mechanism to the Gaussian mechanism.

[0057] In the embodiments of this application, it is understood that gradient The second derivative is bounded: The Lipschitz continuity can be proven using the mean value theorem.

[0058] It should be noted that, for Using Taylor expansion: log(1+x)=x 2x 2 +O(x 3 From this, we can obtain: For all | When |≤M, the remainder term uniformly vanishes, and convergence is uniform. Therefore, when hour, converged to It can be seen that the Student-t loss consistently converges to the MSE loss, ensuring prediction efficiency in conventional scenarios and achieving a balance between robustness and efficiency. This proves that the adaptive Student-t loss function can achieve a smooth transition between heavy-tailed and Gaussian mechanisms, and demonstrates the rationality of using the Relative Difficulty Index (RDI) for dynamic adaptation. These results show that the adaptive framework corresponding to the adaptive Student-t loss function proposed in this application maintains theoretical reliability while possessing the flexibility to cope with varying noise environments.

[0059] Based on this, this application introduces a Relative Difficulty Index (RDI) to achieve a fine-grained quantification of the prediction difficulty for each sample. This allows the model to break away from the traditional one-size-fits-all optimization paradigm and dynamically adjust the shape of the loss function according to the difficulty of the sample itself. For simple samples with low prediction difficulty, a loss approximating MSE is used to maintain efficient learning; for difficult samples with high prediction difficulty, a heavy-tailed loss is automatically switched to enhance robustness to outliers. This differentiated strategy enables the model to better adapt to the inherent heterogeneity in renewable energy data.

[0060] It's also understandable that traditional MSE loss, due to the Gaussian assumption, is prone to gradient explosion (extreme errors lead to unbounded gradients), causing training oscillations. In contrast, the local Lipshitz continuity of the adaptive Student-t loss, by constraining the maximum magnitude of the gradient through L, coupled with the step size constraint of η < 2 / L, fundamentally avoids gradient explosion and training divergence, making optimization more stable in scenarios with heavy-tailed errors. The adaptive Student-t loss utilizes the degree-of-freedom parameter... The dynamic adjustment ensures both the robustness of heavy-tailed samples in difficult samples and the efficient optimization characteristics of easy samples.

[0061] The improvements in this application revolve around heavy-tailed errors. The adaptive Student-t loss allows the model to focus on extreme error samples during training, while the stable convergence of gradient descent ensures the model can fully learn the patterns of these samples, ultimately improving prediction accuracy and robustness in scenarios such as extreme weather and sudden pattern changes—a core requirement for renewable energy prediction. The driving logic of the heavy-tailed mechanism constructed in this application is as follows: the adaptive Student-t loss is responsible for accurately modeling heavy-tailed errors; local Lipshitz continuity ensures the stable optimization of the loss function; and gradient descent with step size constraints is responsible for efficient convergence to the optimal parameters. These three elements work together to implement the heavy-tailed mechanism. This application not only solves the optimization instability of traditional loss functions under heavy-tailed errors but also balances convergence efficiency and prediction robustness, ultimately enabling the model to systematically utilize error distribution characteristics, thereby improving the accuracy and reliability of renewable energy prediction.

[0062] In some embodiments, the adaptive mapping between the relative difficulty index RDI and the degrees of freedom parameters of the Student-t distribution is a negative correlation mapping; the adaptive mapping is implemented using the Sigmoid function, whose input is the relative difficulty index RDI and output is the normalized values ​​of the degrees of freedom parameters; the adaptive mapping process satisfies the expression: ; Where ν(RDI) represents the degree of freedom parameter output based on the relative difficulty index RDI mapping, g( () represents the sigmoid function and supports end-to-end training of the model through its differentiability. It is a positive value and is used to control the transition sensitivity.

[0063] In the embodiments of this application, it can be understood that the higher the Relative Difficulty Index (RDI) value, the higher the degree of freedom parameters. The smaller the value, the more pronounced the heavy-tailed characteristic, and the better it can resist extreme errors; the lower the relative difficulty index (RDI) value, the more degrees of freedom parameters... The larger the value, the closer the loss function approximates the MSE, ensuring optimization efficiency for regular samples. Degrees of freedom parameter These are the core parameters for constructing the adaptive Student-t loss function, specifically the degrees of freedom parameter. It is associated with the Relative Difficulty Index (RDI); this application uses the Relative Difficulty Index (RDI) to distinguish between samples with different prediction difficulties, enabling the model to adopt differentiated optimization strategies for easy and difficult samples.

[0064] Furthermore, the output of the sigmoid function changes relatively smoothly, allowing the degree of freedom parameters to change more gradually. exist and This allows for continuous and smooth adjustments, rather than abrupt changes. The sigmoid function is continuously differentiable, ensuring that the relative difficulty index RDI is consistent with the degree of freedom parameters. The mapping process can participate in backpropagation.

[0065] Furthermore, The degree of freedom parameters can be controlled by adjusting its size. Sensitivity to changes in the Relative Difficulty Index (RDI) The larger the value, the more a small change in the relative difficulty index (RDI) will affect the degree of freedom parameters. Quick switching (suitable for datasets with significant differences in difficulty, such as wind power data with frequent typhoons); The smaller the value, the more degrees of freedom parameters. The smoother the change (suitable for datasets with uniform difficulty distribution, such as solar energy data in stable regions). The technical solution proposed in this application is universal; without modifying the mapping logic, it only requires adjusting... It can adapt to datasets of different renewable energy types and different regions.

[0066] Based on this, this application adapts the heavy-tailed error characteristics through the negative correlation logic between the relative difficulty index RDI and the degree of freedom parameters, and ensures training stability and end-to-end optimization capabilities through the sigmoid function. The parameters enhance scenario adaptability and provide a flexible and reliable parameter adjustment mechanism for the adaptive Student-t loss function.

[0067] In some embodiments, this application provides a new energy power prediction system 300 based on heavy-tailed distribution modeling, such as... Figure 3 As shown, the new energy power prediction system 300 based on heavy-tailed distribution modeling may include the following modules: The data flattening and partitioning module 310 is used to flatten the acquired feature sequence and label sequence into a one-dimensional structure based on time series, and divide them into a first dataset and a second dataset under two consecutive time periods; the feature sequence is the historical meteorological data sequence of renewable energy, and the label sequence is the historical power generation sequence corresponding to the feature sequence; The relative difficulty index calculation module 320 is used to analyze the higher-order moment features and distribution distance of samples between the first and second datasets, and to calculate the relative difficulty index (RDI) for each sample that represents the error distribution characteristics. The loss function construction module 330 is used to obtain the degree of freedom parameters of the Student-t distribution based on the dynamic mapping of the relative difficulty index RDI to characterize the degree of heavy-tailed error, and to construct an adaptive Student-t loss function based on the degree of freedom parameters to model the heavy-tailed error distribution. The integrated driver module 340 is used to integrate the adaptive Student-t loss function with the energy prediction model without changing the forward structure of the preset energy prediction model, and to drive the construction of the heavy-tail mechanism. The heavy-tail mechanism is used to handle the heavy-tail error distribution and includes: determining the loss value of the adaptive Student-t loss function as the optimization objective and replacing the model's MSE loss, and adaptively adjusting the degree of freedom parameter of the Student-t loss through the relative difficulty index. Training module 350 is used to update model parameters through backpropagation of the loss value based on the adaptive Student-t loss function, so as to obtain an optimized target prediction model for multi-step prediction of renewable energy power generation.

[0068] According to embodiments of this application, any and multiple modules among the data flattening and partitioning module 310, the relative difficulty index calculation module 320, the loss function construction module 330, the integration driving module 340, and the training module 350 can be merged into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module.

[0069] Figure 3 Each module in the system shown has the function of implementing each step in the aforementioned new energy power prediction method based on heavy-tailed distribution modeling, and can achieve its corresponding technical effect. For the sake of brevity, it will not be elaborated here.

[0070] In some embodiments, this application provides an electronic device, the structural schematic of which is shown below. Figure 4 As shown.

[0071] The electronic device may include a processor 410 and a memory 420 storing computer program instructions.

[0072] Specifically, the processor 410 may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0073] Memory 420 may include mass storage for data or instructions. For example, and not limitingly, memory 420 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 420 may include removable or non-removable (or fixed) media. Where appropriate, memory 420 may be internal or external to the integrated gateway disaster recovery device. In a particular embodiment, memory 420 is non-volatile solid-state memory.

[0074] Memory 420 may include read-only memory (ROM), random access memory (RAM), disk storage media device, optical storage media device, flash memory device, electrical, optical, or other physical / tangible memory storage device. Therefore, typically, memory 420 includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it can perform the operations described in any of the new energy power prediction methods based on heavy-tailed distribution modeling in the above embodiments.

[0075] The processor 410 reads and executes computer program instructions stored in the memory 420 to implement any of the new energy power prediction methods based on heavy-tailed distribution modeling in the above embodiments.

[0076] In one example, the electronic device may also include a communication interface 430 and a bus 400. For example, Figure 4 As shown, the processor 410, memory 420, and communication interface 430 are connected via bus 400 and communicate with each other.

[0077] The communication interface 430 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application. Bus 400 includes hardware, software, or both, that couples components of an online data traffic metering device together. For example, and not limitingly, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 400 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, any suitable bus or interconnect is contemplated herein.

[0078] Furthermore, in conjunction with the new energy power prediction method based on heavy-tailed distribution modeling in the above embodiments, this application embodiment can provide a computer storage medium for implementation. The computer storage medium stores computer program instructions; when these computer program instructions are executed by a processor, they implement any of the new energy power prediction methods based on heavy-tailed distribution modeling in the above embodiments.

[0079] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0080] The functional blocks shown in the above block diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. Programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0081] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0082] The aspects of this disclosure have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It should be understood that each block in the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by special-purpose hardware performing the specified functions or actions, or can be implemented by a combination of special-purpose hardware and computer instructions.

[0083] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A new energy power prediction method based on heavy-tailed distribution modeling, characterized in that, include: Based on time series, the acquired feature sequences and label sequences are flattened into a one-dimensional structure and divided into a first dataset and a second dataset under two consecutive time periods; the feature sequences are historical meteorological data sequences of renewable energy, and the label sequences are historical power generation sequences corresponding to the feature sequences; Analyze the higher-order moment features and distribution distance of samples between the first dataset and the second dataset, and calculate the relative difficulty index (RDI) for each sample to characterize the error distribution. The degree-of-freedom parameters of the Student-t distribution are obtained by dynamic mapping based on the relative difficulty index RDI to characterize the degree of heavy-tailed error. An adaptive Student-t loss function is constructed based on the degree-of-freedom parameters to model the heavy-tailed error distribution. Without changing the forward structure of the preset energy prediction model, the adaptive Student-t loss function is integrated with the energy prediction model to drive the construction of a heavy-tail mechanism; the heavy-tail mechanism is used to handle the heavy-tail error distribution and includes: determining the loss value of the adaptive Student-t loss function as the optimization objective and replacing the model's MSE loss, and adaptively adjusting the degree of freedom parameter of the Student-t loss through the relative difficulty index. The model parameters are updated by backpropagation based on the loss value of the adaptive Student-t loss function, resulting in an optimized target prediction model for multi-step prediction of renewable energy power generation.

2. The new energy power prediction method based on heavy-tailed distribution modeling as described in claim 1, characterized in that, The adaptive Student-t loss function is determined based on the degree of freedom parameter and the prediction error; the adaptive Student-t loss function is locally Lipschitz continuous with respect to the prediction error; In driving the heavy-tailed mechanism, a step size is used. Gradient descent is used to optimize the adaptive Student-t loss so that the algorithm converges to a stationary point, where η is the step size and L is the local Lipshitz continuity constant of the adaptive Student-t loss function with respect to the prediction error.

3. The new energy power prediction method based on heavy-tailed distribution modeling as described in claim 2, characterized in that, The adaptive Student-t loss function is configured as a plug-and-play loss plugin for use; the adaptive Student-t loss function Satisfying the expression: ； The local Lipschitz continuity constant satisfies the expression: ； in, For prediction error, For degrees of freedom parameters; prediction error Bounded and exists , making M is a positive number; degrees of freedom parameter Bounded and ,in, , and They are respectively The lower and upper bounds; when At that time, the Student-t loss consistently converges to the MSE loss, thus achieving a smooth transition from the heavy-tailed mechanism to the Gaussian mechanism.

4. The new energy power prediction method based on heavy-tailed distribution modeling as described in claim 3, characterized in that, The adaptive mapping between the relative difficulty index RDI and the degree of freedom parameter of the Student-t distribution is a negative correlation mapping; The adaptive mapping is implemented using the Sigmoid function, whose input is the Relative Difficulty Index (RDI) and output is the normalized value of the degree of freedom parameter; the adaptive mapping process satisfies the expression: ； Where ν(RDI) represents the degree of freedom parameter output based on the relative difficulty index RDI mapping, g( () represents the sigmoid function and supports end-to-end training of the model through its differentiability. It is a positive value and is used to control the transition sensitivity.

5. The new energy power prediction method based on heavy-tailed distribution modeling as described in claim 1, characterized in that, The analysis of the higher-order moment features and distribution distance of samples between the first dataset and the second dataset, and the calculation of the relative difficulty index (RDI) for representing the error distribution characteristics of each sample, includes: The normalized Wasserstein distance between the first dataset and the second dataset is analyzed to capture macroscopic pattern differences between data sequences; The skewness and kurtosis differences between the first dataset and the second dataset are analyzed to quantify the non-Gaussian variation of the data sequences; wherein the skewness difference is used to characterize asymmetric pattern changes, and the kurtosis difference is used to characterize tail behavior changes. The relative difficulty index (RDI) is obtained by fusing skewness and kurtosis differences using a modulation function and then weighting the results using a normalized Wasserstein distance.

6. The new energy power prediction method based on heavy-tailed distribution modeling as described in claim 5, characterized in that, Modulation function Satisfying the expression: in, and These are learnable weight parameters. Indicates skewness difference. Indicates kurtosis variation; ； ； in, Represents the first dataset X skewness, Represents the second dataset Y skewness, Represents the first dataset X peak, Represents the second dataset Y Peak value; ， ； In the formula, Z represents a sample in the first dataset or the second dataset. This represents the mean of all samples in either the first or second dataset. represents the standard deviation of each sample in the first or second dataset, and E represents the expected value.

7. The new energy power prediction method based on heavy-tailed distribution modeling as described in any one of claims 1-6, characterized in that, The renewable energy sources include wind or solar energy. The historical meteorological data sequence is derived from the SCADA system for monitoring and data acquisition and meteorological monitoring equipment, and includes air pressure, relative humidity, temperature, dew point, wind direction at the target layer, and wind speed at the target layer. The energy prediction model is a variant of the Transformer model and includes one of the PatchTST model, iTransformer model, and TimeXer model; the energy prediction model is trained using the Adam optimizer, combined with an early stopping strategy and a weight decay mechanism for end-to-end optimization.

8. A new energy power prediction system based on heavy-tailed distribution modeling, characterized in that, include: The data flattening and partitioning module is used to flatten the acquired feature sequence and label sequence into a one-dimensional structure based on time series, and divide them into a first dataset and a second dataset under two consecutive time periods; the feature sequence is a historical meteorological data sequence of renewable energy, and the label sequence is a historical power generation sequence corresponding to the feature sequence; The relative difficulty index calculation module is used to analyze the higher-order moment features and distribution distance of samples between the first dataset and the second dataset, and to calculate the relative difficulty index (RDI) for each sample that represents the error distribution characteristics. The loss function construction module is used to obtain the degree-of-freedom parameters of the Student-t distribution based on the relative difficulty index RDI dynamic mapping to characterize the degree of heavy-tailed error, and to construct an adaptive Student-t loss function based on the degree-of-freedom parameters to model the heavy-tailed error distribution. An integrated driving module is used to integrate the adaptive Student-t loss function with the energy prediction model without changing the forward structure of the preset energy prediction model, and to drive the construction of a heavy-tail mechanism. The heavy-tail mechanism is used to handle the heavy-tail error distribution and includes: determining the loss value of the adaptive Student-t loss function as the optimization objective and replacing the model's MSE loss, and adaptively adjusting the degree of freedom parameter of the Student-t loss through the relative difficulty index. The training module is used to backpropagate and update the model parameters based on the loss value of the adaptive Student-t loss function to obtain an optimized target prediction model for multi-step prediction of renewable energy power generation.

9. An electronic device, characterized in that, include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the new energy power prediction method based on heavy-tailed distribution modeling as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program or instructions that, when executed by a processor, implement the new energy power prediction method based on heavy-tailed distribution modeling as described in any one of claims 1 to 7.

Images