A method and system for correcting power prediction errors of non-stationary wind power data

Abstract

The application discloses a non-stationary wind power data power prediction error correction method and system, relates to the wind power prediction technical field, and comprises the following steps: acquiring meteorological monitoring data, historical power data and a predicted power sequence; performing time sequence change analysis to form a plurality of time evolution segments; extracting segment evolution descriptors, constructing a non-stationary evolution track sequence; acquiring a prediction error sequence, establishing a coupling relationship between the prediction error sequence and the non-stationary evolution track sequence, performing prediction error change characteristic classification analysis corresponding to different non-stationary evolution track sequences, and constructing an error evolution mode set; at a new prediction time, performing error evolution mode set similarity matching according to the non-stationary evolution track of a current node, and performing error correction management by using a matching result. The application solves the technical problem that wind power prediction error is difficult to correct in time in the prior art, and achieves the technical effect of improving wind power prediction error correction accuracy.

Description

Technical Field

[0001] This invention relates to the field of wind power prediction technology, specifically to a method and system for correcting power prediction errors in non-stationary wind power data. Background Technology

[0002] Wind power forecasting typically relies on meteorological monitoring data and historical power data to build predictive models. Statistical analysis or machine learning modeling of historical operating data is used to predict future power output. However, meteorological factors such as wind speed and direction exhibit significant randomness and volatility during wind farm operation, resulting in strong non-stationary changes in wind power operating conditions over time. When wind power operating conditions change rapidly, predictive models based on relatively stable historical data struggle to reflect these dynamic changes in a timely manner, leading to significant discrepancies between predicted and actual power outputs. Furthermore, prediction errors are difficult to identify and effectively correct in a timely manner. Summary of the Invention

[0003] This application provides a method and system for correcting power prediction errors for non-stationary wind power data, which addresses the technical problem in the prior art where wind power prediction is difficult to adapt to non-stationary changes in wind power operating conditions, resulting in difficulty in timely correction of prediction errors.

[0004] In view of the above problems, this application provides a method and system for correcting power prediction errors of non-stationary wind power data.

[0005] A first aspect of this application provides a method for correcting power prediction errors for non-stationary wind power data, the method comprising:

[0006] Acquire meteorological monitoring data, historical power data, and predicted power sequences output by power prediction models for wind farms within a continuous time series. Perform time-series variation analysis on the meteorological monitoring data and historical power data, dynamically segmenting the continuous time series according to wind speed change gradient, wind direction shift rate, and power fluctuation intensity to form multiple time evolution segments reflecting local operating states. Extract segment evolution descriptors representing the change direction, change rate, and fluctuation energy distribution of each time evolution segment, and construct a non-stationary evolution trajectory sequence of the wind farm operating state in chronological order. Compare and analyze the predicted power sequence with the actual power sequence at the corresponding time to obtain the prediction error sequence, and establish a coupling relationship between the prediction error sequence and the non-stationary evolution trajectory sequence. Using the coupling relationship, perform a classification analysis of the prediction error change characteristics corresponding to different non-stationary evolution trajectory sequences to construct an error evolution pattern set. At a new prediction time, perform similarity matching of the error evolution pattern set based on the non-stationary evolution trajectory of the current node, and use the similarity matching results to perform error correction management.

[0007] A second aspect of this application provides a power prediction error correction system for non-stationary wind power data, the system comprising:

[0008] The data acquisition module acquires meteorological monitoring data, historical power data, and predicted power sequences output by power prediction models for wind farms over a continuous time series. The dynamic segmentation module performs time-series variation analysis on the meteorological monitoring data and historical power data, dynamically segmenting the continuous time series based on wind speed change gradients, wind direction shift rates, and power fluctuation intensity to form multiple time evolution segments reflecting local operating states. The trajectory sequence construction module extracts segment evolution descriptors representing the direction, rate, and energy distribution of change in each time evolution segment, and constructs a non-uniform trajectory sequence representing the wind farm's operating state in chronological order. The system comprises: a stationary evolution trajectory sequence; a comparison and analysis module, used to compare and analyze the predicted power sequence with the actual power sequence at the corresponding time, obtain the prediction error sequence, and establish the coupling relationship between the prediction error sequence and the non-stationary evolution trajectory sequence; a classification and analysis module, used to use the coupling relationship to perform classification and analysis of the prediction error change characteristics corresponding to different non-stationary evolution trajectory sequences, and construct an error evolution pattern set; and an error correction management module, used to perform similarity matching of the error evolution pattern set according to the non-stationary evolution trajectory of the current node at a new prediction time, and perform error correction management using the similarity matching results.

[0009] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0010] This application acquires meteorological monitoring data, historical power data, and predicted power sequences output by power prediction models for wind farms within a continuous time series. It performs time-series change analysis on the meteorological monitoring data and historical power data, dynamically segmenting the continuous time series based on wind speed change gradient, wind direction shift rate, and power fluctuation intensity to form multiple time evolution segments reflecting local operating states. It extracts segment evolution descriptors representing the change direction, change rate, and fluctuation energy distribution of each time evolution segment, and constructs a non-stationary evolution trajectory sequence of the wind farm's operating state in chronological order. It compares and analyzes the predicted power sequence with the actual power sequence at the corresponding time point to obtain a prediction error sequence, and establishes a coupling relationship between the prediction error sequence and the non-stationary evolution trajectory sequence. Using this coupling relationship, it performs a classification analysis of the prediction error change characteristics corresponding to different non-stationary evolution trajectory sequences, constructing a set of error evolution patterns. At a new prediction time, it performs similarity matching of the error evolution pattern set based on the non-stationary evolution trajectory of the current node, and uses the similarity matching results to perform error correction management. This invention addresses the technical problem in existing technologies where wind power prediction is difficult to adapt to non-stationary changes in wind power operating conditions, leading to difficulties in timely correction of prediction errors. By constructing a non-stationary evolution trajectory of wind farm operating conditions and establishing a coupling relationship between prediction errors and operating conditions, the invention achieves the technical effect of improving the accuracy of wind power prediction error correction. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 A schematic flowchart of a power prediction error correction method for non-stationary wind power data provided in this application embodiment;

[0013] Figure 2 This is a schematic diagram of a power prediction error correction system for non-stationary wind power data provided in an embodiment of this application.

[0014] Figure labeling: Data acquisition module 11, dynamic segmentation module 12, trajectory sequence construction module 13, comparative analysis module 14, classification analysis module 15, error correction management module 16. Detailed Implementation

[0015] This application provides a method and system for correcting power prediction errors for non-stationary wind power data. It addresses the technical problem in the prior art where wind power prediction is difficult to adapt to non-stationary changes in wind power operating conditions, leading to difficulties in timely correction of prediction errors. By constructing the non-stationary evolution trajectory of the wind farm operating conditions and establishing a coupling relationship between prediction errors and operating conditions, the technical effect of improving the accuracy of wind power prediction error correction is achieved.

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0017] It should be noted that any variation of the terms "comprising" and "having" is intended to cover non-exclusive inclusion, for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such processes, methods, products, or devices.

[0018] Example 1, as Figure 1 As shown, this application provides a method for correcting power prediction errors for non-stationary wind power data, the method comprising:

[0019] Step S100: Obtain meteorological monitoring data, historical power data, and predicted power sequence output by the power prediction model for the wind farm in a continuous time series.

[0020] In this embodiment, meteorological monitoring data of the wind farm over a continuous time series is first acquired. Meteorological parameters such as wind speed, wind direction, temperature, and air pressure are collected according to a preset sampling period using wind measuring devices and environmental monitoring devices in the wind farm. Corresponding time stamps are added to the meteorological parameters at each sampling time. The meteorological monitoring data refers to the data on changes in the meteorological state of the wind farm recorded in chronological order. After the data collection is completed, the meteorological parameters corresponding to the same sampling time are organized and correlated to obtain the continuous time-series meteorological monitoring data.

[0021] Next, historical power data of the wind farm within a continuous time series is acquired. The actual output power values ​​corresponding to each sampling time are read from the wind farm's power metering or monitoring devices and processed according to a sampling period consistent with the meteorological monitoring data. Historical power data refers to the actual power generation records of the wind farm at each sampling time. Then, the historical power data is correlated with the meteorological monitoring data using the same time marker, ensuring that the meteorological monitoring data at each sampling time corresponds to the historical power data at the same time, thus forming the sample data required for model training.

[0022] After obtaining meteorological monitoring data and historical power data, the power prediction model is pre-trained. This model is a power prediction calculation model built based on historical meteorological and power variations. Specifically, historical meteorological monitoring data from multiple consecutive sampling times are used as model input, and historical power data for the corresponding prediction time are used as model output. Training samples allow the power prediction model to learn the mapping relationship between meteorological changes and power output. The training data includes wind speed, wind direction, temperature, air pressure, and historical power data for the corresponding time. The power prediction model can employ a Long Short-Term Memory (LSTM) network model with the following parameters: input time step of 24, 2 hidden layers with 256 units per hidden layer, learning rate of 0.001, batch size of 64, training epochs of 100, Dropout parameter of 0.2, and output layer dimension of 1. This enables the power prediction model to predict wind power output.

[0023] After the power prediction model is trained, the real-time meteorological monitoring data collected from the wind farm is input into the trained power prediction model. The power prediction model outputs the power prediction value at the corresponding prediction time and forms a predicted power sequence in chronological order. The predicted power sequence refers to the predicted power results at multiple prediction times output by the power prediction model based on the real-time meteorological monitoring data.

[0024] Step S200: Perform time-series change analysis on the meteorological monitoring data and historical power data, and dynamically segment the continuous time series according to the wind speed change gradient, wind direction shift rate, and power fluctuation intensity to form multiple time evolution segments reflecting the local operating state.

[0025] In this embodiment, when performing time-series change analysis on meteorological monitoring data and historical power data, the meteorological monitoring data and historical power data are organized according to the same sampling time, so that each sampling time corresponds to wind speed, wind direction, and actual power data, and formed into continuous time data in chronological order. Here, continuous time series refers to data records arranged continuously in chronological order, used to reflect the change process of wind farm operating status over time. After completing the organization, the data changes between adjacent sampling times are used as the basis for analysis, thereby ensuring that subsequent change results can be mapped to specific time positions.

[0026] Next, based on continuous-time data, hourly calculations are performed on the wind speed data to obtain the wind speed change gradient. The wind speed change gradient is the ratio of the wind speed difference between two adjacent sampling moments to the sampling time interval, representing the magnitude of wind speed change per unit time. During the calculation, wind speed values ​​at two adjacent sampling moments are read sequentially, the difference between them is calculated, and then divided by the corresponding sampling time interval to obtain the wind speed change gradient corresponding to the current sampling moment. The wind speed change results for each sampling moment are then generated in chronological order to reflect the rate of wind speed change over continuous time.

[0027] Based on continuous-time data, wind direction data is calculated hourly to obtain the wind direction shift rate. The wind direction shift rate is the ratio of the change in wind direction between two adjacent sampling times to the sampling time interval, representing the speed of wind direction change. During the calculation, wind direction values ​​at two adjacent sampling times are read sequentially, and the difference between them is calculated. Considering that wind direction data is angular, an angle difference correction method is used when calculating the difference. When the difference is greater than 180°, 360° is subtracted from the difference to obtain the actual change; when the difference is less than or equal to 180°, the difference is directly used as the actual change. Then, the actual change is divided by the sampling time interval to obtain the wind direction shift rate at the corresponding sampling time, and the wind direction change results are generated in chronological order to reflect the shift of wind direction over continuous time.

[0028] Based on continuous-time data, historical power data is calculated hourly to obtain the power fluctuation intensity. The power fluctuation intensity is the ratio of the absolute value of the power difference between two adjacent sampling moments to the sampling time interval, representing the strength of power output changes. During the calculation, the actual power values ​​of two adjacent sampling moments are read sequentially, the absolute value of the difference is calculated, and then divided by the corresponding sampling time interval to obtain the power fluctuation intensity corresponding to the current sampling moment. The power change results are then presented in chronological order to reflect the continuous-time fluctuations in wind farm power output.

[0029] After obtaining the wind speed change gradient, wind direction shift rate, and power fluctuation intensity, the three results are integrated to ensure that each sampling moment is associated with a corresponding wind speed change gradient, wind direction shift rate, and power fluctuation intensity. Subsequently, the state changes in continuous-time data are assessed based on these three results. In this process, thresholds for wind speed change gradient, wind direction shift rate, and power fluctuation intensity are pre-set. Then, the three results for each sampling moment are assessed sequentially. If the wind speed change gradient, wind direction shift rate, or power fluctuation intensity at a given sampling moment reaches its corresponding threshold, a significant state change is determined to have occurred before or after that sampling moment. If none of the three results reach their respective thresholds, the operating state is considered relatively continuous during that period.

[0030] After completing the above judgment, the detected state change locations are used as the basis for time segmentation to dynamically divide the continuous time series. Dynamic segmentation refers to dividing the continuous time data based on the actual detected change locations, so that the time segmentation results change with the operating state. Specifically, continuous data between two adjacent state change locations is defined as the same time range. When the wind speed gradient, wind direction shift rate, and power fluctuation intensity do not show significant changes over multiple consecutive sampling times, this time range is maintained continuously. When a new significant change location appears, the current time range ends, and the next time range begins from that change location. Through this processing method, each divided time range has relatively similar change characteristics within its boundaries, while different time ranges exhibit more significant differences in change.

[0031] After completing the dynamic segmentation, the meteorological monitoring data and historical power data corresponding to each time range are organized and saved separately, forming multiple time evolution segments reflecting the local operating status. Among them, the time evolution segment refers to the data set within a local time range obtained from the dynamic segmentation, and each time evolution segment corresponds to the operational changes of the wind farm over a continuous period of time.

[0032] Step S300: Extract segment evolution descriptors representing the direction of change, rate of change, and distribution of fluctuation energy for each time evolution segment, and construct a non-stationary evolution trajectory sequence of the wind field operating state in chronological order.

[0033] In this embodiment, when extracting segment evolution descriptors representing the direction, rate, and energy distribution of change for each time evolution segment, the wind speed, wind direction, and power data in each time evolution segment are read sequentially, and the start time, end time, and data values ​​at each sampling time within the segment are determined respectively. For the direction of change, the difference between the wind speed value at the end time and the wind speed value at the start time, and the difference between the power value at the end time and the power value at the start time are calculated respectively, and the direction of change is determined according to the signs of the wind speed difference and the power difference; when both the wind speed difference and the power difference are greater than zero, the direction of change is recorded as upward; when both the wind speed difference and the power difference are less than zero, the direction of change is recorded as downward; when at least one of the wind speed difference and the power difference is equal to zero, the direction of change is recorded as stable. For the rate of change, the wind speed change rate is obtained by dividing the wind speed difference of the time evolution segment by the segment duration, and the power change rate is obtained by dividing the power difference of the time evolution segment by the segment duration. The wind speed change rate and the power change rate are then combined using an arithmetic mean to obtain the overall rate of change for the time evolution segment. For the fluctuation energy distribution, the absolute value of the power difference between adjacent sampling times within the time evolution segment is calculated sequentially, and all absolute power difference values ​​are arranged in chronological order. The time evolution segment is then divided into three equal intervals based on the number of sampling points: the first third, the middle third, and the last third. The average value of the absolute power difference within each of the three intervals is calculated, and this average is used as the fluctuation energy distribution result for the time evolution segment. Finally, the direction of change, the rate of change, and the fluctuation energy distribution are recorded in a preset order to form the segment evolution descriptor for the corresponding time evolution segment.

[0034] After obtaining the segment evolution descriptors corresponding to each time evolution segment, the segments are arranged according to their starting time in the original continuous time series. The arranged segment evolution descriptors are then sequentially connected to construct a non-stationary evolution trajectory sequence of the wind field's operating state. Specifically, the segment evolution descriptor corresponding to the earliest starting time time segment is taken as the first item in the non-stationary evolution trajectory sequence, and the segment evolution descriptor corresponding to the next adjacent time segment is taken as the next item. This process is repeated for all time evolution segments, resulting in a non-stationary evolution trajectory sequence that unfolds continuously over time. This non-stationary evolution trajectory sequence is used to characterize the continuous changes in the wind field's operating state across different time intervals.

[0035] Step S400: Compare and analyze the predicted power sequence with the actual power sequence at the corresponding time to obtain the prediction error sequence, and establish the coupling relationship between the prediction error sequence and the non-stationary evolution trajectory sequence.

[0036] In this embodiment, when comparing and analyzing the predicted power sequence with the actual power sequence at the corresponding time, time alignment processing is first performed on the predicted power sequence and the actual power sequence to establish a time-by-time correspondence between the two types of data under the same time mark. Specifically, the predicted power value at each prediction time in the predicted power sequence is read, and the actual power value at the same time mark in the actual power sequence is read. The predicted power value and the actual power value are arranged one by one according to the time mark matching method. Here, the predicted power sequence is the power prediction result of the power prediction model over continuous time, and the actual power sequence is the actual power measurement result of the wind farm within the same time range. After completing the time alignment, error difference calculation is performed for each corresponding time. The prediction error value is calculated by subtracting the predicted power value from the actual power value, and all prediction error values ​​are arranged in chronological order to obtain the prediction error sequence. The prediction error sequence is a sequence composed of prediction error values ​​corresponding to each prediction time within a continuous time range, used to represent the deviation between the power prediction result and the actual power.

[0037] After obtaining the prediction error sequence, a time correspondence processing is performed between the prediction error sequence and the non-stationary evolution trajectory sequence to establish a correlation between the two. Specifically, firstly, the time range corresponding to the evolution descriptors of each segment in the non-stationary evolution trajectory sequence is read, and the corresponding error time interval is determined in the prediction error sequence based on this time range, so that each time evolution segment corresponds to a set of prediction error values ​​within the same time range. Subsequently, statistical calculations are performed on the prediction error values ​​within this time range, including calculating the arithmetic mean of the prediction error values ​​within this time range, calculating the difference between the prediction error values ​​at adjacent sampling times, and calculating the average of the absolute values ​​of the prediction errors, thereby obtaining the error change result within this time range. Then, the error change result is associated with the segment evolution descriptors of the corresponding time evolution segments, so that each segment evolution descriptor corresponds to a set of error change results. This correspondence process is completed sequentially according to the time order of the segment evolution descriptors in the non-stationary evolution trajectory sequence, thereby establishing a coupling relationship between the prediction error sequence and the non-stationary evolution trajectory sequence. This coupling relationship is used to represent the correspondence between the characteristics of wind field operation state changes and the changes in prediction errors.

[0038] Step S500: Using the coupling relationship, perform a classification analysis of the prediction error change characteristics corresponding to different non-stationary evolution trajectory sequences to construct a set of error evolution patterns.

[0039] In this embodiment, when performing a classification analysis of prediction error change features corresponding to different non-stationary evolution trajectory sequences using coupling relationships, trajectory change events are first identified based on the change relationships between adjacent time evolution segments in the non-stationary evolution trajectory sequence. These trajectory change events include state rise events, state decay events, and state oscillation events. Subsequently, the error change process corresponding to each trajectory change event is extracted from the prediction error sequence, and error response features that characterize the error change response are calculated. These error response features include error response slope, error response delay, and error oscillation amplitude. Then, the corresponding trajectory-error response feature vector is constructed using the error response features, and a structured classification analysis is performed on the prediction error change patterns corresponding to different non-stationary evolution trajectories based on the trajectory-error response feature vector. Thus, a set of error evolution patterns is constructed based on the classification analysis results.

[0040] Furthermore, in the method provided in the application embodiments, utilizing the coupling relationship to perform a classification analysis of the prediction error change characteristics corresponding to different non-stationary evolution trajectory sequences, and constructing an error evolution pattern set, it further includes:

[0041] Based on the non-stationary evolution trajectory sequence, trajectory change events between adjacent time segments are identified. These trajectory change events include state rise events, state decay events, and state oscillation events. For each trajectory change event, the error response process within the corresponding time window is extracted from the prediction error sequence, and error response features characterizing the error's response to trajectory changes are calculated. These error response features include error response slope, error response delay, and error oscillation amplitude. A trajectory-error response feature vector is constructed using these error response features, and a structured classification analysis is performed on the prediction error change patterns corresponding to different non-stationary evolution trajectories based on the trajectory-error response feature vector. An error evolution pattern set is constructed based on the results of the structured classification analysis.

[0042] In this embodiment, when identifying trajectory change events between adjacent time evolution segments based on the non-stationary evolution trajectory sequence, the segment evolution descriptors of two adjacent time evolution segments are read sequentially, and their change directions, change rates, and fluctuation energy distribution results are compared. When the change direction of the later time evolution segment is upward and the change rate is greater than that of the previous time evolution segment, and the average fluctuation energy distribution of the later time evolution segment is greater than that of the previous time evolution segment, the change relationship between the adjacent segments is identified as a state rise event. When the change direction of the later time evolution segment is downward and the change rate is less than that of the previous time evolution segment, and the average fluctuation energy distribution of the later time evolution segment is less than that of the previous time evolution segment, the change relationship between the adjacent segments is identified as a state decay event. When the change directions of two adjacent time evolution segments are upward and downward, or downward and upward, the change relationship between the adjacent segments is identified as a state oscillation event, thereby obtaining the trajectory change events corresponding to each adjacent time evolution segment in the non-stationary evolution trajectory sequence.

[0043] Next, based on the temporal position of each trajectory change event in the non-stationary evolution trajectory sequence, the prediction error value within the corresponding time window is extracted from the prediction error sequence to form the error response process for the corresponding trajectory change event. Specifically, if the time interval corresponding to a trajectory change event is [t1, t2], then the prediction error values ​​with time markers between t1 and t2 are extracted from the prediction error sequence and arranged in chronological order. Subsequently, the error response process is calculated. The slope of the fitted line between the prediction error value and time is calculated using the least squares linear fitting method, which serves as the error response slope. The error response delay is obtained by calculating the time difference between the occurrence of the trajectory change event and the moment when the error change amplitude first exceeds the preset error change threshold in the prediction error sequence. The error oscillation amplitude is obtained by calculating the difference between the maximum and minimum prediction error values ​​within this time window, thus forming the error response characteristics that characterize the error's response to trajectory changes.

[0044] Subsequently, a trajectory-error response feature vector is constructed using error response characteristics. Specifically, after obtaining the error response slope, error response delay, and error oscillation amplitude corresponding to each trajectory change event, these three error response features are combined in a fixed order to form the corresponding trajectory-error response feature vector. Based on this feature vector, a structured classification analysis is performed on the prediction error change patterns corresponding to different non-stationary evolution trajectories. Specifically, firstly, a structural consistency metric is calculated for each trajectory-error response feature vector in the candidate prediction error change patterns to quantify the consistency level of the corresponding candidate prediction error change patterns in terms of trajectory direction, change rate, and error response amplitude. Then, the error response paths corresponding to the candidate prediction error change patterns are mapped to topological structures, and a topological stability metric is calculated based on the continuity, repeatability, and fluctuation amplitude stability of the topological structure to quantify the stability of the corresponding candidate prediction error change patterns during historical evolution. Finally, the structural consistency metric and topological stability metric are normalized, and a built-in trust index is generated through dynamic weight fusion. This built-in trust index, as the result of the structured classification analysis, is used to characterize the credibility of the corresponding prediction error change patterns.

[0045] After obtaining the built-in trust indices corresponding to each candidate prediction error change pattern, the prediction error change patterns are filtered and classified according to the built-in trust indices to construct a set of error evolution patterns. In this process, firstly, the built-in trust indices corresponding to all candidate prediction error change patterns are read and sorted according to their numerical values; then, a preset trust threshold is set, and candidate prediction error change patterns with built-in trust indices greater than or equal to the preset trust threshold are marked as valid patterns, while candidate prediction error change patterns with built-in trust indices less than the preset trust threshold are marked as invalid patterns. Then, feature interval division processing is performed on the candidate prediction error change patterns marked as valid patterns. The historical value ranges of error response slope, error response delay, and error oscillation amplitude in all trajectory-error response feature vectors are statistically analyzed, and the historical value ranges are segmented according to preset interval division rules to obtain the corresponding error response slope interval, error response delay interval, and error oscillation amplitude interval. Subsequently, the trajectory-error response feature vectors corresponding to each valid candidate prediction error change pattern are mapped to the above interval ranges, and classified according to the trajectory change event type and the interval position corresponding to the trajectory-error response feature vector. Candidate prediction error change patterns with the same trajectory change event type and trajectory-error response feature vectors falling into the same interval combination are grouped into the same category. Finally, the candidate prediction error change patterns corresponding to each category are organized and stored according to the category number, and the statistical results of the built-in trust index corresponding to each category are recorded, thereby forming a set of error evolution patterns used to describe the prediction error change law under different non-stationary operating conditions.

[0046] Furthermore, in the method provided in the application embodiments, the structured classification analysis of the prediction error change patterns corresponding to different non-stationary evolution trajectories based on the trajectory-error response feature vector further includes:

[0047] The structural consistency metric is calculated for each trajectory-error response feature vector in the candidate prediction error change pattern. This structural consistency metric quantifies the consistency level of the corresponding candidate prediction error change pattern in trajectory direction, change rate, and error response amplitude. The error response path of the candidate prediction error change pattern is mapped to a topological structure, and a topological stability metric is calculated based on the continuity, repeatability, and fluctuation amplitude stability of the topological structure. This topological stability metric quantifies the reliability of the corresponding candidate prediction error change pattern in historical evolution. After normalizing the structural consistency metric and the topological stability metric, dynamic weight fusion is performed to generate a built-in trust index. An error evolution pattern set is constructed based on the built-in trust index.

[0048] In this embodiment, after obtaining the trajectory-error response feature vectors corresponding to each candidate prediction error change pattern, a consistency calculation is performed on each trajectory-error response feature vector in the same candidate prediction error change pattern to obtain the corresponding structural consistency metric. Specifically, the trajectory direction, rate of change, and error oscillation amplitude corresponding to all trajectory-error response feature vectors in the candidate prediction error change pattern are read; then, the arithmetic mean of the rate of change and the arithmetic mean of the error oscillation amplitude are calculated respectively, and the absolute average of the difference between each rate of change and the average rate of change, and the absolute average of the difference between each error oscillation amplitude and the average error oscillation amplitude are calculated; then, the above two average deviation values ​​are added together, and their reciprocal is used as the consistency result of the change features. At the same time, the proportion of feature vectors with the same trajectory direction to the total number of feature vectors is counted as the trajectory direction consistency result; finally, the consistency result of the change features and the consistency result of the trajectory direction are weighted and summed according to a preset weight to obtain the structural consistency metric. The structural consistency metric is used to quantify the consistency level of the corresponding candidate prediction error change pattern in terms of trajectory direction, rate of change, and error response amplitude. The weight of the consistency result of change characteristics is set to 0.5, and the weight of the consistency result of trajectory direction is set to 0.5.

[0049] After obtaining the structural consistency metric, the error response processes in the candidate prediction error change patterns are connected in chronological order to form an error response path, and the error response path is converted into a topological structure composed of continuously changing segments in order to calculate the topological stability metric. Specifically, the predicted error values ​​in the error response path are calculated using time-by-time differential calculation. When the difference between the predicted error values ​​of two adjacent sampling times is greater than zero, it is recorded as an ascending segment; when the difference is less than zero, it is recorded as a descending segment; and when the difference is equal to zero, it is recorded as a stable segment. The position of segment change is recorded as a turning point. Then, the ratio of the number of continuous segments in the error response path to the total number of segments is calculated as the continuity result. Next, the proportion of the occurrence of error response paths with the same segment arrangement order in historical samples is calculated as the repeatability result. Then, the average value of all error oscillation amplitudes in the candidate prediction error change pattern is calculated, and the average value of the absolute value of the difference between each error oscillation amplitude and the average value is calculated. The reciprocal of this average value is then used as the fluctuation amplitude stability result. Finally, the continuity result, repeatability result, and fluctuation amplitude stability result are weighted and summed according to preset weights to obtain the topological stability metric. The topological stability metric is used to quantify the reliability of the corresponding candidate prediction error change pattern in historical evolution. The weight of continuous results is set to 0.4, the weight of repeatable results is set to 0.3, and the weight of fluctuation amplitude stability results is set to 0.3.

[0050] After obtaining the structural consistency metric and topological stability metric, normalization and weighted fusion are performed on the two metrics to generate the built-in trust index. In this process, the maximum and minimum values ​​of the structural consistency metric and the topological stability metric are statistically analyzed for all candidate prediction error change patterns. Then, the normalized structural consistency value and normalized topological stability value are calculated using the minimum-maximum normalization method. Finally, a fixed weight is used for fusion, with the weight of the normalized structural consistency value set to 0.6 and the weight of the normalized topological stability value set to 0.4. The built-in trust index is calculated as: Built-in Trust Index = 0.6 × Normalized Structural Consistency Value + 0.4 × Normalized Topological Stability Value, thus obtaining the built-in trust index for the corresponding candidate prediction error change pattern.

[0051] After obtaining the built-in trust indicators corresponding to each candidate prediction error change pattern, an error evolution pattern set is constructed based on these indicators. Specifically, the built-in trust indicators corresponding to all candidate prediction error change patterns are read and sorted from largest to smallest. Then, a preset trust threshold of 0.6 is set, and it is determined whether the built-in trust indicator corresponding to each candidate prediction error change pattern is greater than or equal to 0.6. If the condition is met, the corresponding candidate prediction error change pattern is retained; otherwise, it is removed. The retained candidate prediction error change patterns are then classified according to trajectory change event type, and within the same trajectory change event type, they are grouped according to the range of values ​​for error response slope, error response delay, and error oscillation amplitude. The range of values ​​is determined by statistically analyzing the maximum and minimum values ​​of the corresponding feature values ​​of all retained patterns and using an equal-interval division method. Finally, candidate prediction error change patterns of the same trajectory change event type that fall into the same feature interval combination are grouped into the same category, and the pattern information corresponding to each category is stored, thus forming an error evolution pattern set.

[0052] Furthermore, the method provided in the application embodiments also includes:

[0053] The error evolution pattern set includes a dynamic updater, which performs additions, deletions, weight adjustments, and updates of the built-in trust index of existing error evolution patterns based on the non-stationary evolution trajectory and actual power sequence obtained at the new prediction time.

[0054] In this embodiment, a dynamic updater is set in the error evolution model set to update the existing error evolution models based on the non-stationary evolution trajectory and actual power sequence obtained at the new prediction time. Specifically, meteorological monitoring data, predicted power values, and actual power values ​​within the corresponding time range are obtained at the new prediction time, and the corresponding non-stationary evolution trajectory and prediction error sequence are extracted according to the aforementioned method to calculate the trajectory-error response feature vector corresponding to that time. Subsequently, the similarity of the trajectory-error response feature vector is compared with the pattern center feature vectors corresponding to each pattern in the error evolution pattern set. The pattern center feature vector is obtained by taking the arithmetic mean of the error response slope, error response delay, and error oscillation amplitude of all trajectory-error response feature vectors in the pattern. Then, the feature difference between the new trajectory-error response feature vector and each pattern center feature vector is calculated, and the similarity is calculated according to the formula: similarity = 1 - (0.4 × slope difference + 0.3 × delay difference + 0.3 × oscillation amplitude difference). When the similarity is greater than or equal to the preset matching threshold, the trajectory-error response feature vector is assigned to the corresponding pattern, and the sample number, pattern center feature vector, and built-in trust index of the pattern are updated accordingly.

[0055] When the calculated similarity is less than a preset matching threshold, the trajectory-error response feature vector is added to the error evolution pattern set as a new prediction error change pattern, thus realizing the addition of a new pattern. Subsequently, the number of matches for each pattern is counted within a preset statistical time window, and the total number of matches for all patterns is counted. Then, the number of matches for a certain pattern is divided by the total number of matches for all patterns to obtain the matching ratio corresponding to that pattern. The original weight of the pattern is then updated according to the formula: updated weight = 0.5 × original weight + 0.5 × matching ratio, thus completing the weight adjustment of each pattern. After that, the built-in trust index corresponding to each pattern is recalculated, and when the number of matches for a certain pattern within the statistical time window is less than a preset deletion threshold, the pattern is deleted from the error evolution pattern set, thus completing the addition and deletion of the error evolution pattern set, weight adjustment, and update of the built-in trust index.

[0056] Step S600: At the new prediction time, perform similarity matching of the error evolution pattern set based on the non-stationary evolution trajectory of the current node, and use the similarity matching results to perform error correction management.

[0057] In this embodiment, at a new prediction time, when performing similarity matching of the error evolution pattern set based on the non-stationary evolution trajectory of the current node, the state transition relationship between adjacent time evolution segments is first analyzed on the non-stationary evolution trajectory of the current node to obtain a trajectory transition description sequence characterizing the direction, intensity, and persistence of wind field state changes. Subsequently, in the error evolution pattern set, the corresponding trajectory transition description sequence is extracted for each historical non-stationary evolution trajectory, and a historical trajectory transition structure set is constructed. Then, the trajectory transition description sequence of the current node is matched with the historical trajectory transition structure set to obtain the corresponding similarity matching result. Finally, the confidence prediction analysis of the error change trend is performed using the similarity matching result, and error correction management is implemented accordingly.

[0058] Furthermore, the method provided in the application embodiments, which performs similarity matching of the error evolution pattern set based on the non-stationary evolution trajectory of the current node, further includes:

[0059] The process involves analyzing the state transition relationships between adjacent time segments of the non-stationary evolution trajectory of the current node to obtain trajectory transition description sequences that characterize the direction, intensity, and persistence of wind field state changes; extracting corresponding trajectory transition description sequences from each historical non-stationary evolution trajectory in the error evolution pattern set and constructing a historical trajectory transition structure set; performing transition structure matching between the trajectory transition description sequences and the historical trajectory transition structure set to establish similarity matching results; and using the similarity matching results to perform confidence prediction analysis of error change trends and implement error correction management.

[0060] In this embodiment, when analyzing the state transition relationship between adjacent time evolution segments of the non-stationary evolution trajectory of the current node, the state transition relationship containing wind speed, wind direction, and power change features is first input into the three-dimensional feature extraction channel as input data. In the trajectory change direction extraction sub-channel, the direction deviation of the wind speed and power change features of adjacent time evolution segments is calculated to form a direction change sequence, and the direction change sequence is quantized into discrete direction status codes representing different change direction types. In the trajectory change intensity extraction sub-channel, the wind speed gradient, wind direction offset rate, and power fluctuation amplitude of adjacent time evolution segments are calculated to obtain the change amplitude index, and a change intensity level sequence is generated according to a preset quantization rule. In the trajectory change persistence extraction sub-channel, the duration of continuous change direction and change intensity states in time is statistically analyzed to form a duration sequence, and a corrected duration sequence is obtained through smoothing filtering and outlier correction. Finally, the discrete direction status codes, change intensity level sequences, and corrected duration sequences are fused in chronological order to form a trajectory transition description sequence characterizing the change direction, change intensity, and change persistence of the wind field state.

[0061] After obtaining the trajectory transition description sequence corresponding to the current node, trajectory feature extraction processing is performed on each historical non-stationary evolution trajectory in the error evolution mode set to construct a historical trajectory transition structure set. Specifically, each historical non-stationary evolution trajectory recorded in the error evolution mode set is read, and the state transition relationship between adjacent time evolution segments is analyzed according to the trajectory analysis method consistent with the current node. Then, the discrete direction state code, change intensity level sequence, and correction duration sequence corresponding to each historical non-stationary evolution trajectory are extracted in sequence and combined in chronological order to form the corresponding trajectory transition description sequence. Then, the trajectory transition description sequences corresponding to all historical non-stationary evolution trajectories are classified according to the error evolution mode category, and the trajectory transition description sequences under the same mode category are uniformly stored, thereby obtaining a historical trajectory transition structure set used to describe the historical wind field state change structure.

[0062] After obtaining the set of historical trajectory transition structures, the trajectory transition description sequence of the current node is matched with each trajectory transition description sequence in the set of historical trajectory transition structures to establish a similarity matching result. Specifically, firstly, the discrete direction status codes, change intensity level sequences, and correction duration sequences in the current node's trajectory transfer description sequence are read sequentially in chronological order. Then, the consistency ratio of the current trajectory transfer description sequence with each historical trajectory transfer description sequence in terms of discrete direction status codes is calculated, along with the average absolute value of the differences between change intensity levels and the average absolute value of the differences between correction durations. Next, minimum-maximum value normalization is performed on the difference between change intensity levels and the difference between correction durations. Then, a comprehensive similarity calculation is performed according to preset weights, where the weight of the discrete direction status code consistency ratio is set to 0.4, the weight of the normalized difference between change intensity levels is set to 0.3, and the weight of the normalized difference between correction durations is set to 0.3. The calculation is performed as follows: Comprehensive Similarity = 0.4 × Direction Consistency Ratio + 0.3 × (1 - Intensity Difference) + 0.3 × (1 - Duration Difference). Finally, the structure with the highest comprehensive similarity from all historical trajectory transfer structures is selected as the matching structure, thus forming the similarity matching result corresponding to the current node.

[0063] After obtaining the similarity matching results, confidence prediction analysis of the error change trend is performed using the similarity matching results. In this process, historical error evolution pattern records corresponding to the matched historical trajectory transfer structure are read, and the error response slope, error response delay, and error oscillation amplitude corresponding to all historical samples in the pattern are extracted. Then, the arithmetic mean of the three features is calculated to obtain the average historical error response slope, average historical error response delay, and average historical error oscillation amplitude. Next, the comprehensive similarity between the current node and the matched historical trajectory transfer structure is recorded as the confidence weight, and the change in prediction error is calculated based on the prediction time step Δt, where the change in prediction error = confidence weight × average historical error response slope × Δt. Then, the change in prediction error and the average historical error oscillation amplitude are used together to determine the error change range, where the lower limit of the error change range = change in prediction error - average historical error oscillation amplitude, and the upper limit = change in prediction error + average historical error oscillation amplitude. Finally, the change in prediction error is used as the prediction error change trend value of the current node.

[0064] After obtaining the prediction error trend value, error correction management is performed. The prediction power value corresponding to the current node is read, and the prediction error trend value is added to the prediction power value as the error correction amount. When the prediction error trend value is positive, the correction amount is added to the prediction power value. When the prediction error trend value is negative, the correction amount is subtracted from the prediction power value to obtain the corrected prediction power value. The corrected prediction power value is then used as the final power prediction result corresponding to the current node.

[0065] Furthermore, the method provided in the application embodiment, which analyzes the state transition relationship between adjacent time evolution segments of the non-stationary evolution trajectory of the current node to obtain a trajectory transition description sequence characterizing the direction, intensity, and persistence of wind field state changes, also includes:

[0066] The state transition relationship includes wind speed, wind direction, and power change characteristics. This state transition relationship is used as input data and fed into a three-dimensional feature extraction channel. The process includes: inputting the wind speed and power change characteristics of adjacent segments into a trajectory change direction extraction sub-channel; calculating the direction deviation of the wind speed and power change characteristics to form a direction change sequence; quantizing the direction change sequence into discrete direction status codes, which represent different directional types such as upward surge, downward decay, and offset changes; activating a trajectory change intensity extraction sub-channel; calculating the wind speed gradient, wind direction offset rate, and power fluctuation amplitude of adjacent segments to form a change amplitude index; classifying the change amplitude index according to preset quantization rules to generate a change intensity level sequence, which represents the magnitude of change between different segments; activating a trajectory change persistence extraction sub-channel; statistically analyzing the duration of each continuous change direction and intensity state over time to form a duration sequence; smoothing and filtering the duration sequence and correcting outliers to establish a corrected duration sequence; and fusing the discrete direction status codes, change intensity level sequence, and corrected duration sequence in chronological order to form a trajectory transition description sequence.

[0067] In this embodiment, the state transition relationship between adjacent time segments in the non-stationary evolution trajectory of the current node is first extracted. Wind speed, wind direction, and power data are read from adjacent time segments, and the wind speed difference, wind direction difference, and power difference between the next and previous time segments are calculated using a differential calculation method. This yields wind speed change features, wind direction change features, and power change features characterizing the changes in the wind field's operating state. These features are then input as state transition relationship data into the three-dimensional feature extraction channel. The three-dimensional feature extraction channel consists of a data processing structure composed of three parallel feature calculation modules: a trajectory change direction extraction sub-channel, a trajectory change intensity extraction sub-channel, and a trajectory change persistence extraction sub-channel. These modules are used to extract the directional features, amplitude features, and duration features of the changes in the operating state, respectively.

[0068] In the trajectory change direction extraction sub-channel, wind speed and power change characteristics between adjacent time evolution segments are used as input data to calculate direction deviation. During this process, the direction of change is determined by judging the sign relationship between the wind speed difference and the power difference. When both wind speed and power differences are positive, it is recorded as an upward change state; when both are negative, it is recorded as a downward change state; and when their signs are inconsistent, it is recorded as an offset change state, thus forming a direction change sequence arranged in chronological order. Subsequently, discrete encoding processing is performed on the direction change sequence. Using preset encoding rules, upward change states are encoded as the value 1, downward change states as the value -1, and offset change states as the value 0, resulting in a discrete direction status code sequence composed of integer codes.

[0069] Subsequently, in the trajectory change intensity extraction sub-channel, the change amplitude between adjacent time evolution segments is calculated. In this process, the wind speed gradient between adjacent time evolution segments is first calculated, where the wind speed gradient is obtained by dividing the wind speed difference by the time interval; then the wind direction shift rate between adjacent time evolution segments is calculated, where the wind direction shift rate is obtained by dividing the wind direction difference by the time interval; next, the power fluctuation amplitude between adjacent time evolution segments is calculated, where the power fluctuation amplitude is obtained by the absolute value of the power difference; then, the wind speed gradient, wind direction shift rate, and power fluctuation amplitude are weighted and combined to form a change amplitude index, where the weight of the wind speed gradient is set to 0.4, the weight of the wind direction shift rate is set to 0.3, and the weight of the power fluctuation amplitude is set to 0.3, and the change amplitude index is calculated according to the formula: change amplitude index = 0.4 × wind speed gradient + 0.3 × wind direction shift rate + 0.3 × power fluctuation amplitude; then, the change amplitude index is subjected to interval quantization processing according to a preset quantization rule, for example, divided into three level intervals according to the numerical range, and encoded as numerical values ​​1, 2, and 3 respectively, thus forming a change intensity level sequence composed of numerical codes.

[0070] Subsequently, in the trajectory change persistence extraction sub-channel, statistical processing is performed on the persistence of state changes in the time dimension. In this process, segments with the same discrete direction state code and change intensity level combination in continuous time evolution segments are continuously counted, and the duration value corresponding to this combination state is obtained by multiplying the number of continuous segments by the sampling time interval, thus forming a duration sequence arranged in chronological order. Then, a moving average filter is applied to the duration sequence, using a sliding window averaging calculation to eliminate short-term fluctuation noise. Simultaneously, outlier correction processing is performed on abnormal data in the duration sequence that exceed preset upper and lower thresholds, replacing the abnormal duration with the average value of neighboring time periods, thus obtaining a corrected duration sequence.

[0071] After obtaining the discrete direction status code sequence, the change intensity level sequence, and the corrected duration sequence, the three types of numerical sequences are fused to form a trajectory transfer description sequence. In this process, using the order of time evolution segments as an index, the discrete direction status code, change intensity level, and corrected duration corresponding to the same time position are combined and encoded, and arranged sequentially according to time order to form a trajectory transfer description sequence that simultaneously contains numerical values ​​of direction change characteristics, change intensity characteristics, and change duration. This yields a trajectory transfer description sequence used to describe the direction, intensity, and duration of wind field state changes.

[0072] Furthermore, the method provided in the application embodiments, which utilizes similarity matching results to perform error correction management, further includes:

[0073] Determine whether the similarity matching result meets the preset matching threshold; if the similarity matching result does not meet the preset matching threshold, configure a balance coefficient based on the similarity matching result and the weather forecast confidence level; perform a balanced forecast analysis on the weather forecast result and the similarity matching result based on the balance coefficient, and establish a corrected forecast result.

[0074] In this embodiment, a threshold determination process is first performed on the currently calculated similarity matching result to determine whether the similarity matching result meets a preset matching threshold. During this process, the comprehensive similarity calculated between the current node's trajectory transition description sequence and the historical trajectory transition structure is read, and this comprehensive similarity is numerically compared with a preset matching threshold. When the comprehensive similarity is greater than or equal to the matching threshold, it is determined that the non-stationary evolution trajectory of the current node has a sufficiently high consistency with the historical trajectory transition structure. In this case, the error change trend prediction result corresponding to the similarity matching result is directly used as the basis for prediction correction. When the comprehensive similarity is less than the matching threshold, it is determined that the consistency between the current node's trajectory structure and the historical trajectory transition structure is insufficient. In this case, a balance coefficient is configured based on the similarity matching result and the meteorological prediction confidence level. The confidence level of weather forecasts is calculated by evaluating the prediction performance of the weather forecast model on historical samples. The weather forecast model is pre-trained using a supervised learning method. Specifically, it reads historical weather monitoring data and historical actual power data, using wind speed, wind direction, temperature, and air pressure from the historical weather monitoring data as input features and the corresponding historical actual power as the output label, thus constructing a mapping relationship between weather conditions and power output. The weather forecast model can employ a Long Short-Term Memory (LSTM) network model with the following parameters: input time step of 24, 2 hidden layers, 128 hidden units per layer, learning rate of 0.001, batch size of 64, and training... The model has 100 rounds, a Dropout parameter of 0.2, and an output layer dimension of 1. After training, the current meteorological monitoring data is input into the meteorological prediction model to obtain the corresponding meteorological prediction results. Then, the absolute value of the prediction error of the meteorological prediction model within a preset historical time window is calculated, and the mean absolute error is calculated. The meteorological prediction confidence score is then normalized using the formula: meteorological prediction confidence score = 1 - mean absolute error / historical maximum error. Finally, the balance coefficient is calculated as: balance coefficient = meteorological prediction confidence score / (meteorological prediction confidence score + similarity matching result) to obtain the balance coefficient used to weigh the contribution of the meteorological prediction results and the similarity matching results.

[0075] After obtaining the balance coefficient, a balanced prediction analysis is performed on the meteorological forecast results and the similarity matching results. This process begins by reading the meteorological forecast results output by the meteorological forecast model and the prediction error correction value obtained based on the similarity matching results. Then, the two results are weighted and fused according to the balance coefficient. The weight of the meteorological forecast results is the balance coefficient, and the weight of the prediction results corrected based on the similarity matching results is 1 minus the balance coefficient. The corrected prediction result is calculated as: Corrected Prediction Result = Balance Coefficient × Meteorological Forecast Result + (1 - Balance Coefficient) × (Meteorological Forecast Result + Prediction Error Correction Value), thus obtaining the corrected prediction result that fuses the meteorological forecast results and the similarity matching results.

[0076] Furthermore, the method provided in the application embodiments, which utilizes similarity matching results to perform error correction management, further includes:

[0077] A data correction report is generated based on the error correction management data, and the data correction report is encrypted and stored; the encrypted storage feedback is then displayed visually.

[0078] In this embodiment, a data correction report is first generated based on the correction results produced by the error correction management process. Specifically, the original predicted power data, the corrected predicted power data, and the corresponding prediction error trend values ​​obtained during the error correction management phase are read, and the original predicted power data, the corrected predicted power data, the error correction amount, and the correction time information are organized. Subsequently, the predicted data and corrected data at each time point are written into a structured data table in chronological order, and this structured data table is converted into a data correction report. The data correction report records information such as the predicted power value, the corrected power value, the error correction amount, and the corresponding time node to achieve data recording and result tracking of the error correction process.

[0079] After generating the data correction report, it undergoes encrypted storage. This process involves first converting the report into a standard data file format, such as JSON or CSV; then, encrypting the report using the AES symmetric encryption algorithm with a 128-bit encryption key, and encrypting the data file content using a preset system key to obtain an encrypted correction report file; finally, writing the encrypted correction report file to a database storage unit or file storage server, thus achieving secure storage of the data correction report.

[0080] After encrypted storage is completed, the feedback from the encrypted storage is visualized. This process first reads the storage status information returned by the encrypted storage operation, which includes the data storage time, data file number, and storage success identifier. Then, the storage status information is parsed by the data parsing module, and the parsing results are transmitted to the visualization interface. Next, the graphical interface component plots the corrected power curve, the original predicted power curve, and the changes in error correction amount from the data correction report as a time series curve, and displays the encrypted storage status information on the interface, thus achieving a visual display of the data correction results and storage status.

[0081] Furthermore, the method provided in the application embodiments also includes:

[0082] The data correction report is subjected to correction verification to generate a verification dataset. The verification dataset is then used to configure a feedback database, which is used for verification feedback management of power prediction error correction.

[0083] In this embodiment, the generated data correction report is first subjected to correction verification processing. The time node, original predicted power value, corrected predicted power value, and actual power value at the corresponding time are read from the data correction report, and a corresponding relationship is established according to the time order. Then, the difference between the original predicted power value and the actual power value is calculated to obtain the original prediction error at the corresponding time node. At the same time, the difference between the corrected predicted power value and the actual power value is calculated to obtain the corrected prediction error at the corresponding time node. Then, the original prediction error and the corrected prediction error are compared and calculated. The error improvement value is obtained by subtracting the corrected prediction error from the original prediction error. Afterwards, the time node, original predicted power value, corrected predicted power value, actual power value, original prediction error, corrected prediction error, and error improvement value are organized according to a unified field structure and formed into structured data records according to the time order, thereby generating a verification dataset for verifying the error correction effect.

[0084] The feedback database is then configured using the validation dataset. In this process, a data table structure is created in the database, and fields for time index, original predicted power, corrected predicted power, actual power, original prediction error, corrected prediction error, and error improvement are set. The structured data records from the validation dataset are then written into the database tables one by one, and a time index is created for the time index field. Finally, the records in the database are sorted and stored in chronological order, thus forming the feedback database used to record the validation data for prediction error correction.

[0085] After configuring the feedback database, the verification feedback management for power prediction error correction is performed using the feedback database. During this process, the original prediction error data, corrected prediction error data, and error improvement data for the corresponding time period are read from the feedback database according to a preset time window. The average original prediction error, the average corrected prediction error, and the proportion of data with positive error improvement values ​​within that time window are calculated respectively. Then, the average corrected prediction error is compared with the average original prediction error, and a determination is made based on the proportion of data with positive error improvement values. When the average corrected prediction error is less than the average original prediction error and the proportion of data with positive error improvement values ​​is greater than a preset threshold, the error correction result for that time window is recorded as a valid correction result, and the data for the corresponding time period is written into the feedback database for valid verification. The system records data as follows: When the average corrected prediction error is greater than or equal to the average original prediction error, or when the proportion of data with positive error improvement is less than or equal to a preset proportion threshold, meteorological monitoring data, non-stationary evolution trajectory characteristic data, and predicted power data for the corresponding time period are extracted from the feedback database. This data is then written into the verification feedback record table in the feedback database, and the error statistics for that time window are recorded, thus forming a verification feedback data record. Subsequently, in subsequent prediction cycles, the verification feedback data record in the feedback database is read to compare and verify the error correction results for subsequent time periods, and the verification record in the feedback database is continuously updated, thereby completing the continuous verification feedback management of the power prediction error correction process.

[0086] In summary, the embodiments of this application have at least the following technical effects:

[0087] This application acquires meteorological monitoring data, historical power data, and predicted power sequences output by power prediction models for wind farms within a continuous time series. It performs time-series change analysis on the meteorological monitoring data and historical power data, dynamically segmenting the continuous time series based on wind speed change gradient, wind direction shift rate, and power fluctuation intensity to form multiple time evolution segments reflecting local operating states. It extracts segment evolution descriptors representing the change direction, change rate, and fluctuation energy distribution of each time evolution segment, and constructs a non-stationary evolution trajectory sequence of the wind farm's operating state in chronological order. It compares and analyzes the predicted power sequence with the actual power sequence at the corresponding time point to obtain a prediction error sequence, and establishes a coupling relationship between the prediction error sequence and the non-stationary evolution trajectory sequence. Using this coupling relationship, it performs a classification analysis of the prediction error change characteristics corresponding to different non-stationary evolution trajectory sequences, constructing a set of error evolution patterns. At a new prediction time, it performs similarity matching of the error evolution pattern set based on the non-stationary evolution trajectory of the current node, and uses the similarity matching results to perform error correction management. This invention addresses the technical problem in existing technologies where wind power prediction is difficult to adapt to non-stationary changes in wind power operating conditions, leading to difficulties in timely correction of prediction errors. By constructing a non-stationary evolution trajectory of wind farm operating conditions and establishing a coupling relationship between prediction errors and operating conditions, the invention achieves the technical effect of improving the accuracy of wind power prediction error correction.

[0088] Example 2, based on the same inventive concept as the power prediction error correction method for non-stationary wind power data in the aforementioned examples, such as... Figure 2 As shown, this application provides a power prediction error correction system for non-stationary wind power data. The system and method embodiments in this application are based on the same inventive concept. The system includes:

[0089] Data acquisition module 11 is used to acquire meteorological monitoring data, historical power data, and predicted power sequence output by the power prediction model of the wind farm in a continuous time series. Dynamic segmentation module 12 is used to perform time-series change analysis on the meteorological monitoring data and historical power data, dynamically segmenting the continuous time series according to wind speed change gradient, wind direction shift rate, and power fluctuation intensity to form multiple time evolution segments reflecting local operating states. Trajectory sequence construction module 13 is used to extract segment evolution descriptors representing the change direction, change rate, and fluctuation energy distribution of each time evolution segment, and construct a non-uniform wind farm operating state according to time sequence. The system includes: a stationary evolution trajectory sequence; a comparison analysis module 14, used to compare and analyze the predicted power sequence with the actual power sequence at the corresponding time, obtain the prediction error sequence, and establish the coupling relationship between the prediction error sequence and the non-stationary evolution trajectory sequence; a classification analysis module 15, used to use the coupling relationship to perform classification analysis of the prediction error change characteristics corresponding to different non-stationary evolution trajectory sequences, and construct an error evolution mode set; and an error correction management module 16, used to perform similarity matching of the error evolution mode set according to the non-stationary evolution trajectory of the current node at a new prediction time, and perform error correction management using the similarity matching results.

[0090] Furthermore, the system is also used to implement the following functions:

[0091] The process involves analyzing the state transition relationships between adjacent time segments of the non-stationary evolution trajectory of the current node to obtain trajectory transition description sequences that characterize the direction, intensity, and persistence of wind field state changes; extracting corresponding trajectory transition description sequences from each historical non-stationary evolution trajectory in the error evolution pattern set and constructing a historical trajectory transition structure set; performing transition structure matching between the trajectory transition description sequences and the historical trajectory transition structure set to establish similarity matching results; and using the similarity matching results to perform confidence prediction analysis of error change trends and implement error correction management.

[0092] Furthermore, the system is also used to implement the following functions:

[0093] The state transition relationship includes wind speed, wind direction, and power change characteristics. This state transition relationship is used as input data and fed into a three-dimensional feature extraction channel. The process includes: inputting the wind speed and power change characteristics of adjacent segments into a trajectory change direction extraction sub-channel; calculating the direction deviation of the wind speed and power change characteristics to form a direction change sequence; quantizing the direction change sequence into discrete direction status codes, which represent different directional types such as upward surge, downward decay, and offset changes; activating a trajectory change intensity extraction sub-channel; calculating the wind speed gradient, wind direction offset rate, and power fluctuation amplitude of adjacent segments to form a change amplitude index; classifying the change amplitude index according to preset quantization rules to generate a change intensity level sequence, which represents the magnitude of change between different segments; activating a trajectory change persistence extraction sub-channel; statistically analyzing the duration of each continuous change direction and intensity state over time to form a duration sequence; smoothing and filtering the duration sequence and correcting outliers to establish a corrected duration sequence; and fusing the discrete direction status codes, change intensity level sequence, and corrected duration sequence in chronological order to form a trajectory transition description sequence.

[0094] Furthermore, the system is also used to implement the following functions:

[0095] Based on the non-stationary evolution trajectory sequence, trajectory change events between adjacent time segments are identified. These trajectory change events include state rise events, state decay events, and state oscillation events. For each trajectory change event, the error response process within the corresponding time window is extracted from the prediction error sequence, and error response features characterizing the error's response to trajectory changes are calculated. These error response features include error response slope, error response delay, and error oscillation amplitude. A trajectory-error response feature vector is constructed using these error response features, and a structured classification analysis is performed on the prediction error change patterns corresponding to different non-stationary evolution trajectories based on the trajectory-error response feature vector. An error evolution pattern set is constructed based on the results of the structured classification analysis.

[0096] Furthermore, the system is also used to implement the following functions:

[0097] The structural consistency metric is calculated for each trajectory-error response feature vector in the candidate prediction error change pattern. This structural consistency metric quantifies the consistency level of the corresponding candidate prediction error change pattern in trajectory direction, change rate, and error response amplitude. The error response path of the candidate prediction error change pattern is mapped to a topological structure, and a topological stability metric is calculated based on the continuity, repeatability, and fluctuation amplitude stability of the topological structure. This topological stability metric quantifies the reliability of the corresponding candidate prediction error change pattern in historical evolution. After normalizing the structural consistency metric and the topological stability metric, dynamic weight fusion is performed to generate a built-in trust index. An error evolution pattern set is constructed based on the built-in trust index.

[0098] Furthermore, the system is also used to implement the following functions:

[0099] The error evolution pattern set includes a dynamic updater, which performs additions, deletions, weight adjustments, and updates of the built-in trust index of existing error evolution patterns based on the non-stationary evolution trajectory and actual power sequence obtained at the new prediction time.

[0100] Furthermore, the system is also used to implement the following functions:

[0101] Determine whether the similarity matching result meets the preset matching threshold; if the similarity matching result does not meet the preset matching threshold, configure a balance coefficient based on the similarity matching result and the weather forecast confidence level; perform a balanced forecast analysis on the weather forecast result and the similarity matching result based on the balance coefficient, and establish a corrected forecast result.

[0102] Furthermore, the system is also used to implement the following functions:

[0103] A data correction report is generated based on the error correction management data, and the data correction report is encrypted and stored; the encrypted storage feedback is then displayed visually.

[0104] Furthermore, the system is also used to implement the following functions:

[0105] The data correction report is subjected to correction verification to generate a verification dataset. The verification dataset is then used to configure a feedback database, which is used for verification feedback management of power prediction error correction.

[0106] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0107] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for correcting power prediction errors using non-stationary wind power data, characterized in that, The method includes: Acquire meteorological monitoring data, historical power data, and predicted power series output by the power prediction model for wind farms in continuous time series; The meteorological monitoring data and historical power data are subjected to time series change analysis. The continuous time series is dynamically segmented according to the wind speed change gradient, wind direction shift rate and power fluctuation intensity to form multiple time evolution segments reflecting the local operating status. Extract segment evolution descriptors representing the direction, rate, and energy distribution of change in each time evolution segment, and construct a non-stationary evolution trajectory sequence of the wind field operating state in chronological order; The predicted power sequence is compared and analyzed with the actual power sequence at the corresponding time to obtain the prediction error sequence, and the coupling relationship between the prediction error sequence and the non-stationary evolution trajectory sequence is established. Using the aforementioned coupling relationship, a classification analysis of the prediction error change characteristics corresponding to different non-stationary evolution trajectory sequences is performed to construct a set of error evolution patterns; At the new prediction time, similarity matching of the error evolution pattern set is performed based on the non-stationary evolution trajectory of the current node, and error correction management is performed using the similarity matching results.

2. The method for correcting power prediction errors for non-stationary wind power data as described in claim 1, characterized in that, Perform similarity matching of the error evolution pattern set based on the non-stationary evolution trajectory of the current node, including: The state transition relationship between adjacent time segments is analyzed in the non-stationary evolution trajectory of the current node to obtain the trajectory transition description sequence that characterizes the direction, intensity and duration of wind field state changes; Extract the corresponding trajectory transition description sequence from each historical non-stationary evolution trajectory in the set of execution error evolution modes, and construct a set of historical trajectory transition structures; The trajectory transfer description sequence is matched with the historical trajectory transfer structure set to establish a similarity matching result. The similarity matching results are used to perform confidence prediction analysis on error change trends and to perform error correction management.

3. The method for correcting power prediction errors for non-stationary wind power data as described in claim 2, characterized in that, The state transition relationships between adjacent time segments are analyzed from the non-stationary evolution trajectory of the current node to obtain a trajectory transition description sequence characterizing the direction, intensity, and persistence of wind field state changes, including: The state transition relationship includes wind speed, wind direction, and power change characteristics. This state transition relationship is used as input data and fed into the three-dimensional feature extraction channel, including: The wind speed and power change characteristics of adjacent segments are input into the trajectory change direction extraction sub-channel, and the direction deviation of the wind speed and power change characteristics is calculated to form a direction change sequence. The direction change sequence is quantized into a discrete direction status code, which is used to represent different direction types such as upward surge, downward decay, and offset change. The trajectory change intensity extraction sub-channel is activated, and the wind speed gradient, wind direction shift rate and power fluctuation amplitude of adjacent segments are calculated to form a change amplitude index. The change amplitude index is then classified according to a preset quantization rule to generate a change intensity level sequence. The change intensity level sequence is used to represent the magnitude of change between different segments. Activate the trajectory change persistence extraction sub-channel, count the duration of each continuous change direction and intensity state in time to form a duration sequence, perform smoothing filtering and outlier correction on the duration sequence, and establish a corrected duration sequence; The discrete direction status code, the change intensity level sequence, and the correction duration sequence are fused in chronological order to form a trajectory transfer description sequence.

4. The method for correcting power prediction errors for non-stationary wind power data as described in claim 1, characterized in that, Using the aforementioned coupling relationship, a classification analysis of the prediction error change characteristics corresponding to different non-stationary evolution trajectory sequences is performed to construct a set of error evolution patterns, including: Based on the non-stationary evolution trajectory sequence, trajectory change events between adjacent time evolution segments are identified, including state rise events, state decay events, and state oscillation events; For each trajectory change event, the error response process within the corresponding time window is extracted from the prediction error sequence, and error response features characterizing the response characteristics of the error to trajectory changes are calculated. The error response features include error response slope, error response delay, and error oscillation amplitude. The trajectory-error response feature vector is constructed using the error response features, and a structured classification analysis is performed on the prediction error change patterns corresponding to different non-stationary evolution trajectories based on the trajectory-error response feature vector. A set of error evolution patterns is constructed based on the results of the structured classification analysis.

5. The method for correcting power prediction errors for non-stationary wind power data as described in claim 4, characterized in that, Based on the trajectory-error response feature vector, a structured classification analysis is performed on the prediction error change patterns corresponding to different non-stationary evolution trajectories, including: Calculate the structural consistency metric value for each trajectory-error response feature vector in the candidate prediction error change pattern. The structural consistency metric value is used to quantify the consistency level of the corresponding candidate prediction error change pattern in trajectory direction, change rate and error response amplitude. The error response path of the candidate prediction error change pattern is mapped to a topological structure, and a topological stability metric is calculated based on the continuity, repeatability, and fluctuation amplitude stability of the topological structure. The topological stability metric is used to quantify the reliability of the corresponding candidate prediction error change pattern in historical evolution. After normalizing the structural consistency metric and topology stability metric, dynamic weight fusion is performed to generate a built-in trust index. A set of error evolution patterns is constructed based on the built-in trust index.

6. The method for correcting power prediction errors for non-stationary wind power data as described in claim 5, characterized in that, The error evolution pattern set includes a dynamic updater, which performs additions, deletions, weight adjustments, and updates of the built-in trust index of existing error evolution patterns based on the non-stationary evolution trajectory and actual power sequence obtained at the new prediction time.

7. The method for correcting power prediction errors for non-stationary wind power data as described in claim 1, characterized in that, Using similarity matching results to perform error correction management also includes: Determine whether the similarity matching result meets the preset matching threshold; If the similarity matching result does not meet the preset matching threshold, then a balance coefficient is configured based on the similarity matching result and the weather forecast confidence level. Based on the balance coefficient, a balance prediction analysis is performed on the meteorological forecast results and similarity matching results to establish a corrected prediction result.

8. The method for correcting power prediction errors for non-stationary wind power data as described in claim 1, characterized in that, Using similarity matching results to perform error correction management also includes: A data correction report is generated based on the error correction management data, and the data correction report is encrypted and stored. Visualize the encrypted storage feedback.

9. The method for correcting power prediction errors for non-stationary wind power data as described in claim 8, characterized in that, The data correction report is subjected to correction verification to generate a verification dataset. The verification dataset is then used to configure a feedback database, which is used for verification feedback management of power prediction error correction.

10. A power prediction error correction system for non-stationary wind power data, characterized in that, The system is used to execute a power prediction error correction method for non-stationary wind power data as described in any one of claims 1-9, the system comprising: The data acquisition module is used to acquire meteorological monitoring data, historical power data, and predicted power sequences output by the power prediction model for wind farms in continuous time series. The dynamic segmentation module is used to perform time-series change analysis on the meteorological monitoring data and historical power data. It dynamically segments the continuous time series according to the wind speed change gradient, wind direction shift rate, and power fluctuation intensity to form multiple time evolution segments that reflect the local operating status. The trajectory sequence construction module is used to extract segment evolution descriptors representing the direction of change, rate of change, and distribution of fluctuation energy for each time evolution segment, and to construct a non-stationary evolution trajectory sequence of the wind field operating state in chronological order; The comparative analysis module is used to compare and analyze the predicted power sequence with the actual power sequence at the corresponding time, obtain the prediction error sequence, and establish the coupling relationship between the prediction error sequence and the non-stationary evolution trajectory sequence. The classification analysis module is used to perform classification analysis of the prediction error change characteristics corresponding to different non-stationary evolution trajectory sequences using the coupling relationship, and to construct a set of error evolution patterns. The error correction management module is used to perform similarity matching of the error evolution pattern set based on the non-stationary evolution trajectory of the current node at a new prediction time, and to perform error correction management using the similarity matching results.

Images