A photovoltaic power generation power real-time prediction method based on weather scene recognition

Abstract

The application discloses a photovoltaic power generation power real-time prediction method based on weather scene recognition, relates to the photovoltaic power prediction technical field, and comprises the following steps: calculating the data source credibility score of photovoltaic power data and meteorological observation data by using a time sequence characteristic vector, combining each data source credibility score with the time sequence characteristic vector into a characteristic score vector, inputting the characteristic score vector into a power regression network, extracting the time dependence relationship of a historical power sequence and local meteorological dynamics, and generating a time sequence dependence characteristic sequence.The photovoltaic power prediction residual error is combined with a preliminary confidence interval, and is used to form a closed loop mapping on the combined prediction input data, dynamically responds to the power change and local meteorological fluctuation of each time step, fuses the historical power trend, meteorological change and weather scene characteristics, and outputs a corrected power prediction value and a confidence interval, so that the short-term prediction accuracy and stability are improved.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power prediction technology, and in particular to a method for real-time prediction of photovoltaic power generation based on weather scene recognition. Background Technology

[0002] As the scale of photovoltaic power generation expands, the role of real-time power forecasting in grid dispatching and energy management is enhanced. Photovoltaic power is complexly affected by meteorological factors such as solar radiation, cloud cover, temperature, and wind speed. Its time-varying characteristics make it difficult for traditional statistical forecasting methods to meet the needs of high-precision short-term forecasting. The academic and engineering communities have introduced time series analysis, machine learning, and neural network methods to combine historical power data with local meteorological data, extract time-dependent features and dynamic change patterns, and achieve more accurate power forecasting.

[0003] Existing photovoltaic power forecasting technologies have shortcomings in short-term power fluctuation response and multi-source data reliability processing. Most methods rely only on historical power and a single meteorological variable, and cannot comprehensively evaluate the reliability and time-series dynamic characteristics of multi-source data, resulting in large errors in forecasts under rapid weather changes or abnormal conditions. By calculating the reliability scores of photovoltaic power and meteorological data, extracting time-series dependent features and identifying weather scenarios, dynamic closed-loop adjustment of forecast input can be achieved, thereby improving short-term forecast accuracy and adaptability under rapidly changing weather conditions. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a real-time photovoltaic power generation prediction method based on weather scene recognition to solve the problem of insufficient prediction accuracy of short-term photovoltaic power generation under conditions of dynamic changes in multi-source data and rapid switching of weather scenes.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a real-time photovoltaic power generation prediction method based on weather scene recognition, comprising: collecting real-time power data of photovoltaic power plants and local meteorological observation data, arranging them in chronological order to form a time-series data sequence, and generating a time-series feature vector through feature encoding; calculating the data source credibility score of photovoltaic power data and meteorological observation data using the time-series feature vector, combining the credibility score of each data source with the time-series feature vector to form a feature score vector, and inputting the feature score vector into a power regression network to extract the time dependency relationship between historical power sequences and local meteorological dynamics, generating a time-series dependency feature sequence; identifying the current weather scene category based on the time-series dependency feature sequence, and selecting based on the identification result. The corresponding weather scenario prediction model combines weather scenario categories with time-dependent features to form joint prediction input data. A fully connected mapping is then performed using the weather scenario prediction model to generate preliminary power prediction values ​​and preliminary confidence intervals. Based on a comparison between the preliminary power prediction values ​​and the actual measured photovoltaic power values ​​from the previous time step, prediction residuals are calculated. Adjustment signals for the prediction input features of this time step are generated based on the prediction residuals and the preliminary confidence intervals. These adjustment signals are then scheduled to the joint prediction input data to obtain adjusted joint prediction input data. Finally, the adjusted joint prediction input data is input into the selected weather scenario prediction model for a fully connected mapping to generate corrected power prediction values ​​and corrected confidence intervals.

[0008] As a preferred embodiment of the real-time photovoltaic power generation prediction method based on weather scene recognition described in this invention, the specific steps for generating the time-series feature vector are as follows:

[0009] Based on the operating status of the photovoltaic power station and surrounding meteorological observation information, power data and meteorological data at each time step are acquired, sorted and verified by time tags, and a continuous time series data set is generated.

[0010] Consistency checks and anomaly identification are performed on continuous time series datasets. By filling in missing values, removing outliers, and verifying the time label order, time series data that continuously reflects power fluctuations and meteorological dynamics are obtained.

[0011] Dynamic features are extracted from time series data by capturing power change patterns and meteorological fluctuations, and time series feature vectors are generated through feature encoding.

[0012] As a preferred embodiment of the real-time photovoltaic power generation prediction method based on weather scene recognition described in this invention, the combination is a feature scoring vector, and the specific steps are as follows:

[0013] Based on time-series feature vectors, the recent dynamics of photovoltaic power and local meteorological observation data are obtained. By analyzing the power change amplitude and meteorological fluctuations at the current time step, the credibility of each data source is calculated and a corresponding credibility score is generated.

[0014] Historical prediction errors of photovoltaic power and the variation range of meteorological observation data are extracted to determine the credibility score of each data source. When the power variation range and meteorological fluctuations in the current time step exceed the variation range of the previous time step, they are combined into a feature score vector.

[0015] As a preferred embodiment of the real-time photovoltaic power generation prediction method based on weather scene recognition described in this invention, the specific steps for generating the time-dependent feature sequence are as follows:

[0016] Based on the feature scoring vector, historical photovoltaic power data and local meteorological information are extracted, and then continuously arranged and correlated in time order to obtain an ordered dataset.

[0017] Information is extracted and dependencies are analyzed from historical power sequences and local meteorological dynamics in an ordered dataset, and short-term and long-term change characteristics are captured to generate a dependency dataset.

[0018] The dependent dataset is input into the power regression network for feature integration. In the power regression network, the influence of different features is adjusted according to the power change amplitude and meteorological fluctuations. Historical photovoltaic power data is fused with local meteorological information to generate a time-dependent feature sequence.

[0019] The power regression network is a feature integration submodule of the selected weather scenario prediction model. It is used to integrate and map the time step features, historical power and local meteorological information in the joint prediction input data to generate a time-dependent feature sequence.

[0020] As a preferred embodiment of the real-time photovoltaic power generation prediction method based on weather scene recognition described in this invention, the specific steps for generating the preliminary power prediction value and the preliminary confidence interval are as follows:

[0021] Extract photovoltaic power change information and local meteorological dynamics at continuous time steps from the time-dependent feature sequence, calculate the statistical average value of power change, and obtain the current weather feature data set;

[0022] Based on the comparison of the current weather feature data set with the power change trend and meteorological dynamics of the previous time step, and combined with the weather scene classification standard, the current weather scene category is determined, and the determined weather scene category is combined with the time-dependent features to form joint prediction input data.

[0023] The joint forecast input data is input into the selected weather scenario forecast model and fully connected to generate preliminary power forecast values ​​and preliminary confidence intervals.

[0024] As a preferred embodiment of the real-time photovoltaic power generation prediction method based on weather scene recognition described in this invention, the step of inputting the joint prediction input data into the selected weather scene prediction model for fully connected mapping to generate preliminary power prediction values ​​and preliminary confidence intervals includes, based on the time-dependent feature sequence, integrating and transforming the time step features and historical power information in the joint prediction input data to generate a set of time-series feature vectors of power change trends and dynamic relationships.

[0025] When there are abnormal fluctuations in the amplitude of the mapping vector of the time series feature vector set, the joint prediction input data is numerically corrected to obtain the adjusted joint prediction input data;

[0026] The adjusted joint forecast input data is then input into the selected weather scenario prediction model for fully connected mapping to generate preliminary power forecast values ​​and preliminary confidence intervals.

[0027] As a preferred embodiment of the real-time photovoltaic power generation prediction method based on weather scene recognition described in this invention, the weather scene prediction model is specifically constructed as follows:

[0028] A weather scenario prediction model is constructed based on an input fusion layer, a time-dependent mapping layer, and a prediction output and closed-loop correction layer.

[0029] The input fusion layer integrates time-step photovoltaic power with local meteorological features based on joint prediction input data, analyzes the power change amplitude and meteorological dynamics, and generates a fused integrated input feature vector by combining the power change amplitude and meteorological dynamic range.

[0030] The time-dependent mapping layer uses the fused integrated input feature vector to perform time-dependent relationship analysis on photovoltaic power and local meteorological information at continuous time steps, capture short-term and long-term change trends, and form a set of time-series feature vectors.

[0031] The predicted output and closed-loop correction layer, based on the time series feature vector set, integrates and maps the time series feature vectors of each time step to obtain a preliminary estimate of the photovoltaic power at the current time step, and integrates them to generate a corrected power prediction value and a corrected confidence interval.

[0032] As a preferred embodiment of the real-time photovoltaic power generation prediction method based on weather scene recognition described in this invention, the specific steps for generating the time-step prediction input feature adjustment signal are as follows:

[0033] Based on the preliminary power prediction and preliminary confidence interval of this time step, the actual photovoltaic power measurement value of the previous time step is obtained. By comparing the difference between the current prediction value and the actual measurement value of the previous time step, the power deviation between adjacent time steps is calculated.

[0034] By utilizing the power deviation between adjacent time steps and the upper and lower bounds of the initial confidence interval of the current time step, the deviation of the prediction residual relative to the confidence interval is calculated to generate the prediction input feature adjustment signal for the current time step.

[0035] As a preferred embodiment of the real-time photovoltaic power generation prediction method based on weather scene recognition described in this invention, the step of generating the prediction input feature adjustment signal for the current time step by calculating the deviation of the prediction residual relative to the confidence interval by utilizing the power deviation of adjacent time steps and the upper and lower bounds of the preliminary confidence interval of the current time step includes: combining the difference between the preliminary power prediction value and the actual power of the previous time step to calculate the prediction residual of the current time step, and combining the upper and lower bounds of the preliminary confidence interval of the current time step to form a sequence of deviation ratios of the residual relative to the confidence interval;

[0036] By combining the residual deviation ratio with the predicted residual trend of the previous time step, the numerical adjustment magnitude of the predicted input feature is calculated, and the predicted input feature adjustment signal is generated.

[0037] The predicted input feature adjustment amount of the predicted input feature adjustment signal is integrated with the current joint predicted input data to form the predicted input feature adjustment signal for this time step.

[0038] As a preferred embodiment of the real-time photovoltaic power generation prediction method based on weather scene recognition described in this invention, the specific steps for generating the corrected power prediction value and the corrected confidence interval are as follows:

[0039] The predicted input feature adjustment signal of this time step is numerically corrected and combined with the current joint predicted input data to form the closed-loop input data of this time step;

[0040] The actual measured value of photovoltaic power and local meteorological dynamics are fused and statistically analyzed using the closed-loop input data of this time step to generate fused power information.

[0041] The fused power information is input into the selected weather scenario prediction model and fully connected mapping is performed. The power prediction value and corresponding confidence interval for each time step are calculated through the mapping, and the corrected power prediction value and corrected confidence interval are generated.

[0042] The beneficial effects of this invention are as follows: it combines the photovoltaic power prediction residual with the preliminary confidence interval and applies it to the joint prediction input data to form a closed-loop mapping, dynamically responding to power changes and local meteorological fluctuations at each time step, and integrating historical power trends, meteorological changes and weather scene characteristics to output corrected power prediction values ​​and confidence intervals, thereby improving the accuracy and stability of short-term predictions. Attached Figure Description

[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a flowchart of a method for real-time prediction of photovoltaic power generation based on weather scene recognition.

[0045] Figure 2 This is a flowchart for time-series data processing and dynamic feature generation.

[0046] Figure 3 A flowchart for time-series feature scoring and historical sequence generation.

[0047] Figure 4 A flowchart for weather-driven time-series feature correction and power prediction. Detailed Implementation

[0048] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0049] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0050] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0051] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for real-time prediction of photovoltaic power generation based on weather scene recognition, including the following steps:

[0052] S1: Collect real-time power data of photovoltaic power plants and local meteorological observation data, arrange them in chronological order to form a time-series data sequence, and generate a time-series feature vector through feature encoding.

[0053] S1.1: Based on the operating status of the photovoltaic power station and surrounding meteorological observation information, acquire power data and meteorological data at each time step, sort and verify them through time tags, and generate a continuous time series data set.

[0054] Furthermore, during the operation of the photovoltaic power station, the operating status of the photovoltaic power station is monitored in real time, and photovoltaic power data at each time step is collected. At the same time, local meteorological observation data (including solar radiation, temperature, wind speed, wind direction and cloud cover) are acquired. The power data and local meteorological observation data at each time step are arranged continuously in chronological order through time tags. The data at each time step are checked through time tags, and data points that are discontinuous or duplicated in time tags are eliminated, forming a time series data set that continuously reflects the fluctuation of photovoltaic power and the changes in local meteorological observation data.

[0055] It should be noted that the local meteorological observation data is a collection of meteorological information collected around the photovoltaic power station, including solar radiation, temperature, wind speed, wind direction and cloud cover. Each indicator is a meteorological variable, which is processed or extracted to form local meteorological information. As it changes over time, it forms local meteorological observation data changes or local meteorological observation data sequences.

[0056] S1.2: Perform consistency checks and anomaly identification on the continuous time series data set. By filling in missing values, removing outliers, and verifying the time label order, time series data that continuously reflects power fluctuations and meteorological dynamics can be obtained.

[0057] Furthermore, based on a unified time step, the power and local meteorological observation data timestamps are aligned one by one and cross-compared. After marking the single-source data nodes, they are arranged according to the time axis to form an initial continuous time series framework. The time series framework is traversed to detect gaps. When 1 to 3 consecutive time steps are identified as missing, the numerical difference and time interval of the data before and after the missing point are extracted. By referring to the data change trends of adjacent time steps, the estimated value of the missing time step is calculated and filled in the original position to restore the continuity of the curve. When more than 3 consecutive time steps are missing or more than 10% of the total length of the sampling window, the gap mark is retained to avoid distortion of the data sequence.

[0058] The anomaly detection of the filled sequence is performed by calculating the slope of change of the current time step and the adjacent steps before and after, and comparing it with the average change trend of multiple consecutive steps. When the change of photovoltaic power or local meteorological observation data deviates significantly from the trend and shows abrupt change, it is determined whether it exceeds the reasonable fluctuation boundary by referring to the average value and fluctuation range, quantile interval, rated power change rate or physical change upper limit of each meteorological variable in the same historical time period.

[0059] If the data exceeds the limit, it is identified as noise and removed. The normal data flow before and after the data is used to smooth the transition and eliminate interference. The sequence timestamps are rescanned, and the time index is verified to be strictly increasing and without repetition or reverse order. The breakpoints caused by the removal are closed, and the output time stamps continuously reflect the time series of photovoltaic power fluctuations and changes in local meteorological observation data.

[0060] S1.3: Extract dynamic features from time series data by capturing power change patterns and meteorological fluctuations, and generate time series feature vectors through feature encoding.

[0061] Furthermore, the time series data is strictly sorted and verified according to time labels, missing data in the time series is estimated and filled, and abnormal fluctuations are identified and eliminated to form a complete time series that continuously reflects power fluctuations and meteorological changes. The change range of photovoltaic power is analyzed step by step, and the power at the current time step is compared with the previous time step to determine whether the power is rising, falling, or stable.

[0062] Statistical analysis is performed on the variation range and short-term fluctuations of local meteorological observation data. Data points that are obviously abnormal or exceed the historical variation range are marked. The photovoltaic power variation trend of each time step is compared and integrated with the local meteorological variation pattern of the corresponding time step. By observing the rise or fall of power and the variation range and direction of meteorological variables, dynamic feature vectors for each time step are extracted. The extracted dynamic feature vectors of each time step are feature-coded, and the power and meteorological dynamic features of each time step are converted into a unified dimension while retaining the time dependence and change pattern. Power prediction is then performed. The coded features of all time steps are integrated in chronological order to generate a complete time-series feature vector.

[0063] It should be noted that feature encoding refers to converting the dynamic feature vector generated at each time step into a coded feature of a unified dimension. The input is the photovoltaic power change trend and the corresponding local meteorological variable change information at each time step. The input features are then processed to map the multidimensional original features into a standardized coded vector, eliminating the differences in dimensions and scales between different features.

[0064] S2: Calculate the data source credibility score of photovoltaic power data and meteorological observation data using time-series feature vectors, combine the credibility scores of each data source with the time-series feature vectors to form a feature score vector, and input the feature score vector into the power regression network to extract the time dependence relationship between historical power sequences and local meteorological dynamics, and generate a time-series dependent feature sequence.

[0065] S2.1: Based on time-series feature vectors, obtain the recent dynamics of photovoltaic power and local meteorological observation data. By analyzing the power change amplitude and meteorological fluctuations at the current time step, calculate the credibility of each data source and generate the corresponding credibility score.

[0066] Furthermore, based on the time-series feature vector, the photovoltaic power data and local meteorological observation data of consecutive time steps are checked and sorted according to the time label. The power data of each time step and the corresponding local meteorological observation data are matched in time. A certain window length is traced back from the current time step to extract the recent dynamic information of photovoltaic power data and local meteorological observation data, including the power change range, short-term trend and the increase, decrease and fluctuation range of local meteorological observation data. The traceback window length is set to 5 to 30 time steps or 5 to 60 minutes.

[0067] The photovoltaic power at the current time step is compared with the change in the previous time step. If the change in the current time step is significantly greater than the change in the previous time step or significantly deviates from the historical average change range, the power change is judged to be abnormal. Otherwise, the power change trend is considered to be continuous. For each meteorological variable, the change in the current time step is calculated with the change in the previous time step and compared with the historical average change. If it exceeds the change in the previous time step or shows a sudden jump, it is marked as abnormal. Otherwise, it is considered to conform to a continuous trend.

[0068] In the process of photovoltaic power prediction, the power change amplitude at each time step is matched and analyzed with the fluctuation amplitude of local meteorological observation data and the corresponding time series feature vector to quantify the reliability of each data source.

[0069] The credibility score of each data source is defined between 0 and 1, indicating the level of reliability. When the data missing rate at a certain time step is low, the historical prediction error is small, and the changes in power and local meteorological observation data match the recent trend and are consistent with the trend of continuous time steps, the data source obtains a high credibility score close to 1. When the data at a certain time step is missing, the historical prediction error is large, and the fluctuations in power or local meteorological observation data deviate significantly from the recent trend or are inconsistent with the trend of continuous time steps, the data source obtains a low credibility score.

[0070] For cases falling between high and low, specific score values ​​can be determined based on missing rate, historical prediction error, degree of deviation in change magnitude, and trend consistency. The reliability scores of photovoltaic power data and local meteorological observation data at the current time step are integrated to form the reliability score corresponding to the time step.

[0071] S2.2: Extract the historical prediction error of photovoltaic power and the variation range of meteorological observation data, determine the credibility score of each data source, and combine the power variation range and meteorological fluctuations of the current time step into a feature score vector when they exceed the variation range of the previous time step.

[0072] Furthermore, the historical prediction error of photovoltaic power and the change range of meteorological observation data at each time step are extracted to determine the credibility score of each data source of photovoltaic power data and meteorological observation data.

[0073] During the time-step analysis, the change in photovoltaic power at the current time step is compared with the change at the previous time step to assess the volatility of the data source and to compare whether exceeding the previous time step is insufficient to quantify the credibility.

[0074] To improve accuracy, the change in photovoltaic power at the current time step is compared with the change at the previous time step and the average change and fluctuation range for the same historical time period. When the current change is significantly higher than the previous time step and exceeds the historical fluctuation range, abnormal fluctuation exists, and the reliability score of the photovoltaic power data source is reduced. The same method is used to compare each variable in local meteorological observation data. When the change exceeds the previous time step and deviates from the historical fluctuation range, the reliability score of the corresponding meteorological data source is also reduced. The reduction in reliability score can be determined in segments according to the degree of abnormal fluctuation: a small reduction is given for slight deviations, and a larger reduction is given for significant deviations, while maintaining a range of 0 to 1.

[0075] After independently determining the power and meteorological data sources, the credibility scores of each data source at the current time step are combined with the time series feature vectors of the corresponding time steps. By adjusting and correcting the feature values ​​of each time step, the feature vectors are made to numerically reflect the time dependence of power and meteorology of each data source, and combined into a feature score vector.

[0076] S2.3: Based on the feature scoring vector, extract historical photovoltaic power data and local meteorological information, and perform continuous arrangement and time sequence association to obtain an ordered data set.

[0077] Furthermore, based on historical photovoltaic power data and local meteorological information extracted from feature scoring vectors, the data of continuous time steps are scanned step by step according to time labels to check for time jumps. If a jump is found, the time period is marked as missing. It is also checked for data points with duplicate time labels. For duplicate time steps, the data points that are most consistent with the previous and subsequent trends are retained and duplicate values ​​are removed to eliminate redundancy.

[0078] For the marked missing time steps, the difference between the data of the preceding and following time steps is estimated. A linear interpolation method is used to calculate the approximate value of the missing time step by the power value or meteorological observation value of the preceding and following time steps according to the time ratio. The entire dataset is then sorted again according to the time label, with the time steps increasing and the data of each time step completely corresponding to photovoltaic power and meteorological information. The power data of each time step is gradually correlated with the corresponding meteorological observation information to form an ordered dataset that reflects historical power changes and local meteorological dynamics.

[0079] Specifically, linear interpolation refers to restoring the continuity of a time series by using the effective observations of adjacent time steps before and after the missing point and estimating the value of the missing time step based on the linear trend between the two points when encountering missing data.

[0080] Determine the data values ​​of the preceding and following valid time steps for missing data. Calculate the missing point values ​​by linearly relating the preceding and following data values ​​according to the interval ratio between time steps. By filling in the missing data, the continuity of the photovoltaic power or meteorological variable series is maintained. At the same time, for missing data with a large span, select to retain empty position markers to prevent series distortion or the introduction of excessive errors.

[0081] S2.4: Extract information and analyze dependencies from historical power sequences and local meteorological dynamics in the ordered dataset, capture short-term and long-term change characteristics, and generate a dependency dataset.

[0082] Furthermore, the historical photovoltaic power sequence and local meteorological dynamics in the ordered dataset are arranged in chronological order so that the data at each time step completely correspond to the photovoltaic power and various meteorological variables, forming a continuous historical dataset. Information is extracted from the historical power sequence and meteorological dynamics at each time step, and the power change and meteorological variable increase / decrease at the current time step and the previous time step are read, and the trend of change is recorded to reflect short-term dynamic changes.

[0083] If the change in photovoltaic power or the magnitude of change in local meteorological observation data over a continuous period of time increases or decreases, the historical power sequence is judged to show a clear upward or downward trend, which is a long-term change characteristic. If the fluctuation of the magnitude of change in power or local meteorological observation data is small over a continuous period of time, it is judged to be a short-term stable characteristic.

[0084] The short-term variation features extracted from each time step are integrated and correlated with the determined long-term variation patterns to form a feature description reflecting the dependency relationship between photovoltaic power and local meteorological observation data. Historical power features and local meteorological observation data dependency information are generated for each time step. During the integration process, if the short-term variation direction or magnitude of a time step is significantly inconsistent with the overall variation direction of the long-term trend, or if the variation magnitude deviates significantly from the long-term variation pattern, the time step is marked as an abnormal or unstable time step. By performing continuity and trend matching analysis on the dependency features of all time steps, time steps with large deviations are identified. The dependency data of all time steps are integrated in chronological order to form a complete dependency data set.

[0085] S2.5: Input the dependent data set into the power regression network for feature integration. In the power regression network, adjust the influence of different features according to the power change amplitude and meteorological fluctuations. Integrate historical photovoltaic power data with local meteorological information to generate a time-dependent feature sequence.

[0086] Furthermore, the photovoltaic power sequence and meteorological dynamics at each time step in the dependent dataset are arranged and verified according to time labels, so that the power data at each time step corresponds to the meteorological variables, forming a continuous input dataset.

[0087] The power data for each time step is input into the power regression network. In the power regression network, the variation range of photovoltaic power and the fluctuation range of each meteorological variable are calculated for each time step. The representation of the characteristics of each time step is adjusted according to the variation range. For time steps with larger power changes, the power characteristics are more prominent in the integration. For time steps with larger fluctuations in meteorological variables, the meteorological characteristics are more prominent in the integration. By analyzing the fusion characteristics of continuous time steps, the power regression network arranges and combines the fusion characteristics of all time steps in chronological order to generate a complete time-series dependent feature sequence.

[0088] The power regression network is a feature integration submodule embedded in the weather scenario prediction model, used to extract the time dependence of historical power sequences and local meteorological dynamics. It corresponds to the input fusion layer and time dependence mapping layer of the weather scenario prediction model. The input is a dependency data set, which includes historical photovoltaic power data and local meteorological information arranged in chronological order, as well as dependency features. The power regression network calculates the change amplitude of photovoltaic power and the fluctuation amplitude of each meteorological variable at each time step, and dynamically adjusts the feature representation of different time steps according to the change amplitude. This makes the power features more prominent in the integration of time steps with large power changes and the meteorological features more prominent in the integration of time steps with large fluctuations in meteorological variables. It performs time series correlation analysis on the fused features of continuous time steps, captures the short-term fluctuation response and long-term trend dependence between historical photovoltaic power and local meteorological dynamics, and outputs a time series dependency feature sequence reflecting the time dependence relationship to the prediction output and closed-loop correction layer of the weather scenario prediction model.

[0089] S3: Identify the current weather scene category based on the time-dependent feature sequence, and select the corresponding weather scene prediction model based on the identification result. Combine the weather scene category with the time-dependent features to form joint prediction input data. Use the weather scene prediction model to perform fully connected mapping to generate preliminary power prediction values ​​and preliminary confidence intervals. The weather scene prediction model is constructed based on the input fusion layer, the time-dependent mapping layer, and the prediction output and closed-loop correction layer.

[0090] S3.1: Input fusion layer. Based on the joint prediction input data, it integrates the time-step photovoltaic power with local meteorological features, analyzes the power change amplitude and meteorological dynamics, and generates the fused integrated input feature vector by combining the power change amplitude and meteorological dynamic range.

[0091] S3.2: Time-dependent mapping layer. Using the fused integrated input feature vector, it performs time-dependent analysis on photovoltaic power and local meteorological information at continuous time steps, captures short-term and long-term trends, and forms a set of time-series feature vectors.

[0092] S3.3: Predictive output and closed-loop correction layer. Based on the time series feature vector set, the time series feature vectors of each time step are integrated and mapped to obtain the preliminary estimate of the photovoltaic power of the current time step, and integrated to generate the corrected power prediction value and the corrected confidence interval.

[0093] The power regression network is a feature integration submodule embedded in the weather scenario prediction model, used to extract the time dependence of historical power sequences and local meteorological observation data. It corresponds to the input fusion layer and time dependence mapping layer of the weather scenario prediction model.

[0094] The input is a dependency dataset, which includes historical photovoltaic power data arranged in chronological order, local meteorological observation data, and dependency features. During the integration process, the power regression network dynamically adjusts the degree of feature influence at different time steps based on the variation range of photovoltaic power and the fluctuation range of local meteorological observation data at each time step. That is, in time steps with large variations in photovoltaic power, the power features contribute more to the final fusion result in the integrated representation; in time steps with large fluctuations in local meteorological observation data, the meteorological features contribute more to the fusion result in the integrated representation. The determination of the degree of feature influence is based on the variation range of historical power, the fluctuation range of local meteorological observation data, and the trend continuity of continuous time steps. Through normalization, a feature influence coefficient with a value range between 0 and 1 is obtained, ensuring that the relative contributions of different time steps and different types of features are reasonably reflected in the time-series fusion analysis.

[0095] The power regression network performs time-series correlation analysis on the fusion characteristics of continuous time steps, captures the short-term fluctuation response and long-term trend dependence between historical photovoltaic power and local meteorological observation data, and outputs the generated time-series dependence feature sequence reflecting the time dependence relationship to the prediction output and closed-loop correction layer of the weather scenario prediction model.

[0096] S3.4: Extract photovoltaic power change information and local meteorological dynamics for continuous time steps from the time-dependent feature sequence, calculate the statistical average value of power change, and obtain the current weather feature data set.

[0097] Furthermore, the photovoltaic power data and corresponding meteorological observation information for each time step are read sequentially according to time order, and the power data for each time step is matched with the meteorological data. The numerical difference of power between each consecutive time step is calculated for the photovoltaic power sequence to obtain the amplitude of power change at each time step. The power change amplitude sequence is then formed by organizing the data in chronological order. For each meteorological variable, the change amount for each consecutive time step is calculated sequentially. The change values ​​at each time step are recorded in sequence and organized into a local meteorological dynamic sequence.

[0098] After obtaining the power change amplitude sequence and the meteorological dynamic sequence, the photovoltaic power change amplitude sequence is averaged to obtain the overall level of power change within the entire time window. For the local meteorological dynamic sequence, the average change is calculated for each meteorological variable to reflect the overall change of meteorological conditions within the time window. The average change value of photovoltaic power is combined with the average change value of each meteorological variable to form a complete set of current weather characteristic data.

[0099] S3.5: Based on the current weather feature data set and the power change trend and meteorological dynamics of the previous time step, and combined with the weather scene classification standard, determine the current weather scene category, and combine the determined weather scene category with the time-dependent features to form joint prediction input data.

[0100] Furthermore, photovoltaic power change information and corresponding local meteorological observation information for consecutive time steps are extracted sequentially from the current weather feature data set, so that the power data of each time step is completely correlated with meteorological variables. The power change amplitude of the current time step is compared with that of the previous time step, and the change amplitude of local meteorological observation data of each time step is compared to clarify the power change trend and the direction and magnitude of meteorological condition changes.

[0101] When determining the current weather scenario category, the changing trends of photovoltaic power and local meteorological observation data at the current time step are analyzed. If the photovoltaic power at the current time step continues to rise or fall compared to the previous time step, and the magnitude of the change significantly exceeds the average change level of the same period in recent history, then the power change trend is considered significant; otherwise, it is considered stable. The variables of local meteorological observation data are judged accordingly. If the change magnitude of solar radiation, cloud cover, or meteorological variables is significantly higher than that of the previous time step or exceeds the recent historical fluctuation range, then the meteorological change is considered significant; otherwise, it is considered stable.

[0102] For example, a situation where photovoltaic power continuously increases at the current time step and local meteorological observation data shows significant changes can be used as an example to show that this situation usually occurs in sunny or cloudy weather, while a situation where the power is stable and the meteorological variables change little indicates that it usually occurs in overcast or stable weather.

[0103] The photovoltaic power change trend, local meteorological observation data change magnitude, and historical change patterns at the current time step are used as inputs to the weather scene classification model. The model outputs the probability of each possible weather category, and the category with the highest probability is selected as the deterministic weather scene category for the current time step. After the determination, the weather scene category is combined with the temporal dependency features corresponding to the current time step, and integrated in chronological order to form complete joint prediction input data.

[0104] S3.6: Based on the time-dependent feature sequence, information integration and feature transformation are performed on the time step features and historical power information in the joint prediction input data to generate a set of time-series feature vectors of power change trends and dynamic relationships.

[0105] Furthermore, the features and corresponding historical photovoltaic power information of each time step in the joint forecast input data are arranged in chronological order, so that the power value, meteorological observation value and feature vector of each time step are completely corresponding, forming a continuous time series input. The photovoltaic power change trend information, including power increase, decrease or stable mode, is extracted for each time step in turn, and the change amplitude and direction of local meteorological variables such as solar radiation, temperature, wind speed, wind direction and cloud cover are extracted for the corresponding time step.

[0106] The photovoltaic power change features extracted at each time step are combined with the corresponding meteorological dynamic features to form a temporary feature vector. During the combination process, the magnitude, direction and trend of the changes in photovoltaic power and meteorological variables are encoded simultaneously, so that the temporary vector of each time step can reflect the dependence of power on meteorological conditions. The temporary feature vectors of continuous time steps are integrated in chronological order to maintain temporal continuity and historical dependence, while retaining the response pattern of power to changes in meteorological variables at each time step, thus generating a complete set of time-series feature vectors.

[0107] The formula for integrating the interaction characteristics of power and meteorology is:

[0108] ;

[0109] In the process of integrating the interactive features of photovoltaic power and local meteorological observation data, the variation amplitude features of photovoltaic power, solar radiation, temperature, wind speed, cloud cover, and the residual of photovoltaic power in the previous time step are extracted and normalized to dimensionless features between 0 and 1. Through normalization, the dimensions of each variable are consistent. This indicates the normalized variation range of photovoltaic power. This represents the characteristic of the normalized variation range of solar radiation. This represents the normalized temperature variation range. This represents the normalized variation range of wind speed. This indicates the normalized variation range of cloud cover. This represents the normalized photovoltaic power residual from the previous time step. This represents the comprehensive response feature quantity after the interaction between the power at the current time step and meteorological variables, and is used to generate time-dependent feature vectors.

[0110] S3.7: When there are abnormal fluctuations in the amplitude of the mapping vector of the time sequence feature vector set, the joint prediction input data is numerically corrected to obtain the adjusted joint prediction input data.

[0111] Furthermore, when abnormal fluctuations occur in the amplitude of the mapping vector of the time series feature vector set, the preliminary power prediction value of each time step is compared with the actual photovoltaic power measurement value of the previous time step, and the numerical difference is calculated. The numerical difference is called the power deviation between time steps, which reflects the deviation of power change between adjacent time steps. The power deviation between time steps is compared with the upper and lower bounds of the preliminary confidence interval of the current time step. By analyzing the degree of deviation relative to the confidence interval, it is determined whether the power change exceeds the normal fluctuation range and whether there is abnormal fluctuation.

[0112] After identifying abnormal fluctuations, for each time step feature in the joint prediction input data, numerical adjustments are made sequentially based on the power deviation between time steps (or the degree of deviation of the prediction residual) and the trend of the prediction residual change in the previous time step. If the deviation of the current time step is large and the direction is consistent with the deviation of the previous time step, the correction amount of the corresponding feature is increased; otherwise, the original value is reduced or kept. The magnitude of the adjustment amount can be determined by mapping the degree of deviation proportionally to the feature normalization scale and multiplying it by the adjustment step size coefficient. The adjustment step size coefficient generally ranges from 0.05 to 0.30. The joint prediction input features of each time step are modified to reflect the state of the current prediction deviation. After all time steps are adjusted, the corrected features are recombined to form the adjusted joint prediction input data.

[0113] It should be noted that the normal fluctuation range refers to the numerical interval used to determine whether the current prediction residual is abnormal, including three aspects: First, the typical intraday power fluctuation amplitude calculated based on the statistical characteristics of historical photovoltaic power series, that is, the average change and natural fluctuation range of historical observation data within the same time period; Second, considering the impact of short-term meteorological variable changes on power, including the normal fluctuations of solar radiation, temperature, wind speed and cloud cover within continuous time steps; Third, the upper and lower bounds of the confidence interval calculated in combination with the preliminary power prediction value of this time step, reflecting the allowable error range of the prediction itself.

[0114] When the prediction residual at the current time step falls within the normal fluctuation range, it indicates that the predicted value deviates from an acceptable level and is considered normal fluctuation; if the residual exceeds the normal fluctuation range, it is judged as abnormal fluctuation, and the input data for joint prediction needs to be corrected.

[0115] The mapping vector is a time-dependent mapping layer of a weather scenario prediction model that inputs a set of time-series feature vectors into the model. The model performs element-wise mapping and integration of the time-series feature vectors at each time step through fully connected mapping calculations. The mapping result is generated after capturing the time dependency between photovoltaic power characteristics and meteorological dynamics. The magnitude of the mapping result reflects the changes in the time step mapping result.

[0116] S3.8: Input the adjusted joint forecast input data back into the selected weather scenario forecast model for fully connected mapping to generate preliminary power forecast values ​​and preliminary confidence intervals.

[0117] Furthermore, the features and corresponding historical photovoltaic power information of each time step in the joint prediction input data are read sequentially according to time order. In the input layer of the weather scenario prediction model, the joint prediction input features and historical power information of each time step are integrated and transformed to form a time-series feature vector for each time step, which reflects the changing trend of photovoltaic power and its relationship with local meteorological dynamics. After the generation of the time-series feature vector set is completed, it is input into the fully connected mapping layer of the weather scenario prediction model for calculation step by step. The weather scenario prediction model performs element-wise mapping and combination operations on each vector to capture the continuity and dependence of power changes and meteorological dynamics between time steps. The results of the fully connected mapping are integrated to generate the preliminary power prediction value of the current time step. At the same time, the variation range and distribution of the power prediction results of each time step are statistically calculated to form the corresponding preliminary confidence interval.

[0118] S4: Based on the comparison between the preliminary power prediction value and the actual measured value of photovoltaic power in the previous time step, calculate the prediction residual, and adjust the signal according to the prediction residual and the preliminary confidence interval to generate the prediction input characteristics of this time step.

[0119] S4.1: Based on the preliminary power prediction value and preliminary confidence interval of this time step, obtain the actual measured value of photovoltaic power of the previous time step. By comparing the difference between the current prediction value and the actual measured value of the previous time step, calculate the power deviation between adjacent time steps.

[0120] Furthermore, the actual measured value of photovoltaic power in the previous time step is recorded, so that each measured value corresponds to the power state in the previous time step. The preliminary power prediction value of the current time step is compared with the actual measured value of the previous time step element by element. The numerical difference between the predicted value and the actual measured value is calculated for each corresponding time point, and the power deviation between adjacent time steps is obtained. The power deviation between adjacent time steps reflects the degree of deviation of the predicted power in the current time step from the actual power in the previous time step.

[0121] The judgment criteria are as follows: if the absolute value of the prediction residual is within the range of historical photovoltaic power variation and the confidence interval of the current time step, the prediction residual is considered to be a normal fluctuation; if the prediction residual exceeds the range of historical power variation or the confidence interval, it is judged as an abnormal deviation. The residuals of consecutive time steps are arranged in chronological order to form a complete prediction residual sequence, which continuously reflects the deviation of photovoltaic power prediction from the actual measured value at each time step, and is used as the power deviation of adjacent time steps of the current time step.

[0122] S4.2: Combine the difference between the preliminary power prediction and the actual power at the previous time step, calculate the prediction residual for the current time step, and combine it with the upper and lower bounds of the preliminary confidence interval for the current time step to form a sequence of the deviation ratio of the residual relative to the confidence interval.

[0123] Furthermore, the preliminary power prediction value of the current time step is compared element by element with the actual photovoltaic power measurement value of the previous time step. The difference between the predicted value of each time step and the actual measurement value of the previous time step is calculated to obtain the power deviation of the current time step and the adjacent time step. Each residual value represents the deviation of the predicted value of the time step from the actual power of the previous time step. Each residual value is compared with the upper and lower bounds of the preliminary confidence interval of the current time step. The quotient of the residual with the upper bound represents the positive deviation ratio, and the quotient of the residual with the lower bound represents the negative deviation ratio. The magnitude of the prediction deviation in the confidence interval is quantified to obtain the position and degree of the prediction residual of each time step in the confidence interval, and a complete sequence of the deviation ratio of the residual relative to the confidence interval is formed.

[0124] S4.3: Utilize the residual deviation ratio in conjunction with the residual trend predicted in the previous time step to calculate the numerical adjustment magnitude of the predicted input feature and generate the predicted input feature adjustment signal.

[0125] Furthermore, the prediction residuals of the current time step and the previous time step are obtained and compared over time. By calculating the numerical differences and the direction of change, the trend of residual change is determined, that is, whether the prediction deviation of the current time step is increasing, decreasing or remaining stable.

[0126] Analyze the changing trend of power deviation between adjacent time steps. If the current deviation continuously increases compared to the previous time step and significantly deviates from recent historical change levels, it is determined that the deviation has increased, and the corresponding feature needs to be positively adjusted. If the deviation decreases compared to the previous time step but is still higher than recent historical change levels, it is determined that the deviation has decreased, and the corresponding feature needs to be negatively adjusted. If the deviation is within the range of recent historical change levels or the change amplitude of consecutive time steps is small, the feature value remains unchanged and no adjustment is needed. The feature value should be increased or decreased based on the degree of deviation and trend direction. The calculated adjustment amount is applied sequentially to the features of the joint prediction input data corresponding to the time step to form a complete prediction input feature adjustment signal for the current time step.

[0127] The formula for dynamically adjusting the predicted input features is:

[0128] ;

[0129] in, This represents the dynamic adjustment amount of the predicted input features at the current time step, used to perform closed-loop correction on the joint predicted input data, so that the feature values ​​can reflect the power prediction deviation and time trend. The current time step prediction residual is the difference between the initial power prediction value at this time step and the actual measured photovoltaic power value at the previous time step, reflecting the magnitude and direction of the prediction deviation. The residuals are predicted for the previous time step and used to capture the trend of residual changes. and These represent the upper and lower bounds of the confidence interval for the predicted residuals at this time step, respectively. They are used to normalize the residual amplitude, making the adjustments at different time steps comparable and ensuring consistency with the dimensions of power and meteorological characteristics. To adjust the step size index, normalization is applied to the feature value to control the dynamic adjustment range, which is typically 0.05–0.30.

[0130] Specifically, when the residual It is positive and greater than the residual of the previous time step. When, the corresponding feature is positively adjusted; when the residual When the value is negative and its absolute value is greater than the absolute value of the residual at the previous time step, the corresponding feature is adjusted negatively; otherwise, the feature value remains unchanged.

[0131] S4.4: Integrate the predicted input feature adjustment amount of the predicted input feature adjustment signal with the current joint predicted input data to form the predicted input feature adjustment signal for this time step.

[0132] Furthermore, the photovoltaic power characteristics and local meteorological characteristics of each time step in the joint prediction input data are read sequentially according to time order, and the prediction input characteristic adjustment amount of the prediction input characteristic adjustment signal calculated in this time step is integrated with the characteristics of the corresponding time step of the joint prediction input data.

[0133] When the predicted residual of the current time step exceeds the normal fluctuation range and the trend of residual change shows an increasing deviation, the corresponding feature is adjusted by increasing its value. When the residual exceeds the normal fluctuation range and the trend shows a decreasing deviation, the corresponding feature is adjusted by decreasing its value. When the residual is within the normal fluctuation range or the change is small, the feature value remains unchanged. Based on the judgment result, the adjustment amount is applied to the feature value of each time step. After the feature adjustment of each time step is completed, the continuity of the entire joint prediction input data sequence is checked. By comparing the power change trend and meteorological dynamics of the previous time step, the adjusted features are made to maintain continuity and trend consistency in the time dimension. The features of all time steps after adjustment and correction are integrated in chronological order to generate a complete prediction input feature adjustment signal for the current time step.

[0134] S5: Schedule the predicted input feature adjustment signal of this time step to the joint prediction input data to obtain the adjusted joint prediction input data, and input the adjusted joint prediction input data into the selected weather scenario prediction model for fully connected mapping to generate the corrected power prediction value and the corrected confidence interval.

[0135] S5.1: The predicted input feature adjustment signal of this time step is numerically corrected and combined with the current joint predicted input data to form the closed-loop input data of this time step.

[0136] Furthermore, the photovoltaic power characteristics and local meteorological characteristics of each time step in the joint prediction input data are read in chronological order to ensure that the characteristics of each time step are completely corresponding. The prediction input characteristic adjustment signal calculated in this time step is applied to the characteristics of the corresponding time step in the joint prediction input data. By applying the adjustment amount to the original characteristics according to the deviation direction, that is, increasing or decreasing the characteristic value, the power and meteorological characteristics of each time step numerically reflect the deviation of the current prediction residual from the preliminary confidence interval.

[0137] When the residual of the current time step deviates beyond the normal fluctuation range and the trend shows an increase, the corresponding feature value increases; when the residual deviates beyond the normal range and the trend shows a decrease, the corresponding feature value decreases; when the residual is within the normal range or the change is small, the feature value remains unchanged. The entire joint forecast input data sequence is continuously checked. By comparing the power change trend of the previous time step with the local meteorological dynamics, the adjusted features are kept sequentially continuous in the time dimension and consistent with the historical change trend. The joint forecast input data that has been adjusted and checked is then reintegrated in time step order to form the closed-loop input data for this time step.

[0138] S5.2: Utilize the closed-loop input data of this time step to perform fusion statistics on the actual measured value of photovoltaic power and local meteorological dynamics to generate fused power information.

[0139] Furthermore, the actual measured value of photovoltaic power, local meteorological observation data, and corresponding joint prediction input data features of each time step in the closed-loop input data are read sequentially according to time order. The actual measured value of photovoltaic power at each time step is integrated with the local meteorological observation data at the corresponding time step to form a fusion feature. The power value and each local meteorological observation variable are put into the same data set in turn to form a fusion feature description for each time step.

[0140] Statistical analysis is performed on the changes in photovoltaic power and meteorological dynamics over continuous time steps, including calculating the increase or decrease trend of power over time and the magnitude of changes in meteorological variables, identifying the dependency pattern between photovoltaic power and local meteorological variables, smoothing the power change and meteorological dynamic sequences over continuous time steps to eliminate abnormal fluctuations or short-term noise, ensuring the continuity and stability of the fused data in the time dimension, and integrating the fused features of each time step in chronological order to form complete fused power information (the predicted value of photovoltaic power over continuous time steps, the corresponding local meteorological observation data, and the feature information extracted from each time step in the joint prediction input data).

[0141] Specifically, to ensure the consistency of the real-time prediction logic, the actual measured value of photovoltaic power used can be selected from the measurement value of the previous time step or the observation value within the historical window, rather than the power value that has already participated in feature correction at the current time step.

[0142] S5.3: Input the fused power information into the selected weather scenario prediction model and perform a fully connected mapping. Calculate the power prediction value and corresponding confidence interval for each time step through the mapping, and generate the corrected power prediction value and corrected confidence interval.

[0143] Furthermore, the photovoltaic power characteristics and local meteorological dynamic characteristics of each time step in the fused power information are read sequentially according to time order to form the input vector of each time step, so that the data of each time step completely corresponds to the power and meteorological information. The input vector of each time step is input into the fusion layer of the weather scene prediction model to integrate and transform the photovoltaic power characteristics and local meteorological dynamic characteristics, generate the integrated input feature vector of each time step, and input the integrated features into the time dependency mapping layer to perform mapping calculation on the features of continuous time steps, capture the time dependency relationship between photovoltaic power characteristics and meteorological dynamics, and form a complete set of time-series feature vectors.

[0144] The mapping results are checked for amplitude and continuity. If the photovoltaic power prediction value at a certain time step deviates significantly from the historical trend of the previous time step, or if the change in local meteorological observation data significantly exceeds the recent historical fluctuation range, then the time step is determined to be abnormal and needs to be corrected. When making the determination, the quantile interval, standard deviation range, or historical maximum rate of change of the data in the same period of time can be referenced to determine whether the current change is significant. For the time steps determined to be abnormal, the integrated input features of the corresponding time step are input into the prediction output and the closed-loop correction layer for fully connected mapping calculation. The features of each time step are mapped and integrated element by element. For the time steps not determined to be abnormal, the mapping results are kept unchanged, and the mapping results of all time steps are integrated in chronological order to generate the corrected power prediction value for the current time step. At the same time, the corresponding corrected confidence interval is calculated and generated based on the mapping results.

[0145] It should be noted that the fully connected mapping computation inputs the input vector of each time step into the mapping layer of the weather scene prediction model, calculates and integrates the photovoltaic power characteristics and local meteorological dynamic characteristics, and generates the mapping output for each time step.

[0146] Historical trends can be summarized by the average change direction of continuous time steps, reflecting the normal trend of power or meteorological variables. Historical fluctuation range can be obtained by calculating the quantile interval of historical data, reflecting the normal fluctuation range in the recent period. When determining whether the current time step is abnormal, the actual change is compared with the historical trend and fluctuation range. When the current change significantly exceeds the historical statistical characteristics, it is considered to be abnormal.

[0147] The photovoltaic power characteristics, local meteorological dynamic characteristics, and joint prediction input data characteristics of each time step in the fused power information are formed into an input vector in chronological order and input into the fusion layer for feature integration and transformation to generate an integrated input feature vector. The integrated input features are then input into the time dependency mapping layer in chronological order to process the continuous time step features, capture the time dependency relationship between power changes and meteorological dynamics, and generate a complete set of time-series feature vectors.

[0148] During the mapping process, the amplitude and continuity of the mapping results at each time step are checked. If the power prediction value at a certain time step deviates too much from the historical trend of the previous time step, or if there are abnormal changes in local meteorological characteristics, the time step is determined to be abnormal and needs to be corrected. For the time steps determined to be abnormal, the input features are integrated into the prediction output and the closed-loop correction layer for mapping calculation. The features of each time step are calculated and integrated element by element. For the time steps not determined to be abnormal, the mapping results are kept unchanged, and the mapping results of all time steps are integrated in chronological order to generate the corrected power prediction value for the current time step. At the same time, the corresponding corrected confidence interval is generated statistically based on the mapping results, realizing the dynamic response and closed-loop mapping of photovoltaic power prediction, while reflecting historical power trends, local meteorological dynamics, and prediction deviations.

[0149] In summary, this invention improves short-term prediction accuracy and stability by combining the photovoltaic power prediction residual with the initial confidence interval and applying it to the joint prediction input data to form a closed-loop mapping, dynamically responding to power changes and local meteorological fluctuations at each time step, and fusing historical power trends, meteorological changes, and weather scene characteristics to output corrected power prediction values ​​and confidence intervals.

[0150] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for real-time prediction of photovoltaic power generation based on weather scene recognition, characterized in that, include: Real-time power data from photovoltaic power plants and local meteorological observation data are collected and arranged in chronological order to form a time-series data sequence. Then, a time-series feature vector is generated through feature encoding. The reliability scores of photovoltaic power data and meteorological observation data are calculated using time-series feature vectors. The reliability scores of each data source are combined with the time-series feature vectors to form a feature score vector. The feature score vector is then input into a power regression network to extract the time dependence of historical power sequences and local meteorological dynamics, generating a time-series dependent feature sequence. The current weather scene category is identified based on the time-dependent feature sequence, and the corresponding weather scene prediction model is selected based on the identification results. The weather scene category and time-dependent features are combined to form joint prediction input data. The weather scene prediction model is used to perform a fully connected mapping to generate preliminary power prediction values ​​and preliminary confidence intervals. The prediction residual is calculated by comparing the preliminary power prediction value with the actual photovoltaic power measurement value of the previous time step, and the prediction input characteristics of the current time step are generated based on the prediction residual and the preliminary confidence interval. The predicted input feature adjustment signal of this time step is scheduled to the joint prediction input data to obtain the adjusted joint prediction input data. The adjusted joint prediction input data is then input into the selected weather scenario prediction model for fully connected mapping to generate the corrected power prediction value and the corrected confidence interval.

2. The real-time photovoltaic power generation prediction method based on weather scene recognition as described in claim 1, characterized in that, The specific steps for generating the temporal feature vector are as follows: Based on the operating status of the photovoltaic power station and surrounding meteorological observation information, power data and meteorological data at each time step are acquired, sorted and verified by time tags, and a continuous time series data set is generated. Consistency checks and anomaly identification are performed on continuous time series datasets. By filling in missing values, removing outliers, and verifying the time label order, time series data that continuously reflects power fluctuations and meteorological dynamics are obtained. Dynamic features are extracted from time series data by capturing power change patterns and meteorological fluctuations, and time series feature vectors are generated through feature encoding.

3. The real-time photovoltaic power generation prediction method based on weather scene recognition as described in claim 1, characterized in that, The combination forms a feature scoring vector, and the specific steps are as follows: Based on time-series feature vectors, the recent dynamics of photovoltaic power and local meteorological observation data are obtained. By analyzing the power change amplitude and meteorological fluctuations at the current time step, the credibility of each data source is calculated and a corresponding credibility score is generated. Historical prediction errors of photovoltaic power and the variation range of meteorological observation data are extracted to determine the credibility score of each data source. When the power variation range and meteorological fluctuations in the current time step exceed the variation range of the previous time step, they are combined into a feature score vector.

4. The real-time photovoltaic power generation prediction method based on weather scene recognition as described in claim 1, characterized in that, The specific steps for generating the time-dependent feature sequence are as follows: Based on the feature scoring vector, historical photovoltaic power data and local meteorological information are extracted, and then continuously arranged and correlated in time order to obtain an ordered dataset. Information is extracted and dependencies are analyzed from historical power sequences and local meteorological dynamics in an ordered dataset, and short-term and long-term change characteristics are captured to generate a dependency dataset. The dependent dataset is input into the power regression network for feature integration. In the power regression network, the influence of different features is adjusted according to the power change amplitude and meteorological fluctuations. Historical photovoltaic power data is fused with local meteorological information to generate a time-dependent feature sequence. The power regression network is a feature integration submodule of the selected weather scenario prediction model. It is used to integrate and map the time step features, historical power and local meteorological information in the joint prediction input data to generate a time-dependent feature sequence.

5. The real-time photovoltaic power generation prediction method based on weather scene recognition as described in claim 1, characterized in that, The specific steps for generating preliminary power prediction values ​​and preliminary confidence intervals are as follows: Extract photovoltaic power change information and local meteorological dynamics at continuous time steps from the time-dependent feature sequence, calculate the statistical average value of power change, and obtain the current weather feature data set; Based on the comparison of the current weather feature data set with the power change trend and meteorological dynamics of the previous time step, and combined with the weather scene classification standard, the current weather scene category is determined, and the determined weather scene category is combined with the time-dependent features to form joint prediction input data. The joint forecast input data is input into the selected weather scenario forecast model and fully connected to generate preliminary power forecast values ​​and preliminary confidence intervals.

6. The real-time photovoltaic power generation prediction method based on weather scene recognition as described in claim 5, characterized in that, The step of inputting the joint prediction input data into the selected weather scenario prediction model for fully connected mapping to generate preliminary power prediction values ​​and preliminary confidence intervals includes: based on the time-dependent feature sequence, integrating and transforming the time step features and historical power information in the joint prediction input data to generate a set of time-series feature vectors of power change trends and dynamic relationships. When there are abnormal fluctuations in the amplitude of the mapping vector of the time series feature vector set, the joint prediction input data is numerically corrected to obtain the adjusted joint prediction input data; The adjusted joint forecast input data is then input into the selected weather scenario prediction model for fully connected mapping to generate preliminary power forecast values ​​and preliminary confidence intervals.

7. The real-time photovoltaic power generation prediction method based on weather scene recognition as described in claim 6, characterized in that, The specific construction process of the weather scenario prediction model is as follows: A weather scenario prediction model is constructed based on an input fusion layer, a time-dependent mapping layer, and a prediction output and closed-loop correction layer. The input fusion layer integrates time-step photovoltaic power with local meteorological features based on joint prediction input data, analyzes the power change amplitude and meteorological dynamics, and generates a fused integrated input feature vector by combining the power change amplitude and meteorological dynamic range. The time-dependent mapping layer uses the fused integrated input feature vector to perform time-dependent relationship analysis on photovoltaic power and local meteorological information at continuous time steps, capture short-term and long-term change trends, and form a set of time-series feature vectors. The predicted output and closed-loop correction layer, based on the time series feature vector set, integrates and maps the time series feature vectors of each time step to obtain a preliminary estimate of the photovoltaic power at the current time step, and integrates them to generate a corrected power prediction value and a corrected confidence interval.

8. The method for real-time prediction of photovoltaic power generation based on weather scene recognition as described in claim 1, characterized in that, The specific steps for generating the time-step prediction input feature adjustment signal are as follows: Based on the preliminary power prediction and preliminary confidence interval of this time step, the actual photovoltaic power measurement value of the previous time step is obtained. By comparing the difference between the current prediction value and the actual measurement value of the previous time step, the power deviation between adjacent time steps is calculated. By utilizing the power deviation between adjacent time steps and the upper and lower bounds of the initial confidence interval of the current time step, the deviation of the prediction residual relative to the confidence interval is calculated to generate the prediction input feature adjustment signal for the current time step.

9. The method for real-time prediction of photovoltaic power generation based on weather scene recognition as described in claim 8, characterized in that, The method of generating the prediction input feature adjustment signal for the current time step by using the power deviation of adjacent time steps and the upper and lower bounds of the preliminary confidence interval of the current time step to calculate the deviation of the prediction residual relative to the confidence interval includes: combining the difference between the preliminary power prediction value and the actual power of the previous time step to calculate the prediction residual of the current time step, and combining the upper and lower bounds of the preliminary confidence interval of the current time step to form a sequence of deviation ratios of the residual relative to the confidence interval. By combining the residual deviation ratio with the predicted residual trend of the previous time step, the numerical adjustment magnitude of the predicted input feature is calculated, and the predicted input feature adjustment signal is generated. The predicted input feature adjustment amount of the predicted input feature adjustment signal is integrated with the current joint predicted input data to form the predicted input feature adjustment signal for this time step.

10. The method for real-time prediction of photovoltaic power generation based on weather scene recognition as described in claim 1, characterized in that, The specific steps for generating the corrected power prediction value and the corrected confidence interval are as follows: The predicted input feature adjustment signal of this time step is numerically corrected and combined with the current joint predicted input data to form the closed-loop input data of this time step; The actual measured value of photovoltaic power and local meteorological dynamics are fused and statistically analyzed using the closed-loop input data of this time step to generate fused power information. The fused power information is input into the selected weather scenario prediction model and fully connected mapping is performed. The power prediction value and corresponding confidence interval for each time step are calculated through the mapping, and the corrected power prediction value and corrected confidence interval are generated.

Images