A new energy prediction error online correction method

Abstract

The present application relates to the field of power system new energy, especially to a new energy prediction error online correction method. The method first collects the measured output and corresponding prediction output data of the new energy power station in real time, and calculates the original error time series; the sequence is decomposed by empirical mode decomposition to separate multiple intrinsic mode components of different time scales, and the component error prediction value is obtained by establishing an autoregressive moving average model respectively, and the error trend prediction result is obtained after reconstruction; at the same time, the synchronous measured error data of the adjacent new energy power station is obtained, the spatial correlation weight is calculated, and the spatial error correction amount is obtained by spatial weighted average method; the error trend prediction result and the spatial error correction amount are adaptively fused to obtain the final correction value, which is superimposed on the original prediction output and output. The present application effectively solves the problem of limited correction accuracy of single dimension in the prior art and lack of dynamic adaptability of fusion method, and significantly improves the accuracy of new energy output prediction.

Description

Technical Field

[0001] This invention relates to the field of new energy in power systems, and in particular to an online method for correcting new energy prediction errors. Background Technology

[0002] With the accelerated pace of the global energy transition, the installed capacity of new energy power generation, represented by wind and solar power, has continued to grow rapidly and has become an important component of the power system. However, the output of new energy power generation is characterized by significant intermittency, volatility, and uncertainty. Its output is influenced by a combination of meteorological factors such as wind speed, solar intensity, temperature, and cloud cover, making accurate prediction difficult. Deviations in new energy output prediction directly lead to increased pressure on grid peak shaving and frequency regulation, exacerbate wind and solar curtailment, and in severe cases, even threaten the safe and stable operation of the grid. Therefore, improving the accuracy of new energy output prediction, especially achieving online real-time correction of prediction errors, has become a critical technical issue that urgently needs to be addressed for the large-scale grid integration and consumption of new energy.

[0003] Currently, methods for predicting renewable energy output are mainly classified into three categories: physical modeling, statistical modeling, and machine learning modeling. Physical modeling is based on meteorological principles and the characteristics of power generation equipment, but it requires extremely high accuracy in meteorological data and struggles to capture the influence of complex terrain and micro-meteorological conditions. Statistical modeling predicts output by analyzing the statistical relationship between historical output data and meteorological data, performing well in short-term forecasts, but its ability to handle non-stationary and nonlinear error sequences is limited. Machine learning modeling, with its powerful nonlinear fitting capabilities, has been widely used in recent years, but model training requires a large amount of high-quality historical data, and its generalization ability under extreme weather conditions is insufficient. Even the most advanced prediction models cannot completely eliminate prediction errors; therefore, specialized error correction methods are needed to further improve prediction accuracy.

[0004] Most existing methods for correcting errors in renewable energy forecasting focus solely on the time dimension, using historical error sequences to build predictive models, such as autoregressive moving average models and long short-term memory networks. While these methods can capture the temporal evolution of errors effectively, they neglect the spatial correlation between errors from adjacent renewable energy power plants. In reality, multiple renewable energy power plants within the same region are often affected by similar or identical meteorological conditions, resulting in strong spatial correlations in their output errors. Although some studies have attempted to combine spatiotemporal information for error correction, they typically use fixed weights to fuse temporal prediction results and spatial correction amounts. This fails to dynamically adjust the fusion ratio based on real-time operating conditions and meteorological conditions, leading to significant differences in correction effectiveness across different scenarios and insufficient robustness. Furthermore, traditional time series decomposition methods are prone to confusing high-frequency noise with valid signals when processing complex error sequences, further affecting the accuracy of error correction. Summary of the Invention

[0005] This invention provides an online correction method for new energy prediction errors, aiming to solve the problems of limited accuracy of single-dimensional correction and lack of dynamic adaptability of fusion methods in the existing technology.

[0006] To achieve the above objectives, the following technical solution is adopted.

[0007] A method for online correction of new energy prediction errors includes the following steps: Real-time acquisition of measured power output data and corresponding predicted power output data of new energy power plants; calculation of prediction error sequence; and obtaining original error time series. Empirical mode decomposition is performed on the original error time series to separate multiple intrinsic mode components at different time scales. An autoregressive moving average model is then established for each intrinsic mode component to obtain the component error prediction value. The error prediction values ​​of each component are reconstructed to obtain the error trend prediction result; Synchronous measured power output error data are obtained from adjacent new energy power plants, spatial correlation weights are calculated, and spatial error correction is obtained using the spatial weighted average method. The error trend prediction result is adaptively fused with the spatial error correction amount to obtain the final error correction value, and the final error correction value is superimposed on the original predicted output data to output the corrected output prediction value.

[0008] Optionally, the step of performing empirical mode decomposition on the original error time series to separate multiple intrinsic mode components at different time scales, and establishing an autoregressive moving average model for each intrinsic mode component to obtain the component error prediction value, specifically includes: The local maxima and local minima of the original error time series are fitted by cubic spline interpolation to obtain the upper and lower envelopes. The mean envelopes of the upper and lower envelopes are calculated, and the candidate intrinsic mode function components are obtained by subtracting the mean envelope from the original error time series. Repeatedly perform envelope fitting and subtraction operations until the candidate intrinsic mode function component satisfies the definition condition of intrinsic mode function, separate an intrinsic mode function component from the original error time series and obtain the remaining component; Repeat the envelope fitting, subtraction and separation operations on the remaining components to separate multiple intrinsic mode function components with frequencies from high to low, until the remaining components are monotonic functions or the amplitude is less than the preset threshold, thus obtaining multiple intrinsic mode components at different time scales. For each intrinsic mode component, the autoregressive order and moving average order of the autoregressive moving average model are determined using the Akaike information criterion. After estimating the model parameters, the model is predicted one step forward to obtain the component error prediction value for each intrinsic mode component.

[0009] Optionally, the step of reconstructing the error prediction values ​​of each component to obtain the error trend prediction result specifically includes: The predicted component error values ​​of each intrinsic mode component are arranged in the order in which the intrinsic mode components are separated from the original error time series during the empirical mode decomposition process, thus forming a predicted component sequence. The initial reconstruction error value is obtained by directly adding the error prediction values ​​of all components in the predicted component sequence. The residual components obtained at the end of the empirical mode decomposition process are extended by linear extrapolation to obtain the trend prediction value of the residual components. The trend prediction value of the residual components is added to the preliminary reconstruction error value to obtain the complete error trend prediction result. Calculate the energy percentage of each intrinsic mode component, remove the component error prediction value corresponding to the intrinsic mode component whose energy percentage is lower than the preset energy threshold from the preliminary reconstruction error value, and obtain the reconstruction error value after removing the high-frequency noise component; add the reconstruction error value after removing the high-frequency noise component to the trend prediction value of the remaining component again, and output the filtered error trend prediction result.

[0010] Optionally, the step of obtaining synchronized measured power output error data from adjacent new energy power plants, calculating spatial correlation weights, and obtaining spatial error correction using a spatial weighted average method specifically includes: Taking the current new energy power station to be corrected as the target power station, query all other new energy power stations whose geographical distance from the target power station is less than a set distance threshold from the new energy power station geographic information system, and form a set of adjacent power stations. The measured output data and predicted output data of each adjacent power station in the set of adjacent power stations are obtained from the data acquisition system within the same synchronous time window as the current time, and the measured error sequence of each adjacent power station within the synchronous time window is calculated. Calculate the Pearson correlation coefficient between the historical error sequence of the target power station within the synchronization time window and the measured error sequence of each adjacent power station, and use the Pearson correlation coefficient as the spatial correlation weight between the adjacent power station and the target power station; The spatial error correction amount is obtained by multiplying the measured error value of each adjacent power station at the current time by the corresponding spatial correlation weight, summing the results, and then dividing by the sum of all spatial correlation weights.

[0011] Optionally, the step of adaptively fusing the error trend prediction result with the spatial error correction amount to obtain the final error correction value specifically includes: Obtain the error trend prediction result sequence and spatial error correction sequence for several historical moments in the past, and at the same time obtain the actual error value sequence for several historical moments in the past. Calculate the first root mean square error between the error trend prediction result sequence and the actual error value sequence, and the second root mean square error between the spatial error correction sequence and the actual error value sequence, respectively. The first fusion weight and the second fusion weight are calculated based on the first root mean square error and the second root mean square error, wherein the first fusion weight is inversely proportional to the first root mean square error and the second fusion weight is inversely proportional to the second root mean square error; The error trend prediction result is multiplied by the first fusion weight, the spatial error correction amount is multiplied by the second fusion weight, and the two products are added together to obtain the final error correction value.

[0012] Optionally, it also includes a step of online rolling optimization and updating of the fusion coefficients, specifically including: After each output of the corrected predicted output value, wait for feedback of the actual output data to obtain the true error value at the current moment; The error trend prediction result at the current moment, the spatial error correction amount, and the actual error value are combined to form a new training sample, which is then added to the rolling sample buffer. When the number of samples in the rolling sample buffer reaches the preset window size, the first fusion weight and the second fusion weight are optimized online using the recursive least squares method with the goal of minimizing the sum of squares between the prediction error and the actual error, so as to obtain the optimized first fusion weight and the optimized second fusion weight. The optimized first fusion weight and the optimized second fusion weight are used to replace the original first fusion weight and second fusion weight for error fusion calculation at the next time step.

[0013] Optionally, the step of calculating the Pearson correlation coefficient between the historical error sequence of the target power station within the synchronization time window and the measured error sequence of each adjacent power station, and using the Pearson correlation coefficient as the spatial correlation weight between the adjacent power station and the target power station, specifically includes: Extract the predicted and measured power output values ​​for each moment within the synchronous time window from the historical database of the target power plant, calculate the error value for each moment, and construct the historical error sequence of the target power plant. The measured error values ​​at each moment within the same synchronous time window are extracted from the historical database of each adjacent power station to form an error sequence of adjacent power stations; Calculate the mean and standard deviation of the historical error sequence of the target power station, and calculate the mean and standard deviation of the error sequences of the adjacent power stations; Calculate the covariance between the historical error sequence of the target power station and the error sequence of the adjacent power station, and divide the covariance by the product of the standard deviation of the historical error sequence of the target power station and the standard deviation of the error sequence of the adjacent power station to obtain the Pearson correlation coefficient; The calculated Pearson correlation coefficient is compared with a preset effective correlation threshold. If the Pearson correlation coefficient is lower than the effective correlation threshold, the spatial correlation weight of the adjacent power station is set to zero, that is, the adjacent power station is excluded in the spatial error correction.

[0014] Optionally, it also includes the step of estimating the confidence interval of the corrected power output forecast, specifically including: Collect the final error correction values ​​obtained after each correction within the most recent set time period from historical data to form a correction error sample set; Calculate the mean and standard deviation of the correction error sample set to obtain the statistical distribution parameters of the correction error; using the corrected output prediction value as the center and the multiple of the standard deviation as the interval radius, generate the upper limit and lower limit of the prediction confidence interval corresponding to the confidence level. The upper limit of the predicted confidence interval and the lower limit of the predicted confidence interval are appended to the corrected output prediction value and output for the scheduling system to make uncertainty decisions. Meanwhile, whenever a new true error value is obtained, the corrected error sample set is updated, and the mean and standard deviation are recalculated to achieve online updating of the confidence interval.

[0015] Optionally, the step of calculating the first fusion weight and the second fusion weight based on the first root mean square error and the second root mean square error specifically includes: Obtain the error trend prediction result sequence, spatial error correction sequence, and corresponding true error value sequence for several historical moments in the past; For each prediction result in the error trend prediction result sequence, calculate the absolute value of the deviation between the prediction result and the corresponding true error value, sum the absolute values ​​of the deviations at all historical moments and then divide by the total number of historical moments to obtain the first average deviation magnitude, which serves as an equivalent measure of the first root mean square error. For each correction in the spatial error correction sequence, calculate the absolute value of the deviation between the correction and the corresponding true error value. Sum the absolute values ​​of the deviations at all historical moments and divide by the total number of historical moments to obtain the second average deviation magnitude, which serves as an equivalent measure of the second root mean square error. The ratio of the first fusion weight to the second fusion weight is determined based on the relationship between the first average deviation amplitude and the second average deviation amplitude. The larger the first average deviation amplitude, the smaller the first fusion weight; the larger the second average deviation amplitude, the smaller the second fusion weight. Normalize the first fusion weight and the second fusion weight so that the sum of the first fusion weight and the second fusion weight equals the unit value.

[0016] Optionally, the step of using recursive least squares to optimize the first fusion weight and the second fusion weight online specifically includes: The error trend prediction result of each historical training sample before the current time in the rolling sample buffer is combined with the spatial error correction amount to form an input vector, and the corresponding real error value is used as the target output value. Initialize a weight vector containing two parameters to be optimized, the two parameters to be optimized corresponding to the first fusion weight and the second fusion weight respectively, and simultaneously initialize a covariance matrix; Each training sample in the rolling sample buffer is read sequentially according to time order. For the currently read training sample, the input vector in the training sample is multiplied by the current weight vector to obtain the current prediction error value. Calculate the difference between the current prediction error value and the target output value in the training samples, and multiply the difference by the gain vector calculated based on the covariance matrix and the input vector to obtain the weight adjustment amount; The weight adjustment is accumulated into the current weight vector, and the covariance matrix is ​​updated according to the input vector and the gain vector. After traversing all training samples, the first parameter in the updated weight vector is used as the optimized first fusion weight, and the second parameter is used as the optimized second fusion weight.

[0017] Compared with the prior art, the present invention has the following beneficial effects: This application effectively addresses the core problems of limited accuracy in single-dimensional correction and lack of dynamic adaptability in existing technologies by simultaneously constructing error trend prediction in the time dimension and error correction in the spatial dimension, and adaptively fusing the two. Specifically, by performing multi-scale empirical mode decomposition on the original error time series, the complex non-stationary and nonlinear error series can be separated into multiple intrinsic mode components at different time scales. An autoregressive moving average model is established for prediction based on the characteristics of each component, and the prediction results of each component are reconstructed, significantly improving the accuracy of time dimension error prediction. At the same time, by introducing synchronous measured error data from adjacent new energy power plants, spatial correlation weights are calculated based on the Pearson correlation coefficient, and spatial error correction is obtained using the spatial weighted average method. This effectively captures systematic errors caused by common meteorological factors in the region, making up for the shortcomings of single power plant time prediction. The adaptive fusion mechanism can dynamically adjust the fusion weights according to the historical prediction performance of the two correction methods, ensuring that the optimal correction effect can be obtained under different operating conditions and meteorological conditions.

[0018] Building upon this foundation, the optimal order of the autoregressive moving average model is automatically determined using the Akaike Information Criterion, avoiding the subjectivity and tediousness of manual parameter tuning and improving the efficiency and accuracy of model construction. Setting an effective correlation threshold during spatial correlation calculation automatically excludes adjacent power stations with low error correlation to the target power station, reducing interference from invalid data and further improving the accuracy and computational efficiency of spatial error correction. Online rolling optimization of the fusion weights via recursive least squares enables the model to continuously adapt to changing meteorological conditions and power station operating states, maintaining long-term stable correction performance. Furthermore, by estimating the confidence interval of the corrected output prediction and updating it online, more comprehensive uncertainty information can be provided to the power grid dispatching system, assisting dispatchers in making more scientific and reasonable decisions and effectively improving the safety and economy of power grid operation. Attached Figure Description

[0019] Figure 1 This is a schematic flowchart illustrating the steps of an embodiment of the online correction method for new energy prediction errors according to the present invention. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0021] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0022] like Figure 1 As shown, this method is applicable to online correction of output prediction errors for various new energy power plants such as wind power and photovoltaic power. It can be deployed in the local monitoring system of the new energy power plant or the regional new energy centralized control center, or integrated into the new energy prediction module of the power grid dispatch automation system. The operation cycle of this method can be set to 5 minutes, 15 minutes or 1 hour according to actual needs. A 15-minute operation cycle is recommended to balance correction accuracy and computational resource consumption. Before the system runs, initial configuration needs to be completed, including setting parameters such as geographical distance threshold, effective correlation threshold, energy threshold, rolling sample window size, and confidence level. At the same time, communication connections with the new energy power plant's geographic information system, data acquisition and monitoring control system, and data interaction interfaces with adjacent power plants need to be established.

[0023] First, real-time data acquisition and the generation of the raw error time series are performed. The system acquires the measured output data of the target renewable energy power plant in real time through the data acquisition and monitoring control system. The sampling frequency of the measured output data should be no less than 1Hz. The system performs second-level averaging on the sampled data to obtain minute-level measured output data with the same time resolution as the predicted output data. Simultaneously, the system obtains the predicted output data for the corresponding time moment from the renewable energy prediction module. The time resolution of the predicted output data should be consistent with the measured output data. The system calculates the prediction error value for each time moment, which is equal to the measured output value minus the predicted output value. The system arranges the continuously calculated prediction error values ​​in chronological order to form the raw error time series. The length of the raw error time series should include at least the most recent 24 hours of historical data to ensure the effectiveness of subsequent decomposition and modeling. The system preprocesses the raw error time series, including outlier detection and removal, and missing value imputation. Outlier detection uses the 3σ criterion, identifying error values ​​exceeding the mean ± 3 standard deviations as outliers, and replacing them with linear interpolation of the error values ​​from adjacent time moments. For missing values ​​that occur during data acquisition, linear interpolation is also used to fill them in, ensuring the continuity and integrity of the original error time series.

[0024] Next, empirical mode decomposition (EMD) is performed on the original error time series. The system first identifies all local maxima and local minima in the original error time series. Local maxima are identified using a sliding window method with a window size of three data points. A data point is identified as a local maximum if its value is greater than the values ​​of its two adjacent data points, and as a local minimum if its value is less than the values ​​of its two adjacent data points. The system then uses cubic spline interpolation to fit all local maxima and local minima, obtaining upper and lower envelopes. Cubic spline interpolation ensures that the envelope has continuous first and second derivatives at each interpolation point, making the envelope smoother and avoiding unnecessary fluctuations. The system calculates the average value of the upper and lower envelopes at each time step, obtaining the mean envelope. The system subtracts the mean envelope from the original error time series to obtain the candidate intrinsic mode function components. The system checks whether candidate intrinsic mode function (IMF) components satisfy two defining conditions for IMFs. The first condition is that the number of extreme points is equal to or differs from the number of zero-crossing points by no more than one in the entire data sequence. The second condition is that at any given time, the mean of the upper envelope formed by local maxima and the lower envelope formed by local minima is zero. If a candidate IMF component does not satisfy these conditions, the system treats it as a new original sequence and repeats the envelope fitting and subtraction operations until an IMF component that meets the conditions is obtained. The system separates the first IMF component from the original error time series to obtain the first residual component. The system repeats the envelope fitting, subtraction, and separation operations on the first residual component, sequentially separating multiple IMF components with frequencies from high to low. As the decomposition process progresses, the frequency of the residual components gradually decreases. The decomposition process terminates when the residual component becomes a monotonic function or its amplitude is less than a preset amplitude threshold. The resulting multiple intrinsic mode function components correspond to the fluctuation components at different time scales in the original error time series, and the final residual components represent the long-term trend of the original error time series.

[0025] The system establishes an autoregressive moving average model for each separated intrinsic mode component (EMC) for prediction. For each EMC, the system first uses the Akaike Information Criterion (AIC) to determine the autoregressive order and moving average order of the EMC model. The Akaike AIC automatically selects the optimal model order by balancing the goodness of fit and model complexity, avoiding the subjectivity and overfitting risk of manual order selection. Within a preset order range, the system calculates the Akaike AIC values ​​for different order combinations and selects the order combination that minimizes the Akaike AIC values ​​as the optimal order of the EMC model. The preset order range is typically set to 0 to 10 for both autoregressive and moving average orders, and can be adjusted according to the characteristics of the actual data. After determining the model order, the system uses the least squares method to estimate the parameters of the EMC model. The least squares method obtains the optimal estimate of the model parameters by minimizing the sum of squared errors between the model's predicted and actual values. The system performs significance tests on the estimated model parameters, eliminating insignificant parameter terms to further simplify the model structure and improve its generalization ability. After the model parameters are estimated, the system uses the established autoregressive moving average model to predict one step forward, obtaining the component error prediction value corresponding to that intrinsic mode component. The system repeats the above modeling and prediction process for all intrinsic mode components to obtain the component error prediction value for each intrinsic mode component.

[0026] The system reconstructs the error prediction values ​​of each component to obtain the error trend prediction result. First, the system arranges the component error prediction values ​​of each intrinsic mode component according to the order in which they were separated from the original error time series during the empirical mode decomposition process, forming a predicted component sequence. Components separated earlier correspond to higher frequency components, and components separated later correspond to lower frequency components. The system directly adds all the component error prediction values ​​in the predicted component sequence to obtain the preliminary reconstruction error value. Simultaneously, the system uses linear extrapolation to extend the trend of the remaining components obtained at the end of the empirical mode decomposition process, obtaining the trend prediction value of the remaining components. The linear extrapolation method predicts based on the linear trend of the remaining components and is suitable for predicting the remaining components that represent long-term trends. The system adds the trend prediction value of the remaining components to the preliminary reconstruction error value to obtain the complete error trend prediction result. To further improve the accuracy of the error trend prediction, the system also performs filtering processing on the reconstruction process. The system calculates the energy proportion of each intrinsic mode component. The energy of an intrinsic mode component is equal to the sum of the squares of all data points of that component, and the energy proportion is equal to the energy of that component divided by the total energy of all intrinsic mode components. The system identifies intrinsic mode components with an energy percentage below a preset energy threshold as high-frequency noise components and removes their corresponding component error prediction values ​​from the initial reconstruction error value, resulting in a reconstruction error value after removing the high-frequency noise components. The preset energy threshold is typically set between 0.01 and 0.05 and can be adjusted according to the actual noise level. The system then adds the reconstruction error value after removing the high-frequency noise components to the trend prediction values ​​of the remaining components, outputting the filtered error trend prediction result. This filtering process effectively removes high-frequency random noise from the original error time series, improving the signal-to-noise ratio of the error trend prediction.

[0027] While predicting the temporal error trend, the system simultaneously calculates the spatial error correction. Using the current renewable energy power station to be corrected as the target station, the system queries the renewable energy power station geographic information system for all other renewable energy power stations whose geographical distance to the target station is less than a set distance threshold, forming a set of adjacent power stations. The set distance threshold is typically between 50km and 100km and can be adjusted based on the uniformity of regional meteorological conditions. In plains areas with relatively uniform meteorological conditions, the distance threshold can be appropriately increased; in mountainous areas with complex terrain, the distance threshold should be appropriately decreased. The system obtains the measured and predicted power output data of each adjacent power station in the adjacent power station set from the data acquisition system within the same synchronous time window as the current time. The length of the synchronous time window is typically set between 24 and 72 hours to ensure accurate calculation of the spatial correlation between adjacent power stations and the target power station. The system calculates the measured error value of each adjacent power station at each moment within the synchronous time window. The measured error value equals the measured power output value of the adjacent power station minus its predicted power output value, forming the measured error sequence for each adjacent power station.

[0028] The system calculates the Pearson correlation coefficient between the historical error sequence of the target power station within the synchronization time window and the measured error sequence of each adjacent power station, using the Pearson correlation coefficient as the spatial correlation weight between the adjacent power station and the target power station. First, the system extracts the predicted and measured power output values ​​for each moment within the synchronization time window from the historical database of the target power station, calculates the error value for each moment, and constructs the historical error sequence of the target power station. Then, the system extracts the measured error values ​​for each moment within the same synchronization time window from the historical database of each adjacent power station, constructing the error sequences of the adjacent power stations. The system calculates the mean and standard deviation of the historical error sequence of the target power station, and the mean and standard deviation of the error sequences of the adjacent power stations. Finally, the system calculates the covariance between the historical error sequence of the target power station and the error sequences of the adjacent power stations, reflecting the degree of linear correlation between the two sequences. The system divides the covariance by the product of the standard deviations of the historical error sequence of the target power station and the error sequences of the adjacent power stations to obtain the Pearson correlation coefficient. The Pearson correlation coefficient ranges from -1 to 1. A value closer to 1 indicates a stronger linear correlation between the two sequences, while a value closer to 0 indicates a weaker linear correlation. The system compares the calculated Pearson correlation coefficient with a preset effective correlation threshold. If the Pearson correlation coefficient is lower than the effective threshold, the spatial correlation weight of the adjacent power station is set to zero, effectively excluding it from the spatial error correction. The preset effective correlation threshold is typically set between 0.3 and 0.5 and can be adjusted based on actual conditions. By setting the effective correlation threshold, adjacent power stations with low error correlation to the target power station can be automatically excluded, reducing the interference of invalid data on the spatial error correction results and lowering the computational load. The system multiplies the current measured error value of each adjacent power station by its corresponding spatial correlation weight, sums the results, and then divides by the sum of all spatial correlation weights to obtain the spatial error correction amount. If the spatial correlation weights of all adjacent power stations are zero, the spatial error correction amount is set to zero, and only the error trend prediction results in the time dimension are used for correction.

[0029] The system adaptively fuses the error trend prediction results with the spatial error correction amount to obtain the final error correction value. The system first acquires the error trend prediction result sequence and the spatial error correction amount sequence for several historical time points, and simultaneously acquires the actual error value sequence for several historical time points. The number of historical time points is typically set between 24 and 168, corresponding to historical data from the past 1 to 7 days. The system calculates the first root mean square error (RMSE) of the error trend prediction result sequence and the actual error value sequence, and the second RMSE of the spatial error correction amount sequence and the actual error value sequence, respectively. The RMSE comprehensively reflects the degree of deviation between the predicted value and the actual value and is a commonly used indicator for measuring prediction accuracy. The system calculates the first fusion weight and the second fusion weight based on the first and second RMSEs, where the first fusion weight is inversely proportional to the first RMSE, and the second fusion weight is inversely proportional to the second RMSE. This weight calculation method allows correction methods with higher prediction accuracy to receive larger fusion weights, and correction methods with lower prediction accuracy to receive smaller fusion weights, thus achieving complementary advantages between the two correction methods. The system multiplies the error trend prediction result by the first fusion weight, multiplies the spatial error correction by the second fusion weight, and then adds the two products to obtain the final error correction value.

[0030] As an optional implementation, the system can also use the average deviation magnitude as an equivalent measure of prediction accuracy to calculate the fusion weights. For each prediction result in the error trend prediction result sequence, the system calculates the absolute value of the deviation between the prediction result and the corresponding true error value. The system then sums the absolute values ​​of the deviations at all historical time points and divides this sum by the total number of historical time points to obtain the first average deviation magnitude. For each correction amount in the spatial error correction amount sequence, the system calculates the absolute value of the deviation between the correction amount and the corresponding true error value. The system then sums the absolute values ​​of the deviations at all historical time points and divides this sum by the total number of historical time points to obtain the second average deviation magnitude. The system determines the ratio of the first fusion weight to the second fusion weight based on the relationship between the first and second average deviation magnitudes; a larger first average deviation magnitude results in a smaller first fusion weight, and vice versa. The system normalizes the first and second fusion weights so that their sum equals 1. The average deviation magnitude is less sensitive to outliers than the root mean square error. In cases with many outliers, using the average deviation magnitude to calculate the fusion weights yields more stable fusion results.

[0031] The system also includes a step of online rolling optimization and updating of the fusion coefficients. After each output of the corrected power prediction value, the system waits for feedback from the actual power data. Once the measured power data for the current moment is obtained, the system calculates the true error value for that moment. The system combines the error trend prediction result, the spatial error correction, and the true error value for the current moment into a new training sample and adds it to the rolling sample buffer. The rolling sample buffer is managed using a first-in, first-out (FIFO) mechanism. When the number of samples in the buffer reaches the preset window size, the newly added sample replaces the oldest sample. The preset window size is usually set between 100 and 1000 and can be adjusted according to computing resources and data characteristics. When the number of samples in the rolling sample buffer reaches the preset window size, the system optimizes the first and second fusion weights online using the recursive least squares method, aiming to minimize the sum of squares between the prediction error and the true error. The recursive least squares method is an online parameter estimation algorithm that can continuously update model parameters based on new sample data without reprocessing all historical data, offering advantages such as low computational cost and high real-time performance.

[0032] The specific process of the system's recursive least squares method for fusion weight optimization is as follows: The system combines the error trend prediction results from each historical training sample before the current time step in the rolling sample buffer with the spatial error correction amount to form an input vector, and uses the corresponding true error value as the target output value. The system initializes a weight vector containing two parameters to be optimized, corresponding to the first fusion weight and the second fusion weight, respectively. The initial weight vector can be set to [0.5, 0.5], meaning that initially, the two correction methods have the same weight. The system also initializes a covariance matrix, typically a large diagonal matrix, such as 1000 multiplied by the identity matrix, to ensure rapid convergence of the parameters in the initial stage. The system sequentially reads each training sample in the rolling sample buffer. For the currently read training sample, the system performs a dot product operation between the input vector in the training sample and the current weight vector to obtain the current prediction error value. The system calculates the difference between the current prediction error value and the target output value in the training sample, and multiplies this difference by the gain vector calculated based on the covariance matrix and the input vector to obtain the weight adjustment amount. The system accumulates the weight adjustments to the current weight vector and updates the covariance matrix based on the input and gain vectors. After traversing all training samples, the system uses the first parameter of the updated weight vector as the optimized first fusion weight and the second parameter as the optimized second fusion weight. The system then replaces the original first and second fusion weights with the optimized first and second fusion weights for error fusion calculation in the next time step. Through online rolling optimization, the fusion weights can continuously adapt to changing weather conditions and power plant operating states, ensuring the system maintains optimal correction performance at all times.

[0033] The system also includes a step of estimating the confidence interval for the corrected output forecast. The system collects the final error correction values ​​obtained after each correction within the most recent set period from historical data, forming a correction error sample set. The most recent set period is typically between 7 and 30 days to ensure a sufficient and representative sample size. The system calculates the mean and standard deviation of the correction error sample set to obtain the statistical distribution parameters of the correction error. According to the Central Limit Theorem, when the sample size is sufficiently large, the correction error approximately follows a normal distribution. The system generates upper and lower confidence intervals corresponding to different confidence levels, centered on the corrected output forecast and using multiples of the standard deviation as the interval radius. Commonly used confidence levels are 95%, corresponding to a standard deviation multiple of 1.96; and 90%, corresponding to a standard deviation multiple of 1.645. The system appends the upper and lower confidence intervals to the corrected output forecast and outputs them for the scheduling system to use as a reference for uncertainty decision-making. Dispatchers can assess the uncertainty of forecast results based on the size of the confidence interval and reserve corresponding reserve capacity when formulating dispatch plans, thereby improving the safety of power grid operation. Simultaneously, whenever a new true error value is obtained, the system updates the corrected error sample set, adding the new true error value to the sample set and removing the oldest sample, then recalculating the mean and standard deviation, achieving online updates to the confidence interval. Online updates ensure that the confidence interval always reflects the latest corrected error distribution characteristics, improving the accuracy of confidence interval estimation.

[0034] The system overlays the final error correction value onto the original predicted power output data to obtain the corrected predicted power output value. The corrected predicted power output value equals the original predicted power output value plus the final error correction value. The system outputs the corrected predicted power output value and the corresponding prediction confidence interval for use by renewable energy power plant operators or grid dispatchers. The system stores all data generated during each operation, including the original error time series, intrinsic mode components, component error prediction values, error trend prediction results, neighboring power plant error series, spatial correlation weights, spatial error correction amounts, fusion weights, final error correction values, corrected predicted power output values, and actual error values, in a historical database for subsequent model evaluation, parameter optimization, and data analysis. The system periodically evaluates the correction effect, calculates the corrected prediction accuracy indicators, including mean absolute error, root mean square error, and mean absolute percentage error, and compares them with the prediction accuracy before correction. If a significant decrease in the correction effect is detected, the system automatically readjusts various parameters, including the geographical distance threshold, effective correlation threshold, energy threshold, and rolling sample window size, to restore the system's correction performance.

[0035] This method effectively addresses the limited accuracy of single-dimensional correction by simultaneously utilizing information from both time and spatial dimensions to correct new energy forecasting errors. Empirical Mode Decomposition (EMD) adaptively decomposes complex non-stationary error sequences into multiple stationary components at different time scales, and combined with an autoregressive moving average model, it fully explores the temporal evolution of errors. Spatial correlation analysis effectively captures systematic errors caused by shared meteorological factors within the region by utilizing measured error information from adjacent power plants. An adaptive fusion mechanism dynamically adjusts the fusion weights based on the real-time performance of the two correction methods, achieving complementary advantages. Online rolling optimization further enhances the system's adaptability and robustness. Confidence interval estimation provides more comprehensive information support for scheduling decisions. This method boasts advantages such as high correction accuracy, strong adaptability, and good real-time performance, significantly improving the accuracy of new energy output forecasting and providing strong technical support for large-scale grid connection and consumption of new energy.

[0036] As is known from common technical knowledge, this invention can be implemented through other embodiments that do not depart from its spirit or essential characteristics. Therefore, the disclosed embodiments described above are merely illustrative in all respects and are not the only ones. All modifications within the scope of this invention or its equivalents are included in this invention.

Claims

1. A method for online correction of new energy prediction errors, characterized in that, Includes the following steps: Real-time acquisition of measured power output data and corresponding predicted power output data of new energy power plants; calculation of prediction error sequence; and obtaining original error time series. Empirical mode decomposition is performed on the original error time series to separate multiple intrinsic mode components at different time scales. An autoregressive moving average model is then established for each intrinsic mode component to obtain the component error prediction value. The error prediction values ​​of each component are reconstructed to obtain the error trend prediction result; Synchronous measured power output error data are obtained from adjacent new energy power plants, spatial correlation weights are calculated, and spatial error correction is obtained using the spatial weighted average method. The error trend prediction result is adaptively fused with the spatial error correction amount to obtain the final error correction value, and the final error correction value is superimposed on the original predicted output data to output the corrected output prediction value.

2. The online correction method for new energy prediction errors according to claim 1, characterized in that, The steps of performing empirical mode decomposition on the original error time series to separate multiple intrinsic mode components at different time scales, and establishing an autoregressive moving average model for each intrinsic mode component to obtain the component error prediction value, specifically include: The local maxima and local minima of the original error time series are fitted by cubic spline interpolation to obtain the upper and lower envelopes. The mean envelopes of the upper and lower envelopes are calculated, and the candidate intrinsic mode function components are obtained by subtracting the mean envelope from the original error time series. Repeatedly perform envelope fitting and subtraction operations until the candidate intrinsic mode function component satisfies the definition condition of intrinsic mode function, separate an intrinsic mode function component from the original error time series and obtain the remaining component; Repeat the envelope fitting, subtraction and separation operations on the remaining components to separate multiple intrinsic mode function components with frequencies from high to low, until the remaining components are monotonic functions or the amplitude is less than the preset threshold, thus obtaining multiple intrinsic mode components at different time scales. For each intrinsic mode component, the autoregressive order and moving average order of the autoregressive moving average model are determined using the Akaike information criterion. After estimating the model parameters, the model is predicted one step forward to obtain the component error prediction value for each intrinsic mode component.

3. The online correction method for new energy prediction errors according to claim 2, characterized in that, The step of reconstructing the error prediction values ​​of each component to obtain the error trend prediction result specifically includes: The predicted component error values ​​of each intrinsic mode component are arranged in the order in which the intrinsic mode components are separated from the original error time series during the empirical mode decomposition process, thus forming a predicted component sequence. The initial reconstruction error value is obtained by directly adding the error prediction values ​​of all components in the predicted component sequence. The residual components obtained at the end of the empirical mode decomposition process are extended by linear extrapolation to obtain the trend prediction value of the residual components. The trend prediction value of the residual components is added to the preliminary reconstruction error value to obtain the complete error trend prediction result. Calculate the energy percentage of each intrinsic mode component, remove the component error prediction value corresponding to the intrinsic mode component whose energy percentage is lower than the preset energy threshold from the preliminary reconstruction error value, and obtain the reconstruction error value after removing the high-frequency noise component; add the reconstruction error value after removing the high-frequency noise component to the trend prediction value of the remaining component again, and output the filtered error trend prediction result.

4. The online correction method for new energy prediction errors according to claim 1, characterized in that, The steps of obtaining synchronous measured power output error data from adjacent new energy power plants, calculating spatial correlation weights, and obtaining spatial error correction amounts using a spatial weighted average method specifically include: Taking the current new energy power station to be corrected as the target power station, query all other new energy power stations whose geographical distance from the target power station is less than a set distance threshold from the new energy power station geographic information system, and form a set of adjacent power stations. The measured output data and predicted output data of each adjacent power station in the set of adjacent power stations are obtained from the data acquisition system within the same synchronous time window as the current time, and the measured error sequence of each adjacent power station within the synchronous time window is calculated. Calculate the Pearson correlation coefficient between the historical error sequence of the target power station within the synchronization time window and the measured error sequence of each adjacent power station, and use the Pearson correlation coefficient as the spatial correlation weight between the adjacent power station and the target power station; The spatial error correction amount is obtained by multiplying the measured error value of each adjacent power station at the current time by the corresponding spatial correlation weight, summing the results, and then dividing by the sum of all spatial correlation weights.

5. The online correction method for new energy prediction errors according to claim 1, characterized in that, The step of adaptively fusing the error trend prediction result with the spatial error correction amount to obtain the final error correction value specifically includes: Obtain the error trend prediction result sequence and spatial error correction sequence for several historical moments in the past, and at the same time obtain the actual error value sequence for several historical moments in the past. Calculate the first root mean square error between the error trend prediction result sequence and the actual error value sequence, and the second root mean square error between the spatial error correction sequence and the actual error value sequence, respectively. The first fusion weight and the second fusion weight are calculated based on the first root mean square error and the second root mean square error, wherein the first fusion weight is inversely proportional to the first root mean square error and the second fusion weight is inversely proportional to the second root mean square error; The error trend prediction result is multiplied by the first fusion weight, the spatial error correction amount is multiplied by the second fusion weight, and the two products are added together to obtain the final error correction value.

6. The online correction method for new energy prediction errors according to claim 1, characterized in that, It also includes the step of online rolling optimization and updating of the fusion coefficients, specifically including: After each output of the corrected predicted output value, wait for feedback of the actual output data to obtain the true error value at the current moment; The current error trend prediction result, spatial error correction amount, and the actual error value are combined to form a new training sample, which is then added to the rolling sample buffer. When the number of samples in the rolling sample buffer reaches the preset window size, the first fusion weight and the second fusion weight are optimized online using the recursive least squares method with the goal of minimizing the sum of squares between the prediction error and the actual error, so as to obtain the optimized first fusion weight and the optimized second fusion weight. The optimized first fusion weight and the optimized second fusion weight are used to replace the original first fusion weight and second fusion weight for error fusion calculation at the next time step.

7. The online correction method for new energy prediction errors according to claim 4, characterized in that, The step of calculating the Pearson correlation coefficient between the historical error sequence of the target power station within the synchronization time window and the measured error sequence of each adjacent power station, and using the Pearson correlation coefficient as the spatial correlation weight between the adjacent power station and the target power station, specifically includes: Extract the predicted and measured power output values ​​for each moment within the synchronous time window from the historical database of the target power plant, calculate the error value for each moment, and construct the historical error sequence of the target power plant. The measured error values ​​at each moment within the same synchronous time window are extracted from the historical database of each adjacent power station to form an error sequence of adjacent power stations; Calculate the mean and standard deviation of the historical error sequence of the target power station, and calculate the mean and standard deviation of the error sequences of the adjacent power stations; Calculate the covariance between the historical error sequence of the target power station and the error sequence of the adjacent power station, and divide the covariance by the product of the standard deviation of the historical error sequence of the target power station and the standard deviation of the error sequence of the adjacent power station to obtain the Pearson correlation coefficient; The calculated Pearson correlation coefficient is compared with a preset effective correlation threshold. If the Pearson correlation coefficient is lower than the effective correlation threshold, the spatial correlation weight of the adjacent power station is set to zero, that is, the adjacent power station is excluded in the spatial error correction.

8. The online correction method for new energy prediction errors according to claim 1, characterized in that, It also includes the step of estimating the confidence interval of the corrected power output forecast, specifically including: Collect the final error correction values ​​obtained after each correction within the most recent set time period from historical data to form a correction error sample set; Calculate the mean and standard deviation of the correction error sample set to obtain the statistical distribution parameters of the correction error; using the corrected output prediction value as the center and the multiple of the standard deviation as the interval radius, generate the upper limit and lower limit of the prediction confidence interval corresponding to the confidence level. The upper limit of the predicted confidence interval and the lower limit of the predicted confidence interval are appended to the corrected output prediction value and output for the scheduling system to make uncertainty decisions. Meanwhile, whenever a new true error value is obtained, the corrected error sample set is updated, and the mean and standard deviation are recalculated to achieve online updating of the confidence interval.

9. The online correction method for new energy prediction errors according to claim 5, characterized in that, The steps for calculating the first fusion weight and the second fusion weight based on the first root mean square error and the second root mean square error specifically include: Obtain the error trend prediction result sequence, spatial error correction sequence, and corresponding true error value sequence for several historical moments in the past; For each prediction result in the error trend prediction result sequence, calculate the absolute value of the deviation between the prediction result and the corresponding true error value, sum the absolute values ​​of the deviations at all historical moments and then divide by the total number of historical moments to obtain the first average deviation magnitude, which serves as an equivalent measure of the first root mean square error. For each correction in the spatial error correction sequence, calculate the absolute value of the deviation between the correction and the corresponding true error value. Sum the absolute values ​​of the deviations at all historical moments and divide by the total number of historical moments to obtain the second average deviation magnitude, which serves as an equivalent measure of the second root mean square error. The ratio of the first fusion weight to the second fusion weight is determined based on the relationship between the magnitude of the first average deviation and the magnitude of the second average deviation. The larger the magnitude of the first average deviation, the smaller the first fusion weight, and the larger the magnitude of the second average deviation, the smaller the second fusion weight. Normalize the first fusion weight and the second fusion weight so that the sum of the first fusion weight and the second fusion weight equals the unit value.

10. The online correction method for new energy prediction errors according to claim 6, characterized in that, The step of using the recursive least squares method to optimize the first fusion weight and the second fusion weight online specifically includes: The error trend prediction result of each historical training sample before the current time in the rolling sample buffer is combined with the spatial error correction amount to form an input vector, and the corresponding real error value is used as the target output value. Initialize a weight vector containing two parameters to be optimized, the two parameters to be optimized corresponding to the first fusion weight and the second fusion weight respectively, and simultaneously initialize a covariance matrix; Each training sample in the rolling sample buffer is read sequentially according to time order. For the currently read training sample, the input vector in the training sample is multiplied by the current weight vector to obtain the current prediction error value. Calculate the difference between the current prediction error value and the target output value in the training samples, and multiply the difference by the gain vector calculated based on the covariance matrix and the input vector to obtain the weight adjustment amount; The weight adjustment is accumulated into the current weight vector, and the covariance matrix is ​​updated according to the input vector and the gain vector. After traversing all training samples, the first parameter in the updated weight vector is used as the optimized first fusion weight, and the second parameter is used as the optimized second fusion weight.

Images