Evaluation and optimization model for multiple indexes of new energy power prediction system

Abstract

The invention relates to the technical field of new energy power generation power prediction. The core defect of the prior art is that evaluation dimensions matched with new energy power characteristics cannot be constructed, depiction of error time correlation is lacked, a statistical framework adaptive to non-Gaussian distribution is not established, and dynamic weight optimization of evaluation indexes cannot be realized. The invention provides an evaluation and optimization model for multiple indexes of a new energy power prediction system, and the model carries out the evaluation and weight optimization of the multi-index error characteristics of the new energy power prediction system through the dynamic weighting measurement of a longitudinal error, the time-delay correlation analysis of a transverse error and the tensor representation of a peak-valley value error. The comprehensive score accuracy output by the new energy power prediction system is improved, the longitudinal error, the transverse error and the peak-valley value error are fused to form a multi-index evaluation system, space-time multi-dimensional quantization of the prediction error is achieved, and the evaluation precision bottleneck of a traditional method is broken through.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation prediction technology, and more specifically, to an evaluation and optimization model for multiple indicators of a new energy power prediction system. Background Technology

[0002] The development of renewable energy power prediction and evaluation technology has long been constrained by the disconnect between mathematical models and real-world scenarios. Current mainstream evaluation methods, whether based on single-index systems using the mean or simple combinations of multiple indices, struggle to capture the unique intermittent, time-varying, and non-Gaussian distribution characteristics of renewable energy power. Traditional MAE (Magnitude of Error) indicators can only describe the average level of error, failing to reflect its fluctuation characteristics. According to wind farm measurement data, when prediction errors fluctuate alternately between positive and negative, the actual fluctuation amplitude can reach twice the MAE value. This lack of information leads to misjudgments of the true performance of the prediction system. Furthermore, MRE (Mean Reliability and Evaluation) indicators exhibit computational singularities when renewable energy output approaches zero due to the denominator approaching zero. Tests of photovoltaic power plants during low-wind-speed periods show that when actual output is below 0.1MW, an absolute error of only 0.2MW can cause the MRE indicator to soar to over 300%, completely losing its engineering guidance value.

[0003] Existing statistical evaluation methods often assume that the error sequence follows a normal distribution, but a large amount of measured data refutes this premise. In 2024, the kurtosis of the full error data for a wind power project reached 4.72, far exceeding the eigenvalue of 3 for a normal distribution, exhibiting significant leptokurtosis and heavy tails; the skewness of 1.23 indicates a significant shift in the error distribution. This non-Gaussian property renders traditional methods such as the DM test and adaptive evaluation functions ineffective—when the error distribution rejects the normality assumption, the p-value deviation rate of the statistical test exceeds 40%, making it impossible to reliably determine the true differences between different prediction models. More importantly, existing indicator systems are insensitive to the temporal correlation between the predicted and measured sequences, failing to distinguish between "overall numerical deviation" and "peak time shift." A wind farm case study shows that when the predicted peak value led the actual peak value by 30 minutes, the RMSE increased by only 0.2MW, severely lacking differentiation from longitudinal errors, causing the dispatch system to fail to identify prediction deviations on the time axis.

[0004] The development of multi-indicator evaluation methods also faces bottlenecks. Existing combined indicators often employ equal weighting or static weighting strategies, failing to consider the correlation between indicators and unable to adapt to the evaluation needs of different meteorological scenarios. Data analysis from a photovoltaic power station shows that the correlation coefficient between MAE and RMSE is as high as 0.87, with equal weighting leading to information redundancy. Furthermore, during high wind speed periods, the importance of peak-valley error indicators should be increased to over 40%, but traditional models still use a uniform weighting configuration. This blindness makes it difficult for the accuracy of multi-indicator evaluations to exceed 80%, failing to meet the requirements of power grid assessment. Statistics from a provincial dispatch center show that the misjudgment rate of prediction methods due to evaluation model defects reaches 27%, resulting in an average annual assessment loss of over 5 million yuan.

[0005] In engineering applications, wind and solar power prediction assessment accounts for more than 35% of the scheduling indicator system. However, existing evaluation technologies cannot solve three core contradictions: First, the limitations of single model optimization. For example, although LSTM neural networks can improve the accuracy of numerical prediction, they are difficult to improve the peak time lag problem, and the overall pass rate can only be improved by 5-8%. Second, the lack of optimal selection of multiple prediction systems. When the prediction system indicators of different manufacturers are similar (e.g., MAE of 2.1MW and 2.3MW respectively), traditional methods cannot distinguish their adaptability differences under extreme weather conditions. Third, insufficient real-time performance. Static weight models based on historical data cannot adapt to meteorological changes. For example, when a typhoon passes, the weight of the wind speed influence factor needs to be dynamically increased from 15% to 30%, but existing technologies lack corresponding adjustment mechanisms.

[0006] In summary, the core deficiency of existing technologies lies in their failure to construct evaluation dimensions that match the characteristics of renewable energy power—lacking both characterization of the temporal correlation of errors and a statistical framework adapted to non-Gaussian distributions, as well as the ability to dynamically optimize the weights of evaluation indicators. These shortcomings lead to a severe disconnect between evaluation results and engineering needs, necessitating the establishment of a new generation of multi-indicator evaluation systems that integrate spatiotemporal error characteristics, dynamic weighting mechanisms, and nonparametric statistical theory to overcome the current technical bottleneck in improving the accuracy of renewable energy power prediction. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the present invention aims to provide an evaluation and optimization model for multiple indicators in new energy power prediction systems. This invention mainly solves the core problems in existing new energy power prediction and evaluation technologies, such as the lack of single-indicator evaluation dimensions, the conflict between statistical assumptions and actual data distribution, insufficient characterization of horizontal errors, and the blindness of multi-indicator weights. It integrates vertical errors, horizontal errors, and peak-valley errors to form a multi-indicator evaluation system, realizing the spatiotemporal multi-dimensional quantification of prediction errors and breaking through the evaluation accuracy bottleneck of traditional methods.

[0008] To achieve the above objectives, the present invention provides the following technical solution: An evaluation and optimization model for multiple indicators of a new energy power prediction system includes a multi-dimensional quantification module for longitudinal errors, a spatiotemporal correlation analysis module for lateral errors, and a peak-valley error calculation module. This model evaluates and optimizes the error characteristics of multiple indicators of the new energy power prediction system by dynamically weighting the longitudinal errors, analyzing the time-delay correlation of lateral errors, and representing peak-valley errors using tensors. This improves the accuracy of the comprehensive score output by the new energy power prediction system. Specifically: the multi-dimensional quantification module for longitudinal errors constructs a weighted dynamic MAE model with a forgetting factor to determine the error frequency distribution of the new energy power prediction system; the spatiotemporal correlation analysis module for lateral errors quantifies the time offset between the predicted and measured sequences of the new energy power prediction system to obtain the lateral error time delay; and the peak-valley error calculation module constructs a three-dimensional error tensor to achieve accurate detection of the safety extreme points of the power grid.

[0009] Furthermore, a multidimensional quantization module for longitudinal errors is used to construct a weighted dynamic MAE model with a forgetting factor: In the formula: WMAE is the weighted average absolute error, which reflects the error level with time decay characteristics; n This represents the total number of samples, i.e., the number of time points included in the prediction period. The forgetting factor, with a value range of (0,1], controls the rate at which the weights of historical data decay. i For time indexing, from 1 to n These represent different times in sequence; For the first i The predicted power value at time; For the first i The actual power value at that moment; To address the non-Gaussian distribution characteristics of errors in new energy sources, a Gaussian kernel density estimation model is used to represent the error frequency distribution: In the formula: For bandwidth parameters; e For error variables; For the first i The actual error value of each sample, of which ; It is a natural exponential function.

[0010] Furthermore, the lateral error spatiotemporal correlation analysis module quantifies the time offset between the predicted and measured sequences of the new energy power prediction system, and defines the time delay correlation coefficient matrix: In the formula: Indicates time delay as The correlation coefficient, with a value range of [-1, 1], The time delay parameter represents the time offset between the predicted sequence and the measured sequence; The average value of the predicted power sequence; This is the average value of the actual power sequence; Represents the offset index of the measured sequence, when The time indicates that the predicted sequence leads the measured sequence; A time-delay correlation matrix is ​​constructed. To measure the similarity of non-equal-length sequences, the DTW distance is defined to represent the trend consistency between the predicted and measured sequences. in For time regularization path, satisfy , ,and , For prediction points and measured points The Euclidean distance.

[0011] Furthermore, the peak-valley error calculation module achieves accurate detection of extreme points through the Hessian matrix of the power sequence: in, Let be the second partial derivative of power with respect to time, representing the trend of the rate of change of power; , These are mixed partial derivatives, reflecting the interaction between power and time variations; The second-order partial derivative of power with respect to itself is used to help determine extreme value characteristics; For the first The Hessian matrix at time step 1 is used to detect the extreme points of the power sequence. When the matrix determinant is... And traces The peak point is determined at a certain time, and the accuracy of peak and valley value prediction is quantified by using the three-dimensional error tensor.

[0012] Furthermore, the three-dimensional error tensor In this context, the first dimension is a point in time. The second dimension is the error type: 1. Amplitude error, 2. Time error, 3. Existence error; the third dimension is the error direction: 1. Peak value, 2. Valley value; specific elements are defined as follows: in, For the first Power values ​​at each predicted peak point; : No. The power value at each actual peak point; This refers to the rated power of the power station; For the first The predicted peak point time; For the first The moment of the actual peak point; For time intervals.

[0013] Furthermore, a kernel space similarity measurement model is constructed, and historical data matching is achieved based on the regenerating kernel Hilbert space. A hybrid kernel function is used to map the Hessian matrix to the regenerating kernel Hilbert space RKHS. The kernel function is defined as follows: in This consists of two multi-indicator matrices, one representing historical data and the other representing current data. It is the Euclidean norm. For Manhattan norm, and Using Euclidean distance and Manhattan distance as kernel parameters, the accuracy of historical data matching is improved by fusing them. The fusion weights of Euclidean distance and Manhattan distance are... Construct a PSO intelligent weight optimization system and solve for the optimal weight vector based on the particle swarm optimization algorithm: in , For the first The indicator vector of the group data, with the following constraints: , ; This represents the weight vector that minimizes the objective function. ; This represents the number of training samples; : Transpose of the weight vector; For the first Indicator vectors for group data; : No. The actual assessment score of the group data; The particle update formula is: in For the first The particle in the first The velocity vector of the generation; Inertial weights control the influence of the particle's historical velocity. For the first The particle in the first The velocity vector of the generation; As a learning factor, , A random number in the range [0,1]. For the first The particle in the first The current position of the generation; For the optimal position of an individual, To find the globally optimal position, the optimized weight vector w is obtained. * This will improve the accuracy of the overall score output by the new energy power prediction system.

[0014] In summary, the invention has the following beneficial effects: The multi-index evaluation system model of this invention achieves a comprehensive characterization of errors. The longitudinal error index addresses the one-sidedness of mean statistics, the horizontal error index compensates for the lack of time correlation analysis, and the peak-valley error index strengthens the evaluation of extreme points related to power grid safety. The dynamic weighting mechanism adaptively adjusts the importance of indicators through particle swarm optimization algorithm, automatically increasing the weight of relevant indicators in scenarios such as high wind speeds and extreme weather, making the evaluation results more in line with actual needs. The kernel space similarity model introduces historical data correlation analysis, referring to the prediction performance under similar scenarios, avoiding the randomness of evaluation based on single current data, making the evaluation results more reliable and adaptable to different scenarios. The dynamic weighting mechanism based on particle swarm optimization in this invention automatically adjusts the weight of each indicator according to the characteristics of real-time data, solving the problem of blindness in traditional static weights. Through actual measurement verification, the evaluation accuracy is improved to 91.3%, providing a quantitative standard for the selection of multiple prediction systems in power plants, which can directly reduce assessment losses and improve power generation revenue. It is applicable to the evaluation and selection of new energy power prediction systems with strong intermittent characteristics such as wind power and photovoltaic power. Attached Figure Description

[0015] Figure 1 This is a flowchart of the multi-index evaluation model architecture of the present invention; Figure 2 This is a schematic diagram of the test set results for the multi-index matrix calculation of the present invention; Figure 3 This is a comparative chart of the comprehensive evaluation of five prediction methods. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the accompanying drawings.

[0017] It should be noted that, for ease of description, the descriptions of direction in the following text are consistent with the directions in the accompanying drawings, but they do not limit the structure of the present invention.

[0018] like Figures 1-3As shown, this invention discloses an evaluation and optimization model for multiple indicators of a new energy power prediction system. It includes a multi-dimensional quantification module for longitudinal errors, a spatiotemporal correlation analysis module for lateral errors, and a peak-valley error calculation module. By dynamically weighting the longitudinal errors, analyzing the time-delay correlation of lateral errors, and representing peak-valley errors using tensors, the model evaluates and optimizes the weights of the multi-indicator error characteristics of the new energy power prediction system, thereby improving the accuracy of the comprehensive score output by the new energy power prediction system. Specifically: the multi-dimensional quantification module for longitudinal errors constructs a weighted dynamic MAE model with a forgetting factor to determine the error frequency distribution of the new energy power prediction system; the multi-dimensional quantification module for longitudinal errors constructs a weighted dynamic MAE model with a forgetting factor. In the formula: WMAE is the weighted average absolute error, which reflects the error level with time decay characteristics; n This represents the total number of samples, i.e., the number of time points included in the prediction period. The forgetting factor, with a value range of (0,1], controls the rate at which the weights of historical data decay. The closer it is to 1, the greater the influence of historical data. When it is 1, it is the traditional MAE. In this method, the weight of the most recent third of the data can reach 60%, more sensitively reflecting the real-time performance of the prediction system. In a field test at a wind farm, this invention improved the response speed to sudden errors caused by equipment failure by 3 times compared to the traditional MAE. i For time indexing, from 1 to n These represent different times in sequence; For the first i The predicted power value at time; For the first i The actual power value at any given time.

[0019] To address the non-Gaussian distribution characteristics of errors in new energy sources, a Gaussian kernel density estimation model is used to represent the error frequency distribution: In the formula: For bandwidth parameters; e For error variables; For the first i The actual error value of each sample, of which ; The Gaussian kernel is used as a natural exponential function to achieve smoothing. Data from a photovoltaic power station shows that the goodness of fit of this invention to the error distribution reaches 0.92, which is 0.35 better than the traditional normal distribution assumption, and can accurately describe the characteristics of sharp peaks and thick tails. For the wind speed influencing factor, the wind speed is divided into 8 intervals to construct the conditional error expectation matrix: in Describe the average error for each wind speed range. This indicates the correlation between error and wind speed, providing a basis for error correction in high-wind-speed sections.

[0020] The lateral error spatiotemporal correlation analysis module quantifies the time offset between the predicted and measured sequences of the new energy power prediction system to obtain the lateral error time delay; the lateral error spatiotemporal correlation analysis module quantifies the time offset between the predicted and measured sequences of the new energy power prediction system and defines the time delay correlation coefficient matrix: In the formula: Indicates time delay as The correlation coefficient ranges from -1 to 1, with a larger absolute value indicating a stronger correlation. The time delay parameter represents the time offset between the predicted sequence and the measured sequence; The average value of the predicted power sequence; This is the average value of the actual power sequence; Represents the offset index of the measured sequence, when Time indicates that the predicted sequence leads the measured sequence; construct 7-dimensional matrix ,when and When the predicted peak value is ahead of the actual value by 1 time interval (15 minutes), the present invention solves the problem that traditional indicators cannot distinguish between longitudinal error and lateral time lag.

[0021] A time-delay correlation matrix is ​​constructed. To measure the similarity of non-equal-length sequences, the DTW distance is defined to represent the trend consistency between the predicted and measured sequences. in For time regularization path, satisfy , ,and , For prediction points and measured points The Euclidean distance is used. In scenarios of non-stationary power fluctuations caused by typhoon passage, this invention reflects the trend consistency between the predicted and measured sequences better than the Euclidean distance. One case shows that the matching accuracy of this invention is improved by 27%.

[0022] The peak-valley error calculation module constructs a three-dimensional error tensor to achieve accurate detection of the power grid's safety extreme points. This module utilizes the Hessian matrix of the power sequence to achieve precise detection of these extreme points. in, Let be the second partial derivative of power with respect to time, representing the trend of the rate of change of power; , These are mixed partial derivatives, reflecting the interaction between power and time variations; The second-order partial derivative of power with respect to itself is used to help determine extreme value characteristics; For the first The Hessian matrix at time step 1 is used to detect the extreme points of the power sequence. When the matrix determinant is... And traces The peak point is determined at a given time, and the prediction accuracy of peak and valley values ​​is quantified using a three-dimensional error tensor. When the matrix determinant... And traces The peak point was determined at a certain time, and the actual measurement of a wind farm showed that the accuracy of the extreme point identification of the present invention reached 98%.

[0023] 3D error tensor In this context, the first dimension is a point in time. The second dimension is the error type: 1. Amplitude error, 2. Time error, 3. Existence error; the third dimension is the error direction: 1. Peak value, 2. Valley value; specific elements are defined as follows: Amplitude error: Time error: Existence error: in, For the first Power values ​​at each predicted peak point; : No. The power value at each actual peak point; This refers to the rated power of the power station; For the first The predicted peak point time; For the first The moment of the actual peak point; The time interval is used. The accuracy of peak-valley forecasting is quantified from three dimensions: amplitude error, time error, and existence error. Data from a power grid dispatching system shows that after applying this tensor, the peak-valley forecast accuracy rate increased from 65% to 90%. A kernel space similarity measurement model is constructed, and historical data matching is achieved based on the reproducing kernel Hilbert space. A hybrid kernel function is used to map the Hessian matrix to the reproducing kernel Hilbert space RKHS. The kernel function is defined as follows: in This consists of two multi-indicator matrices, one representing historical data and the other representing current data. It is the Euclidean norm. For Manhattan norm, and Using Euclidean distance and Manhattan distance as kernel parameters, the accuracy of historical data matching is improved by fusing them. The fusion weights of Euclidean distance and Manhattan distance are... Construct a PSO intelligent weight optimization system and solve for the optimal weight vector based on the particle swarm optimization algorithm: in , For the first The indicator vector of the group data, with the following constraints: , ; This represents the weight vector that minimizes the objective function. ; This represents the number of training samples; : Transpose of the weight vector; For the first Indicator vectors for group data; : No. The actual assessment score of the group data; The particle update formula is: in For the first The particle in the first The velocity vector of the generation; Inertial weights control the influence of the particle's historical velocity. For the first The particle in the first The velocity vector of the generation; As a learning factor, , A random number in the range [0,1]. For the first The particle in the first The current position of the generation; For the optimal position of an individual, To find the globally optimal position, the optimized weight vector w is obtained. * , This improves the accuracy of the comprehensive score output by the new energy power prediction system, achieving a 91.3% match between the overall score and the actual assessment results, which is 23% higher than the traditional equal-weight method.

[0024] Comparative verification of five mainstream prediction methods: 1. Prediction Method Settings Method D: ARIMA(2,1,2) model, parameters determined by maximum likelihood estimation. Method E: Bidirectional LSTM neural network with 3 hidden layers (100 neurons per layer), and the loss function is WMAE of equation (1).

[0025] Method F: Wavelet packet decomposition (8 layers) + Transformer model, with the decomposition scaling function being the Daubechies4 wavelet. Method G: Support Vector Machine (SVM) model, using the RBF kernel function, optimized parameters C=10, γ=0.1 Method H: Random Forest (RF) model, containing 100 decision trees with a maximum depth of 15, using feature selection based on the wind speed influence factor matrix of Equation 78. 2. Experimental Data and Scenarios Data source: 15-minute resolution data from a wind farm in 2024. Training set: 19 months (12,960 samples) Validation set: October (1440 samples) Test set: November (1440 samples) Evaluation period: December 1-31, 2024, with 24-hour ultra-short-term pre-tests generated daily. Installed capacity: 50MW, wind speed measurement range: 0-25m / s. 3. Calculation of multi-indicator matrix (test set results) The results are as follows Figure 2 As shown.

[0026] 4. Kernel Space Similarity and PSO Optimization Historical reference period: December 2023 (similar meteorological conditions), constructing a historical multi-indicator matrix. .

[0027] Similarity calculation (Equation 16-17, σ=1.2, γ=0.3): Similarity to corresponding methods D / E / F / G / H.

[0028] PSO optimization parameters: Number of particles N=50, maximum number of iterations T=200 Inertia weight ω =0.7298, learning factors c1=c2=1.49618 Optimal weight vector after convergence: (The corresponding indicators are: WMAE, kernel density, wind speed covariance, correlation coefficient, DTW, peak error, and valley error.) 5. Overall Scoring and Method Ranking Comprehensive scoring formula: The calculation yields: A comparison chart of the comprehensive scores of the five prediction methods is shown below. Figure 3 As shown. The sorting result is: F>H>E>G>D.

[0029] Performance validation: Method F improves upon the traditional ARIMA model by 33.8%, reduces the error by 51.1% in the high wind speed range, and achieves 100% accuracy in peak and valley value forecasts, validating the effectiveness of the multi-index evaluation model.

[0030] This invention is rationally designed and applicable to the evaluation and optimization of power prediction systems for various new energy power plants, such as wind power and photovoltaic power. It combines mathematical modeling with intelligent optimization, preserving error indicators with clear physical meaning while enhancing evaluation adaptability through machine learning algorithms, demonstrating excellent technical and economic efficiency in engineering applications. Through triple innovation—multi-indicator synergy, dynamic weight optimization, and historical data mining—this invention effectively solves the technical challenges of new energy power prediction and evaluation, possessing significant practical application value and broad market prospects.

[0031] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A multi-index evaluation and optimization model for new energy power prediction systems, characterized in that: The system includes a multidimensional quantification module for longitudinal errors, a spatiotemporal correlation analysis module for lateral errors, and a peak-valley error calculation module. By dynamically weighting the longitudinal errors, analyzing the time-delay correlation of the lateral errors, and using tensor representation of the peak-valley errors, the system evaluates and optimizes the weights of the multi-index error characteristics of the new energy power prediction system, thereby improving the accuracy of the comprehensive score output by the new energy power prediction system. Specifically, the multidimensional quantification module for longitudinal errors constructs a weighted dynamic MAE model with a forgetting factor to determine the error frequency distribution of the new energy power prediction system. The lateral error spatiotemporal correlation analysis module quantifies the time offset between the prediction sequence and the measured sequence of the new energy power prediction system to obtain the lateral error time delay. The peak-valley error calculation module constructs a three-dimensional error tensor to achieve accurate detection of the power grid's safety extreme points.

2. The multi-index evaluation and optimization model for new energy power prediction systems according to claim 1, characterized in that: The multidimensional quantization module of the longitudinal error constructs a weighted dynamic MAE model with a forgetting factor: In the formula: WMAE is the weighted average absolute error, which reflects the error level with time decay characteristics; n This represents the total number of samples, i.e., the number of time points included in the prediction period. The forgetting factor, with a value range of (0,1], controls the rate at which the weights of historical data decay. i For time indexing, from 1 to n These represent different times in sequence; For the first i The predicted power value at time; For the first i The actual power value at that moment; To address the non-Gaussian distribution characteristics of errors in new energy sources, a Gaussian kernel density estimation model is used to represent the error frequency distribution: In the formula: For bandwidth parameters; e For error variables; For the first i The actual error value of each sample, of which ; It is a natural exponential function.

3. The multi-index evaluation and optimization model for new energy power prediction systems according to claim 1, characterized in that: The horizontal error spatiotemporal correlation analysis module quantifies the time offset between the predicted and measured sequences of the new energy power prediction system and defines the time delay correlation coefficient matrix: In the formula: Indicates time delay as The correlation coefficient, with a value range of [-1, 1], The time delay parameter represents the time offset between the predicted sequence and the measured sequence; The average value of the predicted power sequence; This is the average value of the actual power sequence; Represents the offset index of the measured sequence, when The time indicates that the predicted sequence leads the measured sequence; A time-delay correlation matrix is ​​constructed. To measure the similarity of non-equal-length sequences, the DTW distance is defined to represent the trend consistency between the predicted and measured sequences. in For time regularization path, satisfy , ,and , For prediction points and measured points The Euclidean distance.

4. The multi-index evaluation and optimization model for new energy power prediction systems according to claim 1, characterized in that: The peak-valley error calculation module achieves accurate detection of extreme points through the Hessian matrix of the power sequence: in, Let be the second partial derivative of power with respect to time, representing the trend of the rate of change of power; , These are mixed partial derivatives, reflecting the interaction between power and time variations; The second-order partial derivative of power with respect to itself is used to help determine extreme value characteristics; For the first The Hessian matrix at time step 1 is used to detect the extreme points of the power sequence. When the matrix determinant is... And traces The peak point is determined at a certain time, and the accuracy of peak and valley value prediction is quantified by using the three-dimensional error tensor.

5. The multi-index evaluation and optimization model for new energy power prediction systems according to claim 4, characterized in that: The three-dimensional error tensor In this context, the first dimension is a point in time. The second dimension is the error type:

1. Amplitude error, 2. Time error, 3. Existence error; the third dimension is the error direction:

1. Peak value, 2. Valley value; specific elements are defined as follows: in, For the first Power values ​​at each predicted peak point; : No. The power value at each actual peak point; This refers to the rated power of the power station; For the first The predicted peak point time; For the first The moment of the actual peak point; For time intervals.

6. The multi-index evaluation and optimization model for new energy power prediction systems according to claim 1, characterized in that: A kernel space similarity measurement model is constructed, and historical data matching is achieved based on the regenerating kernel Hilbert space. A hybrid kernel function is used to map the Hessian matrix to the regenerating kernel Hilbert space RKHS. The kernel function is defined as follows: in This consists of two multi-indicator matrices, one representing historical data and the other representing current data. It is the Euclidean norm. For Manhattan norm, and Using Euclidean distance and Manhattan distance as kernel parameters, the accuracy of historical data matching is improved by fusing them. The fusion weights of Euclidean distance and Manhattan distance are... Construct a PSO intelligent weight optimization system and solve for the optimal weight vector based on the particle swarm optimization algorithm: in , For the first The indicator vector of the group data, with the following constraints: , ; This represents the weight vector that minimizes the objective function. ; This represents the number of training samples; : Transpose of the weight vector; For the first Indicator vectors for group data; : No. The actual assessment score of the group data; The particle update formula is: in For the first The particle in the first The velocity vector of the generation; Inertial weights control the influence of the particle's historical velocity. For the first The particle in the first The velocity vector of the generation; As a learning factor, , A random number in the range [0,1]. For the first The particle in the first The current position of the generation; For the optimal position of an individual, To find the globally optimal position, the optimized weight vector w is obtained. * This will improve the accuracy of the overall score output by the new energy power prediction system.

Images